Technical ReportPDF Available

On the futures of technology in education: Emerging trends and policy challenges

Authors:

Abstract

This report identifies key emerging technologies and discusses their potential impact in education. Drawing on academic research and grey literature, it focuses on a set of ongoing technical developments that could redefine both education and society at large in fundamental ways. It provides insights into the affordances of those technologies and important societal implications, discussing how they may reconfigure education against the background of learning theory. It also considers the socio-material basis of digital technologies, as well as key factors ‒ such as climate change, demographic transitions, environmental concerns, and the growth of mental health problems among the young ‒ that are shaping the emerging educational landscape. The report aims to go beyond the state-of-theart and facilitate richer discussions on the potential impact of emerging technologies in education in order to support long-term strategic thinking. Based on that, it offers recommendations to ensure that future policy actions are aligned with the societal and educational needs.
Joint
Research
Centre
JRC SCIENCE FOR POLICY REPORT
© European Union, 2023
EUR 31652 EN
ISSN 1831-9424
Tuomi, I.
Cachia, R.
Villar-Onrubia, D.
2023
On the Futures of
Technology in Education:
Emerging Trends and
Policy Implications
This publication is a Science for Policy report by the Joint Research Centre (JRC), the European
Commission’s science and knowledge service. It aims to provide evidence-based scientific support
to the European policymaking process. The contents of this publication do not necessarily reflect the
position or opinion of the European Commission. Neither the European Commission nor any person
acting on behalf of the Commission is responsible for the use that might be made of this publication.
For information on the methodology and quality underlying the data used in this publication for which
the source is neither Eurostat nor other Commission services, users should contact the referenced
source. The designations employed and the presentation of material on the maps do not imply the
expression of any opinion whatsoever on the part of the European Union concerning the legal status
of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers
or boundaries.
EU Science Hub
https://joint-research-centre.ec.europa.eu
JRC134308
EUR 31652 EN
PDF ISBN 978-92-68-07327-8 ISSN 1831-9424 doi:10.2760/079734 KJ-NA-31-652-EN-N
Luxembourg: Publications Office of the European Union, 2023
© European Union, 2023
The reuse policy of the European Commission documents is implemented by the Commission Decision
2011/833/EU of 12 December 2011 on the reuse of Commission documents (OJ L 330, 14.12.2011,
p. 39). Unless otherwise noted, the reuse of this document is authorised under the Creative Commons
Attribution 4.0 International (CC BY 4.0) licence (https://creativecommons.org/licenses/by/4.0/). This
means that reuse is allowed provided appropriate credit is given and any changes are indicated.
For any use or reproduction of photos or other material that is not owned by the European Union
permission must be sought directly from the copyright holders.
The European Union does not own the copyright in relation to the following elements:
- cover page, © berCheck, image #629164020, 2023 - stock.adobe.com
How to cite this report: Tuomi, I., Cachia, R. and Villar-Onrubia, D., On the Futures of Technology
in Education: Emerging Trends and Policy Implications, Publications Office of the European Union,
Luxembourg, 2023, doi:10.2760/079734, JRC134308.
Layout: Carmen Capote de la Calle
SCIENCE FOR POLICY REPORT
Tuomi, I.
Cachia, R.
Villar-Onrubia, D.
2023
On the Futures of
Technology in Education:
Emerging Trends and
Policy Implications
Table of contents
06
05
06
62
58
07
82
83
10
32
18
35
Abstract
Foreword
Acknowledgements
References
5. Conclusions and policy implications
Executive summary
List of abbreviations and definitions
List of tables
1. Introduction
1.1 The emerging landscape of education
1.2 Aim of the report
1.3 Methodology
3. Data
3.1 Implications of data spaces for education and
learning
2. The Next Internet
2.1 Towards 6G networks
2.2 Immersive technologies: Extended reality,
holograms and virtual worlds
2.3 Digital credentials and identifiers
2.4 Post-quantum cryptography and zero-knowledge
proofs
4. AI and learning analytics
4.1 Generative AI for teaching and learning
4.2 Open learner models, agency, and humans-in-the-
loop
4.3 Learning analytics and new forms of assessment
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
5
Digital technologies are not new. They have
long permeated our society in the form of
devices and soware applications that we
use for working, studying or personal pur-
poses.
What is new is the dramatic increase in
speed of development and uptake for
these digital technologies. Another novelty
is their disruptive and transformative ca-
pacities: big data, blockchain, 6G and the
new generation internet, artificial intelli-
gence (AI) and the next generation of virtual
worlds, are turning upside down the world in
which we live, and the way we understand
our societies and relations.
We are in the middle of a new socio-tech-
nical revolution that will dramatically
change who we are, how we live and how
we relate to one another. In response to this
revolution, the EU has the ambition to make
this Europe’s ‘Digital Decade’, and to drive
a safe and secure digital transformation
by 2030, succeeding in the twin digital and
green transition.
The education sector needs to adapt to this
new, and rapidly changing, information and
media environment. Education, both formal
and informal, provides fundamental build-
ing blocks that define various keys aspects
of each person’s cognitive development,
personality and life opportunities. That is
the reason why understanding how dig-
ital technologies affect and transform the
educational process, and educational sys-
tems at large, is key for each society and,
in particular, for Europe, in its quest for a
human-centred digital transformation.
At the Joint Research Centre of the Eu-
ropean Commission, we aim to identify
Foreword
emerging technologies and trends that
play a key role in the transformation of
education across Europe. We also work to
understand the ways in which they could
shape the future of teaching, learning, as-
sessment, and all other processes and so-
cial practices underpinning the delivery of
education. Our research aims to anticipate
risks, so that policy makers can mitigate
them with decisive policy action, but also
identify potential opportunities for Europe
to develop more inclusive and high-quality
educational systems.
This report is part of the Joint Research
Centre’s effort to imagine the futures of
technology in education and outline as-
sociated policy implications. It has the
objective to stimulate the discussion, and
identify potential scenarios for the use of
emerging digital technologies, and par-
ticularly AI, in the field of education, going
beyond the state-of-the-art and bringing
some forward-looking ideas.
We believe it contributes to the crucial de-
bate on the digital transformation of ed-
ucation through unfolding some complex-
ities on how these new technologies are
likely to affect education. We also believe
that this report will serve as a guide to Eu-
ropean policy makers to ensure that Euro-
pean educational systems remain effective
and human-centred, while adopting emerg-
ing digital technologies responsibly.
Carlos Torrecilla Salinas
T1 Unit - Digital Economy
Joint Research Centre (JRC)
European Commission
6
SCIENCE FOR POLICY REPORT
This report identifies key emerging technol-
ogies and discusses their potential impact
in education. Drawing on academic research
and grey literature, it focuses on a set of
ongoing technical developments that could
redefine both education and society at large
in fundamental ways. It provides insights
into the affordances of those technolo-
gies and important societal implications,
discussing how they may reconfigure edu-
cation against the background of learning
theory. It also considers the socio-material
basis of digital technologies, as well as key
Abstract
Acknowledgements
The authors of this report would like to
thank the following experts who have re-
viewed, provided input and support, or par-
ticipated in one of the validation workshops
as part of this report: Duuk Baten, Paulo
Blikstein, Simon Buckingham Shum, Teresa
Cerratto Pargman, Maarten de Laat, Keri
Facer, Ulla-Maaria Koivula, Teemu Leinon-
en, Jeremy Roschelle, Demetrios Sampson,
Neil Selwyn, Mike Sharples, Carolina Valla-
dares Celis, Stéphan Vincent-Lancrin and
Riina Vuorikari. Their insights have helped
steer the direction of this report, making it
more relevant for educational policymak-
ers and researchers. We are also grateful
to EC scientific and policy officers, namely
Carlos Torrecilla Salinas, Yves Punie, Judith
Cosgrove, Vicky Charisi, Eleonara Bertoni,
Michelle Vespe, Alexander Kotsev, Nikoleta
Giannoutsou, Michael Horgan, Maria Gkoun-
touma, Anusca Ferrari, Deirdre Hodson, who
supported this work and provided valuable
and detailed feedback on the final dra of
this report. Finally, we would also like to
thank Carmen Capote de la Calle for the
design of the report.
Authors
Ilkka Tuomi
Meaning Processing Ltd.
Romina Cachia
European Commission, Joint Research
Centre, Seville, Spain
Daniel Villar Onrubia
European Commission, Joint Research
Centre, Seville, Spain
factors ‒ such as climate change, demo-
graphic transitions, environmental concerns,
and the growth of mental health problems
among the young ‒ that are shaping the
emerging educational landscape. The re-
port aims to go beyond the state-of-the-
art and facilitate richer discussions on the
potential impact of emerging technologies
in education in order to support long-term
strategic thinking. Based on that, it offers
recommendations to ensure that future
policy actions are aligned with the societal
and educational needs.
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
7
Digital technology currently pervades all di-
mensions of society, playing a central role
in our lives. It shapes our everyday exist-
ence and has redefined the way we engage
in activities related to work, civic participa-
tion, health, sports, entertainment and, of
course, education. At the same time, tech-
nologies are actively shaped by the estab-
lished practices, dynamics, and values of
the social and cultural contexts in which
they are conceived, used or embedded in
any other ways.
This report aims to map emerging tech-
nologies1 and trends with high potential to
influence education2 and provide a critical
perspective on them. It identifies a series of
technologies ‒ currently at different stages
of development and adoption ‒ which have
a high potential to provoke significant so-
cietal transformations. The report focuses
on how such technologies may contribute
to the redefinition of educational practices,
processes, and organizations. While offering
detailed technical information, the report
discusses all these innovations against the
background of learning theory, in order to
properly unpack their potential influence in
teaching, learning and the organization of
education at large.
The report argues that the innovation under
development that will most likely impact
the information and media ecology over
the coming years has to do with the con-
nectivity infrastructure underpinning wire-
less communications. The increase in data
transfer speed and the latency reduction
1. New technologies and the continuing development of existing
technologies that are expected to be available in the coming
years and are could have significant social and economic impact.
2. In this report, the term education is used broadly to cover all
types and levels of education.
Executive summary
afforded by 5G networks will open up new
opportunities, for instance, in relation to im-
mersive virtual environments and extended
reality (XR) applications. Eventually, with the
rollout of the 6G standard, the internet will
operate over a wider spectrum of frequen-
cies and will incorporate a ubiquitous net-
work of highly precise ‘sensing’ devices. The
report anticipates that such developments
will take the internet to a whole new para-
digm of synchronization and blending of the
physical and the digital. The so-called Next
Internet will constitute a new infrastructure
connecting cognition, space, and action in
new ways. This will be primarily realized
through actionable representations of the
physical world, meaning that human action
and interaction will become mediated by a
real-time digital layer.
In terms of pedagogical implications, the
report highlights that the intersections of
cognition, space and action are particularly
relevant to constructivist theories of learn-
ing, which approach practical action as a
key source for learning (as illustrated by the
thinking of Dewey, Vygotsky and Piaget).
Thanks to faster networks, enhanced com-
puting capabilities and lightweight devic-
es, we argue that immersive technologies
‒ such as virtual reality, augmented reality,
and holograms ‒ will be able to enable the
delivery of authentic experiential learning
which would otherwise be unfeasible due,
for example, to the safety risks and high
costs associated with the real-life expe-
riences that these technologies can help
simulate. However, their full potential will
only be realized when the physical, social,
and digital worlds become interlinked in real
time through the connectivity changes de-
scribed in the report.
8
SCIENCE FOR POLICY REPORT
Learning assessment and competence cer-
tification are integral to education, and the
report anticipates that emerging technolo-
gies could redefine the way in which related
processes are organized and how different
stakeholders (e.g., educational institutions,
employers, graduates) interact with each
other. Distributed and decentralised sys-
tems – such as blockchain and micro-cre-
dentials – are not new anymore, and their
disruptive potential was indeed predict-
ed years ago. While high impact remains
largely unrealized, micro- and verifiable
credentials will probably gain ground in a
context where all citizens need to engage
in continuous development and upskilling,
at least to some extent. This is not without
risks as, given the absence of a holistic, in-
clusive, and quality-driven approach, a dig-
ital credentials-driven system also has im-
plications that we need to consider. These
kinds of certifications become particularly
relevant in a context where the curriculum
of formal education needs to be kept up to
date at a faster pace. Modularity, allowing
obsolete parts to be more easily replaced,
is essential here. Nevertheless, when em-
bedding these innovations into educational
systems it will be essential to plan for the
possibility of the current encryption infra-
structures becoming redundant in the not-
so-distant future because of advances in
quantum computing.
Learning analytics and, more generally, the
mining of data in education is already a pri-
ority for institutions, policymakers, and edu-
cational technology vendors. The so-called
datafication of education is only expected
to keep gaining ground across all levels and
settings, meaning that large-scale data on
education will be at the centre of knowledge
society transformations. Such data are ex-
pected to reinforce and inform the devel-
opment of new pedagogical approaches,
learning technologies, and education policy;
however, it will be necessary to determine
which data is required for this and find ways
to ensure that all key stakeholders have ac-
cess to them. The aspiration is to be able
to record learning processes, instead of just
outcomes, and by doing so to ultimately im-
prove student engagement and attainment.
Indeed, datafication of education will not
necessarily add value unless the focus re-
mains on the quality of outcomes. However,
ethical concerns about personal data pro-
cessing and their potential misuse need to
be carefully addressed, as does the current
reliance on infrastructures that are largely
in the hands of just a few commercial pro-
viders and may pose vendor lock-in or lack
of interoperability risks.
AI systems have reached such a level of
maturity that there are already products
available to the general public that can
mimic – and even outperform – humans
at certain tasks. In particular, there seems
to be broad agreement that generative AI
and foundation models will have a lasting
impact on education, and there are indeed
many examples of how they may positive-
ly (and negatively) affect teaching, learning
and assessment. Automatic video caption-
ing, translation, video summarizing and
highlight extraction, along with text-to-an-
imation and voice to synthetic video all
provide many new opportunities for peda-
gogical innovation and enhanced accessi-
bility. In this emerging context – where AI
systems can support the development of
capabilities for learning as they become
‘smart companions’, ‘learning partners’ or
‘cognitive tools’ – metacognition and re-
flection gain importance over learning by
rote. The resulting assemblages will entail
a redistribution of agency among humans
and machines. It will be essential to make
informed decisions on the activities that we
could ‘delegate’, and the ones we should
retain, in order to maximize learning oppor-
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
9
tunities. Moreover, the report recognizes the
need for critical evaluation of existing dis-
courses regarding the impact of emerging
technologies on education and learning, for
example, to recognise their role in widening
the digital divide and the risks of commer-
cial interests unduly dominating the sector.
The report highlights the importance of reg-
ulating technological innovation to allow
diverse educational visions to thrive, also
noting the need to connect industrial poli-
cy with theories of learning and education
policy.
Open learner models that support the learn-
ers in reflecting and regulating their learning
processes are particularly relevant from a
learning theory perspective, and their im-
portance will increase as agency becomes
distributed between AI systems and hu-
mans. Self-reflection and self-regulation
are essential for learning to happen, and
better understanding of the development
of agency is also important for AI govern-
ance models beyond education.
Likewise, the capability approach becomes
relevant in new ways for educational pol-
icy when the emerging technologies are
widely used for learning and teaching. In
this context, the development of agency ‒
what people are free to do and achieve in
pursuit of goals and values they regard as
important ‒ can be viewed as the ultimate
goal of education. Such a capability-based
approach is expected to gain relevance in
educational theories and policies as do-
main-specific and ‒ more generally ‒ epis-
temic skills become less important than
they used to be. Overall, skills and com-
petences are concepts that will need to be
reconsidered in the light of emerging hu-
man-computer interaction entanglements.
The impact of emerging technologies on
assessment, as a factor that heavily in-
fluences teaching and learning, deserves
particular attention. Formative assessment
may increasingly gain prominence thanks
to learning analytics developments, while
AI will force a rethink about what should be
assessed, how assessment data should be
collected and used, and how new assess-
ment practices might enhance learning. In
any case, a key priority should be to avoid
automating assessment practices that are
no longer relevant in the new context.
All the emerging technologies and asso-
ciated trends described in this report are
profoundly interconnected with each other,
while underpinned to a considerable extent
by data and the increasing datafication of
all aspects of society that characterises the
postdigital age. The most significant policy
challenge is how to best harness the po-
tential of the new emerging socio-technical
ecology of education without compromising
the wellbeing of students and educators. AI
must be trustworthy before it is fully em-
bedded into the curriculum. Understanding
the flows of personal data that result from
the interaction between students and com-
puter systems is an essential requirement
for anyone making decisions regarding the
introduction of emerging technologies in ed-
ucational settings. Understanding the envi-
ronmental impact of new technologies (e.g.,
in the energy-intensive training and use of
AI systems) is equally important, allowing
informed decisions to be taken about the
extent to which they should be deployed in
education and other areas of society. The
digital transformation of education and so-
ciety should not be accomplished at the ex-
pense of the green transition and sustain-
ability.
The interplay – increasingly mediated by
AI – between the digital ecosystem and the
education sector will require an unprece-
dented interlinking of educational, digital,
environmental, and industrial policy, espe-
cially the regulation around the structuring
of markets and innovation processes.
10
SCIENCE FOR POLICY REPORT
EU policymaking relies on scientific research
and evidence-based recommendations to
positively change society. Understanding
current practices, behaviours, sentiments,
opinions, and capabilities – just to mention
a few key factors – is essential for the de-
sign of effective policies in any sector. It is
equally important to examine the patterns
that connect past and present socio-ma-
terial dynamics. Apart from looking at the
past and the present, the development of
timely policy actions that can enhance the
readiness for potential challenges and op-
portunities requires also imagining future
scenarios. The future is yet to happen and
is therefore by definition impossible to re-
search empirically; but how potential fu-
tures come to be imagined and who takes
part in such speculative endeavours active-
ly shape the unfolding of events and con-
tribute to defining the versions of the future
that will eventually come to pass.
Speculative works on the futures of educa-
tion tend to gravitate around digital tech-
nology (Facer, 2011; Ross, 2023; Selwyn at
al., 2020). It is important to bear in mind
that the visions of those behind the creation
of technology heavily influence the way we,
as a society, get to imagine what education
might look like in the future. This means
that the views, interests, and priorities of
unprivileged populations are largely ab-
sent from dominant narratives and visions
on the future of education. Such a lack of
diversity ‘leads to narrow and unimagina-
tive discussions about what AI ought to be,
and the education issues that are deemed
worthy of being addressed, and what spe-
cific social responsibilities are chosen above
others’ (Selywn, 2022, p. 628).
Over the last three decades, digital infor-
mation and communication technologies
– taking the form of soware, data, algo-
rithms, connectivity infrastructures, and
computing devices of different shapes and
sizes – have increasingly permeated all di-
mensions of society. Indeed, such a level of
pervasiveness has rendered the term ‘digi-
tal’, and related prefixes such as ‘cyber-’ or
‘e-’, somehow meaningless as it is no longer
possible to approach any fields of human
activity without considering the very cen-
tral role digital technologies play. The term
‘postdigital’ was already proposed before
the end of the 20th century (Cascone, 2000;
Pepperell & Punt, 2000), but it has gradu-
ally gained ground in the academic litera-
ture as a way of characterizing the current
state of entanglement of technology and
society. Present in a rather diverse range of
disciplines, postdigital refers to ‘a state of
becoming where the human and the digi-
tal are interacting, co-creating, and merg-
ing in ways that are beyond imagining’ (Ball
and Savin-Baden, 2022, p. 754). However,
instead of thinking of it as a chronological
term that suggests that we now live ‘aer
the digital’, it is best understood as a crit-
ical attitude or philosophy concerned with
the constitution, theoretical orientation, and
consequences of the so-called digital world
(Peters and Besley, 2019). The postdigital
condition signals ‘our raising awareness of
blurred and messy relationships between
physics and biology, old and new media,
humanism and posthumanism, knowledge
capitalism and bio-informational capital-
ism’ (Jandrić et al., 2018, p. 896).
The concept of media and information
‘ecologies’ can also help us to refine our
1 Introduction
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
11
understanding of the role that technology
plays in mediating the relationships be-
tween humans and non-human actors and
their environments. Adopting an ecological
perspective in this sense entails looking at
systems made of highly heterogeneous en-
tities, including living beings as well as nat-
ural and socially constructed elements, all
of them constantly coevolving in the course
of continuous and non-teleological dynam-
ics (Taffel 2019, Nardy & O’Day 2009).
Technologies exist within the socio-material
contexts in which they are invented, devel-
oped, and put into use. It is through situated
practices that they become interesting, im-
portant, and relevant. Therefore, it is impos-
sible to separate technological functionality
from the human and social settings in which
they are embedded. This takes the form, for
example, of technological standardization,
regulatory environments, economic ar-
rangements, culturally shared expectations,
both rational and unfounded fears, and cur-
rent social practices.
Innovation becomes real only when it is used
and appropriated in social practice (Tuomi,
2002). Technology is never neutral because
every technology provides some affor-
dances and enables some ways of using it,
at the same time making other uses diffi-
cult or impossible (Miller, 2021). Technical
affordances, however, are also underpinned
by culture and thus influenced by norms,
values, and current knowledge. The tech-
nological advances that we discuss below,
therefore, need to be considered in a broad-
er socio-technical context of development,
driven by certain objectives, and uses. Tech-
nology is, by definition, instrumental and we
need to ask what it is for3 and whose inter-
ests it serves best. This in itself suggests
that innovation and technological ‘progress’
3. Tuomi (2001) has argued before that technical change implies
development only if it expands human capability for action that
we have reason to value.
are inherently political topics that require a
variety of viewpoints to be discussed and
debated, always bearing in mind that they
will benefit different populations unevenly.
In this report we address new emerging
technological developments and trends,
specifically emerging technical innovations
and usage-related developments, with a
focus on their educational implications and
policy action. These include, most notably,
important standardization initiatives that
could shape technology use in the educa-
tional domain. Emerging technologies are
constantly perceived as vectors of change
with, presumably, the capacity to fix or neu-
tralize the most pressing challenges faced
by education. They are designed, sold,
bought, and adopted with the hope that
they will radically transform, for the bet-
ter, learning within formal education and
beyond. However, as highlighted by Fac-
er & Selwyn (2021), the last four decades
of research reveal that technologies alone
have not been able to transform education,
improve learning, resolve inequalities, or re-
duce teachers’ workloads, while also show-
ing that the unintended consequences that
may result from technology adoption can
be hardly predicted and that actual impact
always depends on socio-technical contexts
and factors.
Overall, the perceived potential of technol-
ogies tends to eclipse notions of what can
be realistically expected based on facts and
lessons learnt from the past about the ac-
tual implications of technology for society.
In the case of science and education, while
advances in information and communica-
tion technologies have certainly prompted
a reconfiguration of the way we produce,
share and access knowledge, excitement
around new developments has oen led
to initial hyperbole and subsequent disap-
pointment, for example, regarding radio
12
SCIENCE FOR POLICY REPORT
and TV (Buckingham, 2007) and computers
(Cuban, 2001). The history of technology
shows that it is impossible to accurately
foresee the technologies of the future and
their social uses, with abundant examples
of predictions about technologies such as
the telephone and the personal computer
that ‘were little more accurate than flipping
a coin’ (Nye, 2006: 211).
Still, there is value in attempting to antici-
pate the role that emerging socio-technical
trends might play in education, as well as in
trying to understand the ways in which they
could foster or erode different visions for
the future of education. In this regard, the
formulation of relevant, effective, and time-
ly educational policies in support of such
visions requires drawing on history, examin-
ing current developments, and looking into
possible futures that may result from the
potential evolution of incipient trajectories.
Our current digital infrastructure rests on
connectivity provided by the internet, and
the massive transformation it is currently
undergoing means that the overall digi-
tal landscape will soon look very different.
Over the next ten years, this connectivity in-
frastructure will see fundamental changes
with the introduction of 5G-advanced and
6G networks. In particular, 6G networks will
connect physical, digital and social worlds
in fundamentally new ways, with pro-
found implications for learning, knowledge
sharing, and knowledge use. This change
is systemic: 6G, in combination with other
significant developments and trends (e.g.,
the reduction of costs of computing capac-
ity and edge computing) will bring together
several key technologies that are now rap-
idly advancing in parallel, with limited con-
vergence for now. These include immersive
and augmented reality technologies, artifi-
cial intelligence (AI) and machine learning
technologies, sensors, and blockchain. Per-
haps most importantly, data are becoming
a crucial component that underpins both
the development and the use of these
emerging technologies. We introduce some
key developments in these areas below for
further discussion.
It has been suggested that data-driven AI is
about to generate a socio-technical ‘Cam-
brian explosion’ in the coming years (Liang,
2023; Tuomi, 2022). This is because ‒ even
without much further technical develop-
ment ‒ generative AI and related foundation
models are already having a major impact
on many dimensions of society, including
education and learning, work, and everyday
life. Existing AI technologies can be used as
tools in the AI development process itself,
and the impact is cumulative. Such posi-
tive (i.e. amplifying) feedback is typical for
those general-purpose technologies, such
as steam power, electricity, and computers,
which have become the key technologies
of emerging techno-economic paradigms
of the past. In AI, the development is ex-
traordinarily fast, and social and econom-
ic consequences can already be observed.
Because of this, the report also highlights
the importance of understanding innova-
tion dynamics in this area. As Blikstein et
al. (2022) show in their empirical analysis
of the futures of education, technology
firms that develop AI for education spend
considerable effort in redefining what intel-
ligence, education, and learning are. These
definitions are then fed back to educators
and policymakers as images of the future,
shaping their expectations and vision on the
role that technology could and should play
in education.
It is, therefore, essential to critically as-
sess existing discourses about the impact
of emerging technologies in education, so
as to ensure that in Europe our priorities for
an inclusive and high-quality education ‒
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
13
as defined in the Digital Education Action
Plan ‒ remain unchallenged. It is important
that policymakers understand, for example,
how commercial interests shape these dis-
courses (Davies et al., 2021). In the case of
data and AI, innovation dynamics can lead
to natural monopolies unless regulation can
shape the industry structure so that multi-
ple visions of the future of education and
technology can compete for prominence,
making sure that ‘AI works for people and
is a force for good in society’; as defined by
the European approach to artificial intelli-
gence.4 Understanding critical perspectives
from diverse stakeholders is vital (e.g. see
Selwyn 2023 in relation to AI in education).
Within this technological landscape, we can
see how industrial policy becomes linked
with theories of learning and education pol-
icy.
The report introduces key emerging trends,
at different stages of development, from a
technology-oriented point of view. The goal
is to facilitate a rich and wide-ranging dis-
cussion, which could be complemented with
further research drawing on methods from
futures and foresight research.5 It starts by
looking at technology from a rather tradi-
tional engineering-oriented standpoint, to
then focus on social impact and implica-
tions for education with the main aim being
to offer relevant insights for policymaking in
education.
4. https://digital-strategy.ec.europa.eu/en/policies/european-ap-
proach-artificial-intelligence
5. On possible methodological approaches, see, for example,
Miller & Tuomi (2022), Miller (2018), Tuomi (2019), and their
references.
1.1 The emerging landscape
of education
New technologies6 are oen viewed as
solutions to widely recognized societal
problems. In the case of education, they are
actively proposed by different stakeholders
as a ‘fix’ to pressing issues, for example
the vision of AI helping teachers in marking
student homework. Teachers spend many
working hours in this task (OECD, 2019b),
and it appears to be a natural opportunity
for automation. Although the situation var-
ies greatly across countries, teachers are
oen overloaded with administrative tasks
and marking (Vuorikari et al., 2020). Policy-
makers in many countries are keen to find
cost-effective ways of reducing these work-
loads in an attempt to make teaching more
attractive and keep teachers teaching.
Many visions of the future of AI in education
also claim that AI will radically transform
formative assessment. For example, AI sys-
tems could provide continuous feedback for
students and make real-time assessment
and personalised or adaptive learning pos-
sible. In that vision of AI-supported learning,
there remains little need to mark student
homework.
Future visions of technology are rarely vi-
sions about any realistic or imaginable fu-
ture. They are frequently based on past
experiences. More oen, they are predic-
tions about the impact of new technologi-
cal functionality while the rest of the world
stays the same. For predictive models, this
assumption of ‘other things being equal’
is a practical necessity. In real life it is, of
course, fiction. Educational systems are
composed of complex interdependent sys-
6. In this report the term ‘new technologies’ is not determined
by the date when a given technological innovation was first
invented or made publicly available. Instead, it considers the
extent to which group of users have found a meaningful way
to integrate latent innovative opportunities in the current social
practice (Tuomi, 2002).
14
SCIENCE FOR POLICY REPORT
tems, and their change requires a systemic
approach. The adoption and appropriation
of a new technology in education has a sys-
tem-wide impact, much of which remains
invisible when technology is understood as
a solution to a given problem.
Computers have oen been used to auto-
mate apparently well-defined tasks, such
as ‘marking homework’, at the same time
‘hard-wiring’ work practices that reflect the
constraints of the past. More generally, digi-
tal technologies can easily be used to scale-
up pedagogic practices that resulted from
the constraints of the past. Automation,
by necessity, starts from a static view on
a given system, assuming ceteris paribus,
and successful automation makes system
change difficult. To avoid being stuck with
the needs of the past, we should, therefore,
ask what the constraints of the future will
be, and how emerging technologies could
address the emerging opportunities.
For example, it is now clear that AI will have
a profound impact on labour-market skill
demand. A recent study by economists from
Goldman Sachs (Briggs & Kodnani, 2023)
estimated that roughly two-thirds of cur-
rent jobs are exposed to some degree to AI
automation, and that generative AI could
provide a substitute for up to a quarter of
current work. Similar findings were found by
Sostero & Tolan (2022) in their classifica-
tion of over 13,000 different skills required
by employers, whereby several clusters that
emerged were related to advance digital
skills, including AI. For such skills, an associ-
ated 10.8% higher wage was offered. Many
non-routine and knowledge-based work
tasks will soon be supported by AI systems.
Likewise, climate change and the decline in
biodiversity pose large-scale challenges to
humanity. The energy consumption of data
centres and both the training and use of AI
systems are becoming potentially impor-
tant sources of global warming. In addition
to environmental consequences, AI is pos-
sibly amplifying societal and geographical
divides while limiting sovereignty because
computing capacity is unevenly distributed
across the world (Crawford, 2021; OECD,
2022). For education, this means that those
who have the material capacity to develop
AI systems will be able to shape learning
processes globally.
Lack of digital competence among teach-
ers – and indeed the entire population – are
seen as key challenges in Europe and be-
yond. These problems may, however, look
very different as digitalization and develop-
ments in AI continue. At the same time ‒
although causal links with digital media are
still unclear (Odgers & Jensen, 2020) ‒ de-
pression, anxiety, and behavioural disorders
are rapidly increasing among adolescents
(Braghieri, Kevey & Makarin, 2022). Educa-
tional policymakers are now focusing on the
development of social, emotional, and me-
ta-cognitive competences (Chernyshenko
et al. 2018, Council of the European Union,
2018, Sala, Punie, Garkov & Cabrera, 2020).
Economically useful skills and knowledge
are increasingly being developed outside
formal education. In short, the emerging
landscape of learning and access to
knowledge looks quite different from
what it was when the existing institu-
tions of education were formed over the
past centuries.
In addition, control and power are shiing
across educational systems from some ac-
tors to others. What people learn and how
they learn is increasingly shaped by com-
mercial actors that remain largely outside
the control of regional, national, and Euro-
pean policymakers. The EdTech industry ac-
tively influences formal education as it gets
digitized. At the same time, digitization is
making informal and non-formal learning in-
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
15
creasingly important for society, and educa-
tional institutions are looking for new sourc-
es of legitimation. Formal and non-formal
education is increasingly putting learners in
control of their learning, at the same time
this might be seen as burdening individuals
with responsibilities that used to belong to
governments and the state (Biesta, 2015,
p. 76).
Ensuring that education and training sys-
tems in Europe are fit for the digital age is a
European Commission priority, set out in the
Digital Education Action Plan 2021-2027.
The plan is aimed at promoting high quality,
inclusive and accessible digital education in
Europe. The recent publication of the pro-
posals for both the Council Recommenda-
tion on the key enabling factors for success-
ful digital education and training and the
Council Recommendation on improving the
provision of digital skills in education and
training addresses the growing importance
of technology-mediated learning, a process
which was accelerated by the sudden main-
streaming of remote learning during the
COVID-19 pandemic. This unforeseen tran-
sition came with opportunities to modernize
educational practices, but also with an ex-
acerbation of pre-existing socio-economic
gaps, the emergence of new challenges to
equal participation in education and the un-
even distribution of technological resources
(Blaskó et al., 2022, Cachia et al., 2021).
Likewise, emerging technologies such as AI
bring both opportunities and risks, prompt-
ing the need for a better understanding of
the impact of technological innovations
and for improved digital skills among ed-
ucational stakeholders. In this respect, AI
was addressed in the last update of the Eu-
ropean digital competence framework for
citizens: DigComp 2.2 (Vuorikari, 2022). A
wide range of policy developments are rel-
evant to the educational sector, reflecting
Europe’s commitment to enhancing digital
education and skills to nurture a digitally lit-
erate and competitive society. Most notably,
these include A European approach to artifi-
cial intelligence7; the European strategy for
data8 (with the European Data Governance
Act9 and the Data Act10); the European skills
agenda11; the proposal for an Artificial In-
telligence Act, the Digital Services Act, and
Digital Markets Act12; the Ethical guidelines
on the use of artificial intelligence and data
in teaching and learning for educators13; the
Digital Decade policy programme14; as well
as the proposed European digital identity
framework15.
Below we focus on technology, but it is im-
portant to keep in mind the broader soci-
etal changes and historical trajectories that
will keep shaping the education systems of
the future. In this report we show how nov-
el and emerging technologies already are
(and could become even more) important
for learning and education. It is, therefore,
essential to understand these technologies
better in order to assess the extent to which
their use could redefine certain aspects of
education and potentially help to address
well-known existing problems. More inter-
estingly, the emerging technologies dis-
cussed below will reveal their true potential
as tools that will influence the future of ed-
ucation.
7. https://digital-strategy.ec.europa.eu/en/policies/european-ap-
proach-artificial-intelligence
8. https://digital-strategy.ec.europa.eu/en/policies/strategy-data
9. https://digital-strategy.ec.europa.eu/en/policies/data-govern-
ance-act
10. https://digital-strategy.ec.europa.eu/en/policies/data-act
11. https://ec.europa.eu/social/main.jsp?catId=1223&langId=enai
12. https://digital-strategy.ec.europa.eu/en/policies/digital-servic-
es-act-package
13. https://op.europa.eu/en/publication-detail/-/publication/
d81a0d54-5348-11ed-92ed-01aa75ed71a1/language-en
14. https://commission.europa.eu/strategy-and-policy/priori-
ties-2019-2024/europe-fit-digital-age/europes-digital-dec-
ade-digital-targets-2030_eneuropean
15. https://commission.europa.eu/strategy-and-policy/priori-
ties-2019-2024/europe-fit-digital-age/european-digital-identi-
ty_en
16
SCIENCE FOR POLICY REPORT
1.2 Aim of the report
The report looks at emerging trends and
technologies that are already or may in the
future, contribute to the redefinition of ed-
ucational practices, processes, and organi-
zations. The aim of the report is to inform
European policy stakeholders of important
technological innovations and develop-
ments in the context of learning theory, in
order to unpack their potential influence in
teaching, learning and the organization of
education. The scope of the report is limit-
ed to a series of digital technologies ‒ cur-
rently at different stages of development
and adoption ‒ with high potential to be
accompanied by significant societal trans-
formations.
International institutions such as OECD
(2022) and UNESCO (2021) have recently
released reports on the impact of technol-
ogy in the future of education. This report
adds and complements existing work in this
area by reviewing technological develop-
ments such as the emerging next genera-
tion internet, next generation virtual worlds,
and AI technologies – in particular genera-
tive AI – that have the potential to disrupt
the ways in which we understand and or-
ganize education.
Considering the extremely fast pace of
technological innovation and evolution of
the current media and information ecology,
it is of the utmost importance to imagine
a variety of possible alternative futures. In
this regard, as a society we need to ensure
that diverse voices are heard and to make
sure that we go beyond industry-driven per-
spectives on the role that technology could
and should play in education.
1.3 Methodology
There have been many attempts to imagine
speculative scenarios for the future of ed-
ucational institutions, with particular atten-
tion to teaching and learning. Technological
trends, including digitalization, play a central
part in these attempts. Some have relied on
established scenario development methods
(e.g., Baker & Smith, 2019; Facer & Sand-
ford, 2010; OECD, 2020; Pelletier, 2021).
Others have extracted key trends and used
social science and design fiction in specific
technology domains such as AI and robotics
(e.g., Bai, Zawacki-Richter, et al., 2022; Cox,
2021; Selwyn et al., 2020), summarized
the outcomes of expert opinions and stake-
holder consultations (e.g., OECD, 2019a;
Roschelle et al., 2020; UNESCO, 2021;
Vuorikari et al., 2020), or presented more
idiosyncratic visions of the future of educa-
tion and learning based on expected tech-
nological and socio-technical developments
(e.g., Pinkwart, 2016; Schiff, 2021; Tuomi,
2007; Unwin, 2019; Woolf et al., 2013).
Future visions of educational technologies
have also been extensively discussed in the
context of science and technology studies,
oen critically highlighting the assumptions
that underpin these visions (e.g., Selwyn,
2019; Williamson, 2017).
One way to study emerging technological
trends would be to deconstruct the visions
produced in such studies and analyse their
technical requirements. For example, many
visions of future digital education are based
on the idea that instruction can be tailored
and personalised for individual students us-
ing computer-based adaptive learning en-
vironments. Such a vision of personalised
instruction generates clear design require-
ments for intelligent tutoring systems. They
need to have a model of the knowledge
domain, a similar model of student’s knowl-
edge state, a pedagogic model that moves
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
17
the student towards mastery, learning con-
tent that helps the student learn, and user
interfaces that allow the system to inter-
act with the student (Luckin et al., 2016).
For each system component we could ask
whether there are expected technical ad-
vances that could lead to breakthroughs
that could realize the proposed vision. In-
deed, some technology experts have sug-
gested that personalised intelligent tutoring
systems will soon revolutionize education
(e.g., Lee & Qiufan, 2021).
Although this report is informed by these
earlier discussions on the futures of educa-
tion and learning, below we approach the
future from a more idiosyncratic point of
view. There have been few previous stud-
ies that integrate future-oriented techno-
logical analysis with educational theories.
Below we try to do this basing our analysis
on academic and grey literature and earli-
er research by the report authors, with the
aim to highlight those aspects of emerging
technologies that seem particularly rele-
vant for education, learning, and education
policy. In this sense, the present report aims
to go beyond the state-of-the-art and fa-
cilitate richer discussions on the potential
impact of emerging technologies in educa-
tion in order to support long-term strategic
thinking in education. At the same time, we
have tried to validate the relevance of the
report content by involving various experts
during its development. Two workshops
have been organised to discuss the dra
report developed by the first author, and
validate its ideas, structure, and content.
For the workshops, we invited key innova-
tive thinkers from three continents, able to
cover the topics in the report and known for
their high quality of research. The outcomes
of these workshops have been integrated in
our analysis. In addition, we have received
feedback and comments from many educa-
tional experts who did not participate in the
workshops, but whose input is reflected in
the referenced literature.
One starting assumption in this report is
that technological developments will open
up new possibilities with regard to the or-
ganisation and motivation of learning, at
the same time generating new needs and
societal objectives for education. The tech-
nical change that we see possible in the
coming decade touches the foundations of
society, its educational systems, and the
processes of learning in ways that can ap-
propriately be called disruptive. In simple
terms, emerging technologies and associ-
ated social dynamics are expected to re-
configure some aspects of the educational
field. The present report aims to explore this
landscape, highlighting some of its salient
features. As we explore a future that does
not exist yet, we don’t have empirical facts
about it (Tuomi, 2012; 2019). In the spirit
of abductive reasoning that generates hy-
potheses that reveal what evidence would
be important, we simply try and interpret
anticipated technical developments from
the point of view of their implications for
future learning and education. The aim is
not to list all potentially important emerg-
ing technologies; instead, the objective is to
help in locating and prioritizing trends that
should be studied and further researched
in more detail to gather evidence for policy
development.
18
SCIENCE FOR POLICY REPORT
Sixth generation networks (6G) are ex-
pected to link the physical and the virtual
worlds and enable and operate immersive
technology-mediated experiences by the
end of the decade. Although 6G provides
similar wireless communication services as
5G, it also represents a conceptually new
approach to digital networks. This expected
technological change ‒ which we describe
as the ‘Next Internet’ ‒ will have important
social and cognitive consequences. The
reason to highlight 6G and call the result-
ing socio-cognitive and technological in-
frastructure the Next Internet is that it will
potentially generate a disruption that is in
many ways like the one generated by the
internet and the World Wide Web over the
last three decades. The internet disrupted
established forms of human communication
and access to knowledge, impacting social
interaction and cognition. On different lev-
els, the impact of 6G will be similar, or even
more profound. Beyond social relations, 6G
will redefine our relationship with material
reality and its time-space organization. It
is not clear yet how this change should be
conceptualized, but it is clear that it has di-
rect relevance for learning and our current
theories about how learning happens. To
understand this disruption, it is useful to put
the internet in its sociological context. We
will return to this aer first briefly reviewing
the more technical aspects of 6G in the next
section.
2.1 Towards 6G networks
A variety of use cases have been proposed
for the 6G networks. Generic terms such
as virtual worlds and extended reality are
commonly used, although there are many
interpretations of what they would mean in
practice. For 6G, relatively straightforward
extensions of existing 5G use cases have
been proposed. These include, for example,
high quality real-time 3D meetings during
high-speed train travel, remote surgical
operations using virtual reality interaction,
and more exotic applications such as holo-
presence. From the point of view of learning
theories, the core use case for 6G networks
will, however, be ‘digital twins’. In the Next
Internet, these will be digital and actiona-
ble representations of the physical world at
different levels (from the whole Earth to the
human body). Although the concept is not
new, 6G networks enable the synchroniza-
tion of the physical and the digital in quali-
tatively new ways.
In the Next Internet, the various functions
of digital networks will be virtualized and
implemented in a dynamically orchestrated
network, where the control of the network
itself will require very fast communication
capabilities. Instead of centralized cloud
platforms, much of the processing will oc-
cur at the edge of the network, close to the
users, on what is being called ‘edge com-
puting’. This will require dynamic allocation
of data and processing across the network.
The users will experience a very high-speed
and low-latency network, where user devic-
es can rely on many different situation-de-
pendent access technologies. These are ex-
pected to use fast radio technologies, first
at the 7-20GHz range and later towards
millimetre wavelengths and terahertz fre-
quencies. 6G networks will extensively use
machine learning and AI, both in its founda-
tional technologies ‒ for example to beam
directed radio signals from the antennas
to the user location in a three-dimensional
2 The next Internet
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
19
space ‒ and in predicting data needs and
user movement. They will also to switch
user devices across many different types of
network access technologies as the user’s
data and processing requirements change.
In contrast to the 4G and 5G networks,
which rely on cell towers where base sta-
tions connect the wireless network to the
fixed networks and to the internet, 6G will
move beyond the cellular network paradigm
(Giordani et al., 2020). The very high speeds
that the network is expected to provide
require new wireless access technologies
(Polese et al., 2021). These may include
wide-spectrum access points embedded
in homes, offices, and classrooms, optical
communication using LED luminaries, and
low- or zero-power devices that scavenge
energy from available radio signals or other
sources.
The vast increase in short-distance wire-
less access points implies that energy con-
sumption, data security, and privacy are key
design criteria for the 6G networks. Identity
and identification infrastructures, as dis-
cussed later in this report, will therefore be
integral elements in the Next Internet. A
potentially important challenge will be the
development of commercial quantum com-
puting, which is expected to be available
at the beginning of the next decade. This
will make the current public-key encryp-
tion-based privacy and security approaches
vulnerable. Although it is not clear how this
challenge will be solved, semiconductor de-
sign tool providers are already starting to
roll out chip architectures that increase the
complexity of encryption algorithms so that
they would be more resilient to quantum
attacks over their expected lifetimes (Neu-
stadter, 2022).
2.1.1 The emerging infrastructure
of knowing, action, and learning
The material dimension of mediated inter-
actions and information technologies have
traditionally attracted much less attention
than the content of messages. However, the
physical settings in which actors are em-
bedded, as well as the infrastructures avail-
able to them, play a key role in enabling
or otherwise limiting their opportunities to
access and share information. In relation to
education, the term ‘virtual’ ‒ as in Virtual
Learning Environment or Virtual Universi-
ty ‒ somehow implies that digital systems
operate as realms for learning completely
detached from the material world. Educa-
tional researchers writing at different times
have confronted that myth (e.g., Gourlay,
2021; Robins & Webster, 2002) and the
Covid-19 pandemic made more explicit the
impact on education of socio-economic dif-
ferences, as manifested by lack of access to
dedicated spaces for study or the need to
share devices with others.
In the Next Internet, the physical and the
digital will become increasingly intercon-
nected in more profound and complex
ways. Beyond the obvious instrumental
uses of new technical possibilities such as
holographic presence and immersive sim-
ulations, the fusion of digital and physical
creates what could be regarded as a new
infrastructure for knowing and action. Inter-
action with the material environment has
been a critical element in several influential
learning theories.
To understand the difference between 6G
and earlier communication technologies,
it is important to note that spatially and
temporally organized human-to-human in-
teractions provide the foundation for social
life. Recurring interactions make expecta-
tions possible and generate social habits,
routines, and norms. Historically, these in-
20
SCIENCE FOR POLICY REPORT
teractions have required that people are
physically present in a specific location at
the same time. In Durkheim’s (1933) clas-
sical description of the emergence of mod-
ern capitalist world, value-centric medieval
villages transform into global networks of
commerce as money enables transactions
across geographies and cultures. Since
Thorstein Hägerstrand’s quantitative work
in the 1950s on time geographies of social
interaction, sociologists have increasingly
emphasized the role of spatial structures
as enablers and constraints in shaping soci-
eties and their practices (e.g., Crang & Thri,
2000; Hillier & Hanson, 1984; Larsen et al.,
2016). In terms of Castells (1996, pp. 410–
411), ‘space is the expression of society’ and
‘crystallized time’ that provides the material
support for time-sharing social practices. In
a sociological interpretation, physical struc-
tures such as roads, bridges, airports, cities,
and the architectures of dwellings and offic-
es represent the outcomes of a long history
of spatially and temporally organized hu-
man interactions. As such, they can benefit
certain social groups while embodying bias-
es and oppression dynamics against others.
The industrial-age organization of time and
space was disrupted by the emergence of
the internet. Very fast increases in optical
network capacity at the turn of the 1990s
(Hugill, 1999; Tuomi, 2002) created a new
social infrastructure, and wireless networks
have now further expanded outside the
global hubs. The geography of this emerg-
ing ‘cyberspace’ was mapped by Castells in
his landmark studies (Castells, 1989, 1996,
2001), and its characteristics have been
extensively studied by sociologists, philoso-
phers, cultural scholars, and philosophically
oriented computer scientists over the last
decades. Faster 5G networks will further so-
lidify this interaction infrastructure, but 6G
will be different. Whereas Castells focused
on networks and the flows between the
hubs of information, technology, ideas, and
organizational interactions, and their social
consequences, 6G more directly connects
cognition, space, and action. In this sense,
we are moving towards a ‘post-Castellsian’
world.
In learning theories, this linkage between
cognition, space and action is particular-
ly prominent in those constructivist theo-
ries that emphasize practical action as a
source of learning. John Dewey (1991), for
example, argued that learning is rooted in
concrete experience and requires practical
experimentation with the material world.
According to Dewey, learning starts when
our habitual action does not lead to the
anticipated outcomes and action is moved
to an intellectual level. Similarly, Vygotsky
(1986) argued that children learn by inter-
nalizing practical action. A central idea in
Vygotskian learning theories is that action
is mediated by practical tools that reflect
the current historical level of technology,
which can also become instruments of cog-
nition (Bruner, 1986; Cole, 1996; Engeström,
1987; Leont’ev, 1978; Luria & Vygotsky,
1992). In this practical action, the physi-
cal characteristics of concrete objects both
constrain and enable thought and its devel-
opment. The interaction between the ma-
terial and cognitive worlds was the starting
point in Schön’s (1987) analysis of learning
processes, and also in Mead’s (1967) sym-
bolic interactionism.
Whereas the digital world until now has
been mainly a representation and a reflec-
tion of social and material realities, human
action and interaction are now becoming
mediated by a real-time digital layer. Those
individualistic variants of constructivism
that understand knowledge development
as the construction of mental or cognitive
models of objective reality, assume that the
learner is an observer of the world. These
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
21
are oen inspired by the studies of Piaget,
although Piaget himself argued that cog-
nition and reality are mutually constructed
(Furth, 1981). In social variants of construc-
tivism, the learner, in turn, is a participant
in a process of collective knowledge crea-
tion and building where systems of know-
ing are understood as fundamentally cul-
tural products. For the Dewey, Piaget, and
Vygotskian cultural-historical variants of
constructivism, practical interaction with
the physical world is the fundamental start-
ing point for learning as well the source of
meaning. Under this new perspective of
framing learning, pedagogical approaches
such as situated learning (Brown, Collins &
Duguid, 1989) and problem-based learning
(Barrows & Tamblyn, 1980) have emerged,
prompting educators to shi their teaching
to make learning more relevant for students
by creating learning in realistic or simulated
environments (Herrington, Reeves, & Oli-
ver (2007). While the social transformation
discussed by Castells focused on the new
time-space organization of social interac-
tions, in the context of such action-oriented
theories of knowledge creation and learn-
ing, the technical characteristics of 6G net-
works imply a reorganization at the deeper
level of human cognition and action.
The ‘sensorization’ of the network means
that spatiality and body become important
elements in the digital future. This opens
up, for example, new lines of research on
extended (Gibson, 1977; Vygotsky & Luria,
1994), embodied (Dourish, 2001a; Varela
et al., 1991), situated (Suchman, 1987),
and distributed (Hutchins, 1995; Pea, 1985;
Salomon, 1993) cognition, with potential
implications for theories of learning and
pedagogic practice.
2.2 Immersive
technologies: Extended
reality, holograms and
virtual worlds
Extended Reality (XR) is a term commonly
used to describe a set of technologies that
augment or extend human perception of
the real world. These technologies include
Augmented Reality (AR), Virtual Reality (VR),
and Mixed Reality (MR) (Chang et al., 2022;
Kaplan et al., 2021). Although the idea of
using extended reality and immersive virtu-
al worlds for learning and education is an
old one,16 and their use in education has
been explored from different aspects (Vour-
ikari et al., 2020), technical requirements
for XR have limited their development. Fast
networks, such as 5G and Wi-Fi 6 are now
unlocking some key constraints, and the
emerging Next Internet, combined with
AI-supported content production tools, can
also be expected to lead to expansion of XR
in education.
VR is a fully artificial environment that a
user can interact with. Early examples of VR
include virtual game worlds and platforms
such as Second Life and, more recently,
Minecra and Roblox, but oen the term is
understood to imply immersive experience
that requires specialized VR headsets. Aug-
mented Reality, in contrast, involves super-
imposing digital content onto the real world.
Mixed Reality, in turn, synchronizes the dig-
ital and physical worlds, and combines ac-
tionable computer-generated objects in an
augmented world. With fast networks and
processing, these actionable digital objects
can be directly linked with physical arte-
facts, blurring the boundaries between the
representation and the represented.
16. Some examples were discussed, e.g., in Tuomi (2005,
Chapter 6).
22
SCIENCE FOR POLICY REPORT
Over the last decade, there have been great
expectations about the transformative im-
pact of XR. For example, Meta has been
spending about $10-12 billion annually on
its XR initiatives, and about $50 billion by
the end of 2022, against $6 billion in reve-
nue since 2019 (Ball, 2023). The global AR
market was estimated to be $25 billion in
2021, with a compound annual growth rate
of over 40 per cent from 2022 to 2030.17
The developers of XR technology have,
however, realized that immersive XR has
surprisingly demanding technical require-
ments that may delay the wide adoption
of XR for several years, particularly for con-
sumer use.18
AR can be experienced through a smart-
phone camera that provides a ‘magic win-
dow’ or ‘portal’ to the augmented reality or
using a specialized AR headset for a more
immersive experience. Detached display
devices, such as mobile phones, can use
their sensors to track the device pose us-
ing three degrees of freedom and provide a
monoscope simulation of a 3D world. Low-
cost stereoscopic implementations have
been built using ‘cardboard boxes’ that im-
itate the classic View-Master device with
a standard mobile phone. This approach is
used, for example, in Google’s Cardboard.19
In contrast to XR cardboard boxes that can
be purchased for a few euros, high-quality
head mounted devices adjust the perceived
world in six degrees of freedom20, adapting
both to the user’s movements and to rota-
17. https://www.grandviewresearch.com/industry-analysis/aug-
mented-reality-market
18. These include very high-density displays, processing power
that is needed to manage XR applications and user and envi-
ronment monitoring, as well as the need for light-weight power
sources. It is commonly expected, for example, that VR headsets
may require 12 cameras to track the user’s eyes, fingers, and
location. Related technical and economic trends and remaining
challenges are analyzed in detail in Ball (2023).
19. https://arvr.google.com/cardboard/get-cardboard/
20. This refers to the specific number of axes that body can
move in a three-dimensional space.
tional head movements for an immersive
experience with three-dimensional sur-
round sound. It can be expected that head-
set tracking will be increasingly common
in the future when low-cost AI processors
can be used to map the environment. With
these advanced technologies, not only are
user movements tracked by fixed ‘base sta-
tions’, but ‘inside-out-tracking’ is also possi-
ble, where the headset itself tracks the user.
One of the areas where XR has been suc-
cessfully used is in applications that allow
students to explore human and animal
anatomy and practice surgical procedures.
XR has also been used to enable immersive
visits to historical periods and outer space,
for interactive language learning, and the
visualization of mathematical functions
(XRA, 2022). XR simulations have also been
applied, for example, for employee train-
ing for dangerous work tasks in oil rigs and
supermarkets, and for training professional
athletes (Bailenson, 2018).
XR is also used in many industrial and man-
ufacturing settings. In the health sector, XR
provides a training environment that adds
value when compared to observation-based
ways of acquiring procedural knowledge.
For example, in 2020, surgeons at the
John Hopkins University School of Medicine
performed spinal fusion for a patient and
removed a tumour from the spine of an-
other patient using AR. In these operations
the headset overlaid the patient’s internal
anatomy (such as bones and other tissue),
recorded using CT scans with a see-through
image.21 Adoption of XR in health education
can have a significant impact on the qual-
ity of education in providing simulated en-
vironments for training and practicing, skill
assessment, and procedural knowledge ac-
quisition (Iop, El-Hajj, Gharios et al., 2022).
21. https://www.hopkinsmedicine.org/news/articles/johns-hop-
kins-performs-its-first-augmented-reality-surgeries-in-patients
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
23
In more consumer-oriented applications, XR
is also becoming widely used in gaming and
entertainment, e-commerce, and retail.
Although consumer headsets have main-
ly been marketed for VR gaming, it is ex-
pected that adding video feed-through
will make them usable also for AR and
MR.22 High-quality headsets are still rela-
tively expensive, with a typical consumer
setup costing several hundred euros and
a high-performance PC or a game station.
More importantly, it has been expensive to
develop content for XR. It can be expected
that generative AI tools will soon disrupt the
economics of XR content creation, including
for educational applications.
The size, weight, and cost of XR headsets
will most likely diminish quickly in the com-
ing years. Beyond the bulkier headsets,
there have been several attempts to com-
mercialize light-weight AR eyeglasses. The
first wireless AR contact lens was demon-
strated in 2022, although its development
was discontinued soon aer.23 The state-of-
the-art now includes light-weight stereo-
scopic full-colour augmented reality glass-
es that use micro-LED optical waveguide
displays. State-of-the-art glasses support,
for example, GPS-based mapping of the
environment, gesture recognition, and re-
al-time translation when speaking face-to-
face by showing translated subtitles on its
screen.24
A variant of immersive technologies is the
hologram. Three-dimensional holograms
have been proposed for educational uses
22. https://www.linkedin.com/pulse/23-augmented-reality-
trends-keep-eye-2023-tom-emrich
23. Due to tight capital markets in 2022 and the ‘yet-to-be
proven market potential’, the developer of the AI contact lens,
Mojo Vision, decided in 2023 to discontinue its work on this tech-
nology and focus on Micro-LED technology https://www.mojo.
vision/news/a-new-direction
24. https://www.tcl.com/global/en/news/tcl-unveils-groundbreak-
ing-augmented-reality-glasses-at-ces-2023
at least since the early 1990s (e.g., Ghu-
loum, 2010; Yoo et al., 2022). In recent
years, holograms have been used to put
deceased superstars on stage, including
Whitney Houston, Maria Callas, Buddy Hol-
ly, Ronnie James Dio, and Glen Gould. Al-
though the technology used has oen been
based on image projection, there have also
been experiments with real 3D holograms.
A well-known early example was the ho-
lographic news report in CNN’s 2008 Wolf
Blitzer -anchored news program, where a
reporter was ‘beamed in’ to the studio floor
to comment on the results of the U.S. pres-
idential elections. Commercial holography
devices are now marketed as replacements
for online meetings, for e-commerce, hospi-
tality, and for education. These include the
life-size communication platform offered by
Proto Inc.25 and Solid Light technology from
Light Field Lab.26 The latter uses a similar
electromagnetic beaming approach to 6G
antennas.
The impact of VR and, more generally XR
in educational settings has been stud-
ied in many small-scale research projects
(e.g., Pellas et al., 2021; Yiannoutsou et al.,
2021), with various studies conducted in
the medical field, but the rapid advances
in technology make generalizations about
learning outcomes still difficult. Another
limitation is the geographical distribution of
such advanced technology. Most research in
this field, for instance in medical education,
is conducted in Canada and USA, where
most XR systems employed in the studies
are devised (Iop et al, 2022). Bailenson
(2018) has suggested that XR applica-
tions can be particularly valuable in rela-
tion to learning situations characterized by
the DICE principle (Dangerous, Impossible,
Counterproductive, Expensive). According to
25. https://protohologram.com/about/
26. https://www.lightfieldlab.com/
24
SCIENCE FOR POLICY REPORT
this, XR could replace experiences that are
difficult, impossible, or non-desirable in the
real world. In educational settings such ex-
periences are oen excluded by design. This
suggests that XR could expand the space
of learning, instead of replacing already ex-
isting learning activities. Based on a review
of literature on neurosurgical education, Iop
et al., 2022 highlight the relevance of using
XR systems for skills assessment, where-
by a system could be developed to enable
self-assessment or evaluation by experi-
ence neurosurgeons. Likewise, those forms
of simulations will be particularly relevant
to the vocational education and training
(VET) sector, particularly work-based learn-
ing. For example, by having a simulated
environment to learn how to operate dan-
gerous or hazardous machines, as close as
reality as possible but without real safety
risks. Such innovations are part of the main
focus of a council recommendation on im-
provement of the provision of digital skills
to the Member States that was published in
April 2023 (European Commission, 2023b).
On the other hand, it is important to bear
in mind that the simulation of dangerous,
impossible, and counterproductive experi-
ences can also be psychologically harmful,
in particular, for children and adolescents.
For example, it may be educational to have
immersive experience of learning history
by burning Rome as Nero or to understand
the impact of schoolyard bullying through a
XR game, but this also raises many ethical
questions. The ethics of XR use will, there-
fore, be an important topic also in educa-
tional settings. XR devices and applications
will create vast amounts of fine-grained
data on user behaviour and environment,
and there are already indications that pri-
vacy and the ethics of educational XR need
to be addressed soon (e.g., Christopoulos et
al., 2021; Steele et al., 2020).
Virtual worlds have oen been used to re-
fer to an immersive virtual place and the
next frontier for gaming, social media, and
advertising (Wunderman Thompson, 2022).
In the context of futures of learning, virtu-
al worlds are perhaps best understood as a
form of mixed reality. As discussed above,
its true impact will be seen when the physi-
cal, social, and digital worlds become inter-
linked in real time. Although there has been
much hype around the term, this integration
of digital with the social and physical gen-
erates some new foundations for the social
world. Existing social, economic, or cognitive
theories have little to say about develop-
ments in this emerging universe. However,
the emergence of new generation virtual
worlds and so-called metaverses comes
with important societal implications, op-
portunities, and challenges (Hupont et al.,
2023).
2.3 Digital credentials and
identifiers
Credentialing and certification are impor-
tant social functions of education (Biesta,
2010). In many professions, credentials
and educational certificates are required
for entry to an occupation or for performing
specific job tasks. Such credentials are of-
ten important constituents of systemic trust
which, for example, allow us to expect that
people in the cockpit know how to fly the
plane, or that an electrician can connect an
electric oven safely. In other words, paper
qualifications have traditionally worked as
a proof that someone has the competences
defined by the intended learning outcomes
established in the curriculum of a given ed-
ucational programme or level.
Digital technology is reconfiguring the way
educational institutions issue credentials in
diverse ways. For example, by supplement-
ing or replacing paper certificates with elec-
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
25
tronic versions, verification becomes easier
to key stakeholders such as employers or
other educational institutions for further
studies. At the same time, while creden-
tials were traditionally issued to students
only aer completing an entire academic
programme, there is now a growing level of
granularity in certification practices. In this
regard, micro-credentials (e.g. open badg-
es) have proliferated as a way of evidenc-
ing skills, the completion of short courses or
modules, as well as other achievements in
formal and non-formal learning, especially
in higher education and VET.
Partly because of the rapid technical
change and the possibility to generate skill
definitions from online job announcements
using natural language processing, it is now
common to generate increasingly detailed
lists of skills that could be verified and val-
idated. Micro-credentials have emerged as
a natural answer to the need to manage in-
formation about such detailed skills. In the-
ory, such skill lists can provide useful views
on labour market skill gaps and information
on skill development needs, albeit inherent
to some biases towards more professional
occupations (Sostero & Fernández-Macias,
2021). Such skill lists and taxonomies are
now widely used around the world for la-
bour market analysis, educational guidance,
and in educational planning.27
2.3.1 Micro-credentials
In recent years, micro-credentials have been
widely implemented for many different pur-
poses, with the support of governments
and intergovernmental organization (OECD,
2023). The interest in micro-credentials at
least partly reflects ongoing technological
27. This approach has been used, for example, in Cedefop’s
Skills-OVATE system that analyzes online job advertisements.
Skills-OVATE is now jointly developed by Cedefop and Eurostat
as part of the Web Intelligence Hub. A detailed review on exist-
ing initiatives in this area can be found (in Finnish) in Tuomi et
al. (2021).
change that has reduced the labour-mar-
ket relevance of traditional educational
degrees (European Commission, 2021c;
OECD, 2021; Oliver, 2022; van der Hijden
& Martin, 2023). Micro-credentials enable
flexible learning paths and the recognition
of prior learning, at the same time facilitat-
ing broad access to short courses that can
close labour market skill gaps (Council of
the European Union, 2022a). They can be
beneficial as standalone certifications, or
as a way to complement other established
programmes for greater employability, but
also to improve lifelong learning (Orr et al.,
2020).
Many digital-era skills are related to new
tools and technologies that oen have short
lifetimes or that require people to be able
and ready to upskill or reskill at a fast pace.
This has led to an interest in packaging
training and education in short courses and
personalised segments of ‘microlearning’
that could be accumulated to cover wid-
er areas of expertise. The premise is that
these alternative credentials could then be
used to address labour market needs, and
to help individuals in forming meaningful
lifelong learning paths.
Micro-credentials can record such learning
paths for three different purposes. Many
current ‘micro-credentials’ address regula-
tory requirements. For example, a ‘hygiene
passport’ may be required from restaurant
workers that certify that they are able to
handle food items. A plumber may not be
allowed to solder copper pipes without a
firework permit. Similarly, to legally drive a
car on a public road requires the driver to
have a driver’s license that certifies that the
driver knows traffic signs and rules and has
verified skills in handling the car in practical
settings. Such micro-credentials are essen-
tially permits. They have high labour market
value as they open access to restricted jobs.
26
SCIENCE FOR POLICY REPORT
Micro-credentials are also widely used as
a signalling mechanism in the labour mar-
ket. For example, many technology provid-
ers offer certificates that aim to prove that
a person is familiar with that technology.
Signals relevant for the labour market tend
to be both domain and technology specif-
ic. They can, for instance, show that a per-
son knows how to form SQL queries, is able
to waterproof a renovated bathroom with
the products of a specific vendor, or knows
how to operate a particular machine. While
there is still limited evidence on the per-
ceived value of micro-credentials, they play
an increasingly relevant role in the context
of growing competency-based recruitment
and the challenges of traditional talent de-
velopment (Gauthier, 2020; Hollands, 2023,
OECD, 2023). Micro-credentials are oen
viewed as a tool to address labour-market
skill gaps and as a mechanism that could
guide job seekers to complementary train-
ing and education that improves their em-
ployability, as they can help employers in
the process of filtering and sorting potential
applicants.
Micro-credentials are also used in social
representation, self-reflection, and profes-
sional identity construction. Open badg-
es, for example, are a means of building
one’s online identity. They are oen liter-
ally badges and emblems that showcase
achievements and signal membership of
professional associations. In lifelong learn-
ing, micro-credentials can represent pro-
gress towards self-imposed objectives. The
organizers of massive open online courses
(MOOCs), for example, oen make effort
visible by providing a badge for successful
participants.
As micro-credentials are assumed to reflect
knowledge, skills, and competences, the dy-
namics of competences underpinning these
three types of micro-credentials suggest
that their implementation may need dif-
ferent infrastructures that can also ensure
quality assurance and standardisation in the
way micro-credentials are recognised. This
remains a complex issue because there are
still limited digital solutions that are able to
validate, recognise and store micro-creden-
tials. In addition, some technical skills are
directly associated with specific technolo-
gies and tend to become obsolete fast and
the content of skills changes rapidly.
The fast-paced obsolescence of skill defini-
tions could pose a significant challenge to
human resource departments. Large organ-
izations are particularly vulnerable to this
issue, as they may well keep skill terms in
their competence-management systems
for decades, even when their content sub-
stantially changes. In general, the appar-
ent stability of skill definitions in many skill
taxonomies can be viewed as an artefact
generated by the taxonomy itself (Bowker
& Star, 1999). This decoupling of skill defi-
nitions from actual work tasks may limit
the value of micro-credentials in certifying
labour-market relevant skills. However, one
of the benefits of digital credentials is that
they may include examples of the work that
the holder submitted for assessment before
earning.
Online job advertisements oen include long
lists of expected skills and accumulated ex-
perience that are used to generate skill tax-
onomies. Skill definitions extracted from on-
line job advertisements may easily become
reified. Such skills definitions might also be
used as the basis for micro-credentialing.
It is not clear how important micro-creden-
tials would be in the actual hiring process.
However, it is widely known that machine
learning systems are already extensively
being used to sort applicants, and that HR
professionals believe that AI will radically
change recruiting processes (Fatemi, 2019;
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
27
Hunkenschroer & Luetge, 2022). Moreover,
further research might be needed to explore
the extent to which the description of tasks
and skills in job descriptions reflect the real
needs of employers.
As the development and use of machine
learning systems requires data, it will be
natural to think that detailed data on cer-
tified knowledge and skills would be use-
ful and necessary. The epistemological
assumptions that underpin such thinking,
however, are complex and it is important to
make them explicit.28 The most requested
competences in job advertisements include
‘so skills’ such as ‘team working’, ‘prob-
lem-solving’, and ‘communication’ skills
that are very different from domain specif-
ic technical skills and knowledge typically
learned at educational institutions.
2.3.2 Verifiable credentials
According to the World Wide Web Consor-
tium (W3C), a verifiable claim is a qualifi-
cation, achievement, quality, or piece of
information about an entity’s background
such as a name, government ID, payment
provider, home address, or university de-
gree (Sporny et al., 2022). The W3C Verifi-
able Credentials standard defines methods
to issue, assert, verify, store and move, re-
trieve and revoke claims. In their use cases
(Otto et al., 2019), W3C highlights four use
cases for education: digital transcripts of
student grades and competences issued by
an educational institution; proof of student
identity for high-stakes tests; student-con-
trolled storing of accumulated credentials;
and identification of student identity in on-
line learning systems such as massive open
online courses (MOOCs).
28. These epistemological assumptions have been extensively
discussed in organizational knowledge management and knowl-
edge creation literature (Brown & Duguid, 2001; Nonaka, 1994;
Tuomi, 1999; von Krogh et al., 2000).
Over the last decade, blockchain has be-
come an established technology in differ-
ent sectors, prompting interest also for its
possible applications in education (Bosch
et al., 2022; Grech et al., 2022; Smolenski,
2021; Camilleri et al., 2017). Although from
a technical point of view blockchains are
just a way to implement distributed digital
ledgers using public-key encryption, they
are oen claimed to have the potential to
revolutionize learning and education. Ac-
cording to such views, they could transform
individual skills development and life-long
learning, provide new models for funding
educational institutions, and, for example,
allow billions of potential students in less
developed countries to gain low-cost ac-
cess to learning (Tapscott & Kaplan, 2019).
Nevertheless, it can be argued that the ed-
ucation sector has not found blockchain as
a useful response to their most pressing
needs and there are indeed several chal-
lenges to its adoption beyond pilots; includ-
ing philosophical, practical and legal chal-
lenges (Park, 2021; Steiu, 2020).
The idea that blockchains could in the fu-
ture be used to register micro-credentials
is conceptually founded on the view that
reliable information about increasingly spe-
cific skills can be upgraded and presented,
for example, to future employers. The link
of cultural factors to blockchain technology
with concepts such as ‘democracy’, ‘lack of
central authorities’, ‘data self-sovereignty’,
‘decentralised autonomous organizations’,
and the idea that individuals could con-
trol their identities, is a partial explanation
why such technologies have such a strong
presence in the public discourse, especially
in education. From a sociological point of
view, the proposed models of democracy,
authority, and identity deserve further elab-
oration. From an educational point of view,
the technical and standardization work on
verifiable credentials (VC) and decentralised
28
SCIENCE FOR POLICY REPORT
identifiers (DIDs) will, however, be important
in the coming years.
The EU has been among the leading devel-
opers of large-scale infrastructures in this
area, with the European Blockchain Services
Infrastructure (EBSI) and the Regulation on
electronic identification and trust services29
for electronic transactions in the internal
market (eIDAS). EBSI is also used by the
Digital Credentials for Europe (DC4EU) pi-
lot30 for the European digital wallet, which
‘will be available to EU citizens, residents,
and businesses who want to identify them-
selves or provide confirmation of certain
personal information’.31 The digital wallet
will, for example, support electronic attes-
tations of attributes such as the existence
of educational diplomas and profession-
al certificates. The European Digital Cre-
dentials for learning (EDC) infrastructure,32
which underpins Europass,33 already offers a
standard for tamper-evident electronic doc-
uments that allows providers of credentials
to describe and show the achieved learning
outcomes (knowledge, skills) of learners.
Although blockchains have been claimed to
be extremely relevant to education, the full
educational potential of blockchain tech-
nology remains to be fully realized. There
have been attempts to build new block-
chain-based ‘educational ecosystems’ and
‘blockchain universities’ but so far these
initiatives have not been successful. Typi-
cal expected benefits in research literature
on educational uses of blockchains include
improved security, better control of data
access, enhanced accountability and trans-
29. http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=uriserv:O-
J.L_.2014.257.01.0073.01.ENG
30. https://www.dc4eu.eu/
31. https://commission.europa.eu/strategy-and-policy/priori-
ties-2019-2024/europe-fit-digital-age/european-digital-identi-
ty_en
32. https://year-of-skills.europa.eu/news/new-version-eu-
ropean-digital-credential-learning-infrastructure-re-
leased-2023-07-26_en
33. https://europa.eu/europass
parency, enhanced trust, student authenti-
cation, and cost savings (Alammary et al.,
2019). At present, the most frequent appli-
cation types are certification and validation
processes and authentication of self-sover-
eign digital ID services (Bosch et al., 2022).
Several common use-case scenarios, how-
ever, have been proposed. Crech et al. (2022;
2017), for example, list eight scenarios for
blockchains in education. These include the
permanent secure storage of education-
al certificates, automatic recognition and
transfer of credits, student identification,
and verification of multi-step accreditation.
Smolenski (2021) suggests that the main
impact of blockchain technology will be as
a new credentialing system, with the poten-
tial to eliminate records fraud, streamline
and reduce the costs of record sharing and
verification, and reducing institutional risk
by returning the control of personal data
to individuals. While relevant to many ad-
ministrative processes central to education,
those developments seem to be rather far
from the main pedagogical aspects that
shape teaching, learning and even assess-
ment. In this regard, it is not clear whether
blockchain is the most suitable way of ad-
dressing the needs of educational institu-
tions, although its potential value in relation
to interoperability, mobility or security, to
mention a few aspects, cannot be ignored.
At the same time, it would be irresponsible
to ignore the heavy ecological footprint as-
sociated with blockchain (Schinckus, 2021).
The possibility to use blockchains to store
fine-grained information on micro-creden-
tials may allow individuals to own and con-
trol their personal learning histories. How-
ever, the impact of these technologies is
still limited and is not necessarily the most
optimal way for addressing the needs of
the educational sector. It is important that
such technology innovations are grounded
in theories of learning and skill develop-
ment. As learning is fundamentally about
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
29
change, the dynamics of micro-credentials
is a highly important topic in this learn-
ing-theoretical context.
In many blockchain-oriented visions of the
future of education, technology is under-
stood to enable fine-grained management
of skills and related certificates. In the con-
text of learning theories, such an atomistic
model of skills and competences looks in-
adequately developed. As there are many
ongoing national, EU-level, and interna-
tional initiatives in this area, it would be
important to gain better understanding of
the potential of micro-credentials in certi-
fying competences and learning in differ-
ent application areas and also to identify
which skills lend themselves better to their
use. Also, a good understanding of the data
infrastructures that are needed to support
such developments is important. For such
an understanding we must clarify the na-
ture of various types of skills and perhaps
also anticipate the impact of emerging
technologies on their development. We will
return to this topic in the context of AI-ena-
bled learning later in this report.
2.3.3 Decentralised identifiers
While it is easy to understand potential uses
of verified credentials in education, perhaps
a more fundamental and disruptive speci-
fication has been developed for decentral-
ised identifiers, or DIDs.
DIDs generalize the core idea of the World-
Wide Web. The current web relies on uni-
versal resource identifiers (URIs) that point
to documents and computational resources.
A decentralised identifier, in contrast, can
point to anything, including persons, organ-
izations, material objects, or concepts. For
each DID, there is an associated ‘DID docu-
ment’ that stores keys and data that can be
used to identify the ‘subject’ referenced by
the DID, as well as verify any claims about
it. DIDs are ‘decentralised’ because the DID
documents are stored in trusted distribut-
ed data storages, for example using block-
chains.
DIDs are a core technology of the Next In-
ternet. They will fuse cognitive, social, and
material dimensions in a homogeneous dig-
ital universe. As noted above, this will make
the digital realm actionable in fundamen-
tally novel ways. Although this emerging
universe is oen called metaverse or virtual
worlds, suggesting that it would be a par-
allel description of our everyday reality, the
combination of Internet of Things (IoT) and
the social world will be something that so-
ciological or educational thinkers have not
seen before. Although it is easy to imagine
ways in which immersive social worlds can
be used in existing practices of instruction,
more philosophical and conceptual ques-
tions like what we actually mean by know-
ing and learning in such environments, will
become relevant (Dourish, 2001a; Furth,
1981; Nonaka et al., 2008; Piaget, 1970;
Rosen, 1998). A practical consequence is
that as their epistemological foundations
are reconsidered, many concepts that un-
derpin education may, at least partially, be-
come obsolete.
It is important to note that the emerging
technologies are not just ‘applied’ to solve
existing problems or to address expected
challenges. There are good reasons to ex-
pect that these technologies change the
foundations of societies, including the ways
in which they produce and use knowledge.
This is nothing new. Leading social theorists
have highlighted the importance of new
communication technologies, such as the
printing press (Habermas, 1989; McLuhan,
1962), information technologies (Giddens,
1984; Thompson, 1995; Webster, 1995),
and communication networks (Castells,
1996) for social life. The linking of the ma-
terial, social and cognitive realities through
30
SCIENCE FOR POLICY REPORT
data can also be expected to require new
concepts for understanding how policies
should be developed for this emerging
world. Whereas many of the earlier classi-
cal discussions focused on social interac-
tion and communication, DIDs go beyond
communication by linking material objects
to this emerging network of interaction.
‘Computational embodiment’, in this sense,
means ‘a presence and participation in the
world, real-time and real-space, here and
now’ (Dourish, 2001b).
While the above considerations may lead
to conceptual discussions that may be un-
familiar to some technology developers,
blockchains also have some more imme-
diate technical and regulatory challenges
that will have implications for their use. For
example, core blockchain design objectives
challenge key assumptions of European
data protection law, such as data minimiza-
tion, purpose limitation, and right to be for-
gotten (Finck, 2019). It is not clear whether
certificates that are stored in blockchains
should be considered personal data. It is
also not obvious how the requirement that
under specific circumstances data must be
amended or erased could be interpreted
in the context of blockchains, where such
changes are impossible by design. As one
of the claimed benefits of blockchains is
that they cannot be controlled by any single
actor, it is difficult to find a data controller
as required by the GDPR. Due to their dis-
tributed nature, the updating of blockchains
also becomes very slow if the number of
transactions to be recorded grows. Also,
the exorbitant amounts of energy needed
to keep the distributed trust mechanisms
operating is a well-known challenge for the
green transition.34 Beyond the more foun-
dational conceptual problems related to the
potential benefits of blockchains, micro-cre-
dentials, and decentralised identifiers, there
are also very concrete engineering prob-
lems that may limit the use of these tech-
nologies.
One such technical problem is the possibili-
ty that blockchain technologies will become
obsolete in the coming decade as quantum
computing becomes able to disrupt the
trust infrastructure that makes blockchains
possible.
2.4 Post-quantum
cryptography and zero-
knowledge proofs
Over the next decade, blockchain technol-
ogies need to implement post-quantum
cryptography (PQC). Existing blockchain ar-
chitectures rely on public-key cryptography
that is expected to become obsolete in the
next 10-15 years. At present, it is not known
how blockchain-based trust platforms will
adapt to the post-quantum era, or how the
existing transactions and contracts recorded
in them can remain immutable. PQC meth-
ods and communication protocols such as
Quantum secure layer (QSL) are, therefore,
actively studied at present.35
To some extent, these challenges are en-
gineering problems, and problems with the
interpretation and adaptation of law, but
34. The Cambridge Bitcoin Electricity Consumption index
estimates the annual Bitcoin power demand to be at 115 TWh
and at 13 GW at the time of writing: https://ccaf.io/cbeci/index.
The upcoming proof-of-stake trust mechanism planned for the
Ethereum will consume much less energy. The Bitcoin block-
chain can make about four transactions per second, using 707
kWh per transaction This is about 11 times that of the current
Ethereum.
35. See, for example https://www.nist.gov/news-events/
news/2020/07/nists-post-quantum-cryptography-program-en-
ters-selection-round and (Beullens et al., 2021).
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
31
they also suggest that some earlier op-
timistic visions of the forthcoming block-
chain revolution may lack important detail.
In general, it will be important to integrate
sociological, legal, ethical, developmental,
and cybersecurity knowledge in assessing
the impact of technologies such as verified
credentials and decentralized identifiers in
the educational domain.
If educational certificates are stored in a
permissioned blockchain, where the ac-
cess can be linked with separate contrac-
tual terms of use, ‘zero-knowledge proofs’
could address some regulatory and pri-
vacy-related challenges. Zero-knowledge
proofs (ZKPs) enable data-minimization
and selective disclosure (Goldwasser et al.,
1989). Zero-knowledge methods allow the
data subject to choose which attributes to
reveal and which attributes to withhold on
a case-by-case basis. The approach uses
predicate proofs that allow the verifier to
ask true-or-false questions, for example
whether the person is over the age of 18 or
whether the person has a given profession-
al certificate. In the context of the EU dig-
ital wallet (European Commission, 2023c),
such zero-knowledge proofs are known as
‘attestable attributes’. At present, there are
only relatively generic use cases developed
in this area, including the possibility to up-
load digital diploma attestations and pro-
fessional qualifications in a personal wallet.
Whether a similar approach could be used
for micro-credentials remains a topic for
further research and policy development.
32
SCIENCE FOR POLICY REPORT
Specifications such as DIDs will enable the
Next Internet, but perhaps the most impor-
tant technology for education in the next
years will be data. One of the most impor-
tant consequences of the increasing perva-
siveness of computerization and informat-
ics has been an exponential growth in the
amount of data generated as a by-product
of our daily activities. As we engage with
online services and apps, whether actively
or simply by carrying with us our smart-
phones, we leave behind valuable traces
of data that are the basis of the current
economic system (Cohen, 2019; Zuboff,
2019). Moreover, as we transit through our
increasingly monitored surroundings, large
amounts of data are also generated upon
us, and self-tracking enabled by wearable
devices has become a common activity
(Lupton et al., 2018). Data is the corner-
stone of the current media and information
ecology and the amount of data generated,
stored, and processed is expected to keep
growing at an even faster pace, partly due
to the growth of IoT.
Data are used to train data-driven AI, for
learning analytics, and for research on
learning. Many arising concerns in educa-
tional settings are related to the production,
use, and dissemination of data (Day, 2021;
Livingstone & Pothong, 2022; William-
son, 2017). We are generating enormous
amounts of data while using digital tech-
nologies.
Privacy and ethics of data use are key po-
litical priorities in Europe. Moreover, access
to high-quality data on learning and edu-
cation will be a key factor for research and
development of AI applications in and for
education (McNamara, 2023).
The EU is a global front-runner in data reg-
ulation and governance, as shown by the
General Data Protection Regulation (GDPR)
and the ambitious European Strategy for
Data that includes the already adopted
Data Governance Act aimed at improving
trust between different actors in the data
economy and the proposed Data Act aimed
at improving the fairness of the data shar-
ing. Large-scale data on education and
learning will be at the core of the knowledge
society transformation. Such data extend
beyond records on learning outcomes, cre-
dentials, and statistical data. Digital learn-
ing environments also enable the collection
of data on learning itself. This is something
that has been impossible until recently, ex-
cept in small-scale research. Many educa-
tional collaborative platforms and learning
environments used in Europe belong to
non-European companies and this cannot
be ignored.
In recent years, detailed data on learning
processes has increasingly been accumu-
lated by the providers of digital learning
platforms. Much debate has therefore con-
centrated on potential problems associated
with the uncontrolled flows of data to com-
mercial actors, and the more conceptual
need to critically assess the benefits and
harms of ‘datafication’ of education (e.g.,
Lupton & Williamson, 2017). Among other
things, this has generated debates on the
role of commercial interests in the develop-
ment of educational practices and policies
(e.g., Perrotta et al., 2021; Perrotta & Sel-
wyn, 2020). Datafication and commercial-
ization of education have been widely dis-
cussed in recent years but deserve further
attention from all key stakeholders in the
sector in the light of new regulatory frame-
works.
3 Data
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
33
3.1 Implications of data
spaces for education and
learning
Datafication can, however, also be viewed
as an engineering or an industry policy
problem. If large-scale detailed data on
learning processes and outcomes is im-
portant for the development of new peda-
gogical approaches, learning technologies,
and education policy, we need to ask what
data are needed for this, and how these
data can be made available. Much of these
data already flow on the global information
networks, and the emerging digital technol-
ogies will rapidly increase the amount of
data that could be used to develop educa-
tion and educational services. As a result,
there are major economic and policy inter-
ests in defining access to data on learning.
Regulatory approaches that focus on data
privacy only partially address the emerging
challenges and opportunities in this area.
Re-use and secondary uses of data gen-
erated on digital learning platforms will,
therefore, be an important topic for policy-
makers in the coming years. In Europe, there
has been various efforts to ensure data
sharing, in particular the Data Act and the
Data Governance Act, both aimed at estab-
lishing a single market for data. To achieve
this vision, the focus cannot solely be on
technical aspects, because other factors
such as trustworthiness and effective data
governance are also crucial (Farell, E., et al,
2023).
From a technical point of view, enabling and
regulating access to learning data requires
the definition of information architectures
that address such concerns. Current learn-
ing analytics specifications, such as xAPI36
and Caliper Analytics,37 define some basic
36. https://www.adlnet.gov/projects/xapi/#resources
37. https://www.imsglobal.org/spec/caliper/v1p2
data structures that can be used to record
learner activity on learning platforms. These
architectures, however, need to be comple-
mented by knowledge about learning, so
that data that are collected will be useful
and relevant for learning and education.
Data spaces for learning and education,
therefore, are conceptually different from
European data spaces that are currently
being developed for skills data.38 Learning
data consist to a large extent of trace data
that records learning processes instead of
outcomes. Although the EU Data Space for
Skills (DS4Skills) project is at present de-
veloping conceptual approaches and some
use cases for learning-related data, it will
only address topics that are important for
learning analytics, educational innovation,
and the development of AI systems for edu-
cation and learning in a limited way. As part
of the broader European data governance
and federation initiative Gaia-X, the public
interest association Prometheus-X is in the
process of bringing together actors in the
educational sphere to develop a govern-
ance system for education and skills data.
One of the use cases of Prometheus-X is
for enabling actors to pool aggregated data
in order to train artificial intelligence algo-
rithms.39 This work, however, is still in its
early stages.
Learning is about personal development,
and fine-grained data on learning process-
es are inherently personal. This means that
existing regulations on personal data, data
governance, and privacy are important fac-
tors in shaping the information architecture.
More generally, what data are collected
and how they are used needs to be justified
based on socially accepted understanding
about the aims and objectives of education.
This links the ethics of education to the in-
38. https://www.digitaleurope.org/data-space-for-skills/
39. https://prometheus-x.org/
34
SCIENCE FOR POLICY REPORT
formation architectures that support educa-
tion. Data structures, in this sense, are in-
herently political and those structures used
for education will require specific attention
on how they should be regulated and what
kind of data can be collected. Efforts aimed
at empowering individuals to gain owner-
ship on their data are particularly relevant
in this regard. Solid has already developed a
specification enabling people to store their
data securely in decentralized data stores40
and there are examples of prominent initi-
atives within the EU, such as the Flemish
Data Utility Company.41
New federated computational and data
architectures (Kaissis et al., 2020) can pre-
serve personal data in large-scale machine
learning and data-analytics, and they could
become important for educational data in-
frastructures. They could, for example, ad-
dress many of the requirements of GDPR.
The concepts of personal data and privacy,
however, are complex (e.g., Etzioni, 1999;
Roessler & Mokrosinska, 2015; Sax, 2018).
To better understand what privacy means
in practical educational contexts, we would
need to elaborate and study a variety of use
cases. For instance, existing uses of AI such
as online proctoring (Coghlan et al., 2021;
Henry & Oliver, 2021; Mutimukwe et al.,
2023) would benefit from EU level policies.
40. https://solidproject.org
41. https://www.vlaanderen.be/digitaal-vlaanderen/athu-
mi-het-vlaams-datanutsbedrijf/the-flemish-data-utility-company
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
35
Digital learning has dramatically increased
the amount of data that can be used to an-
alyse learning processes and outcomes. As
a result, interest in fields such as Learning
Analytics (LA) and Educational Data Min-
ing (EDM) has grown rapidly during the last
decade.
‘Learning Analytics’ commonly refers to
the measurement, collection, analysis and
reporting of data about learners and their
contexts, for purposes of understanding and
optimizing learning and the environments
in which it occurs (Lang et al., 2017). ‘Ed-
ucational Data Mining’, in turn, focuses on
the development of methods for exploring
the unique types of data that come from
an education context (Romero et al., 2010).
There are many overlaps between LA and
EDM (Siemens & Baker, 2012), and as ma-
chine learning is increasingly being used in
both, these overlaps are growing. Educa-
tional institutions also generate data for
many regulatory and management purpos-
es, ranging from web usage statistics and
telemetry data to scheduling, planning, and
monitoring (Kitto et al., 2020), and open
school data (OSD) policies are becoming
commonly adopted around the world (Pois-
son, 2021). It is reasonable to assume that
‘in the future, administrative and real-time
learning data will be updated and analyzed
in real time’ (De Witte & Chénier, 2023).
In general, the amount of data collected
on learning and education is growing fast.
As noted above, this has also been charac-
terized as the datafication of learning and
learners (Lupton & Williamson, 2017). On
the other hand, data-driven AI requires data
both for system development and opera-
tion, and the potential benefits of AI in ed-
ucation can only be realized with access to
data on learning, not only by engineers and
data developers but also by educational
stakeholders.
Until recently, AI in education (AIED) has
mainly relied on knowledge-based AI in-
stead of machine learning approaches
(Holmes & Tuomi, 2022). Whereas da-
ta-driven AI uses oen large amounts of
data and machine learning to come up with
predictive models, knowledge-based AI re-
quires explicit representation of knowledge
structures. This usually means that human
experts need to describe and define domain
knowledge, for example, in terms of rules,
expectations, or semantic relations.
This history of AIED is very rich both in
pedagogical ideas and computational ap-
proaches. Much AIED research has focused
on intelligent tutoring systems (ITS), where
personalised learning has been supported
by individualized sequencing and spacing
of content to be mastered (Tuomi, 2023a).
Although prototypical ‘screen-level’ intelli-
gent tutoring systems have been extended
to ‘classroom-level’ AIED and beyond, and
many alternative pedagogic approach-
es have been studied also in the context
of intelligent tutoring systems (du Boulay,
2019), rapid advances in data-driven AI
over the last decade have now introduced AI
in educational settings in unexpected ways
too. In particular, the recent easy access
to generative AI tools has highlighted the
4 AI and learning
analytics
36
SCIENCE FOR POLICY REPORT
potential of AI in education for the general
public, while prompting concerns about the
use of student data to improve its function
The state-of-the-art in AI is advancing very
fast. Whereas AI systems have until now
been limited to well-defined and special-
ized tasks, large language models, such as
OpenAI’s GPT-4, now seem to be able to
emulate some human performance in rela-
tion to a diverse range of intellectual tasks.
Whether such achievements can actually
be counted as general AI remains a hotly
debated topic in the field (Bubeck et al.,
2023; Knight et al., 2023). There is, how-
ever, a general agreement that AI will have
a transformative impact on education and
learning in the coming years.
Policy efforts, most notably the EU Artifi-
cial Intelligence Act (European Parliament,
2023), will play a key role in both harness-
ing the opportunities and mitigating the
risks of AI, particularly in education, signal-
ling some use cases that could be consid-
ered as ‘high-risk’ and highlighting the need
for AI Literacy.
4.1 Generative AI for
teaching and learning
AI can be used in learning, for learning, for
teaching, and for education administration
(Holmes et al., 2019). Fast improvements
in data-driven large language models have
led some observers to believe that AI-sup-
ported personalised learning is now be-
coming possible, and education, in fact, will
soon become a major driver in commercial-
izing consumer (Skates, 2023).
Applications such as language translation,
search, speech-to-text, and text-to-speech,
rely on data-driven AI, and they are fre-
quently utilized by both students and teach-
ers. In recent discussion most attention has
been paid to generative AI foundation mod-
els that can be adapted to various tasks.
These models include, for example, the
large language models that underpin gen-
erative AI systems such as Google’s BERT
(Devlin et al., 2018), OpenAI’s GPT-3 (Brown
et al., 2020), and the BLOOM open-source
language model (Scao et al., 2022). Dia-
log-based modifications of these systems,
such as OpenAI’s ChatGPT, Microso’s GPT-
4 derivative Bing Chat, Google’s Bard, and
Meta’s LLaMa have become widely popu-
lar with speeds that greatly exceed earlier
technologies.
There are now many general-purpose AI
models available for experimentation and
commercial use.42 The development of
state-of-the-art data-driven AI systems,
however, oen requires internet-scale data
and extraordinary amounts of computing
power. This has resulted in dynamics that
will link not only innovation and education,
but also environmental policies in novel
ways.
4.1.1 The emerging AI ecosystem
At present, it is not clear what the emerg-
ing AI ecosystem will look like in the medi-
um-term or even near future. It is possible
that it will be dominated by a small number
of very large firms that have been charac-
terized as ‘google-sized natural monopo-
lies of the Internet’ (Tuomi, 2020). It is also
possible that the recent fast progress in AI
will hit technical barriers that allow many
small players to catch up with the leading
AI developers. For example, state of the art
data-driven AI systems require very large
amounts of energy, potentially making AI an
important driver of global warming (Bender
et al., 2021; Luccioni et al., 2022; Masanet
et al., 2020; Strubell et al., 2019; Tuomi,
42. In April 2023, the machine learning platform Hugging Face
Hub (https://huggingface.co) provided access to over 187,000
models and 31,000 datasets.
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
37
2020). More recently, the water usage of
data centres has also become a problem
(Siddik et al., 2021). AI researchers, govern-
ments, and others have called for tighter
regulation and a total moratorium on AI de-
velopment (Klein, 2023).
Unless the AI industry hits roadblocks, ex-
isting models of industrial innovation sug-
gest that the large variety of AI system
designs that follow recent technological
breakthroughs will become consolidated
in a small number of ‘dominant designs’
(Utterback, 1994; Utterback & Abernathy,
1976). One could argue that this has al-
ready happened with generative AI models
as they are typically based on the same
‘transformer’ architecture. According to the
Abernathy-Utterback model of product and
process innovation, the ‘fluid phase’ of in-
novation, where many different alterna-
tive designs are introduced and search for
a market, is then superseded by a phase
where process improvement dominates. If
this is the case, it can be expected that a
small number of organizations will control
the future AI ecosystem.
During the last years, improvements in AI
performance have been associated with
larger models, more data, and higher pro-
cessing requirements. This has made even
the largest AI developers worried that they
will run out of compute capacity soon
(Tuomi, 2020). A recently published fund-
raising pitch deck of the OpenAI spin-off
Anthropic stated that the company will
need up to $5 billion to become compet-
itive in the generative AI market (Wiggers
et al., 2023). More interestingly, Anthropic
argued that the companies that are able
to train the best models in 2025/6 will be
so far ahead that no-one will be able to
catch up with them in subsequent cycles.
Anthropic expects that it would be able to
develop ten times more powerful AI mod-
els than the current state-of-the-art using
compute clusters that would contain tens
of thousands of GPUs. OpenAI itself has
become allied with Microso, which in ear-
ly 2023 invested several billion dollars in
OpenAI. Much of this cost is for compute
capacity that is needed to train large lan-
guage models.
Training foundation models requires com-
pute power that only few organizations
can afford. For example, the open-source
BLOOM model was trained for 3.5 months
on 384 A100-80GB enterprise level GPUs
(Scao et al., 2022). The current price for
these processors is above €15,000 each.
Application-specific fine-tuning of existing
foundation models can be done with less
compute, but as the models are very large
(e.g., 176 billion parameters in BLOOM),
only large AI developers have the required
infrastructure for this. BLOOM was therefore
created by a large group of AI researchers
with the aim to democratize access to large
language models.
What such a democratization means in
practice is an important question which also
has implications for education policy.43 From
the point of view of innovation dynamics,
an open-source large language model such
as the BLOOM provides equal access to a
‘foundation model’ for many innovators.
This model can then be used to develop a
large variety of applications. At the same
time, it becomes a shared platform and a
technological artefact where the interests
of many developers and users meet. The
resulting inter-dependencies mean that in-
43. Beyond computing capability, the training of large language
models also requires extremely large amounts of data. This is a
challenge for smaller languages and domain-specific applica-
tions. For example, the TurkuNLP research team at the University
of Turku has released an open-source language model for the
Finnish language. This is a 13 billion parameter model developed
using the LUMI supercomputer, which is the fastest supercom-
puter in Europe. The researchers note that they are now starting
to run out of Finnish language resources (Jakobsson, 2023).
38
SCIENCE FOR POLICY REPORT
novation in the foundation itself needs to
slow down and become controlled.
A similar innovation dynamic made possi-
ble the historical development of the Linux
open-source ecosystem (Tuomi, 2002). The
tight control of the Linux core operating
system, or its ‘kernel’, enabled very fast
expansion of Linux and the related appli-
cations. Under such innovation dynamic, it
can be expected that a small number of
domain-specific large language models will
emerge in the near future. An early exam-
ple of such a domain-specific model is the
BloombergGPT, developed for the financial
sector (Wu et al., 2023).
If Anthropic is right, and massive invest-
ments guarantee sustainable dominance
in the emerging AI ecosystem, or if open-
source foundation models such as BLOOM
can avoid the emergence of natural mo-
nopolies and oligopolies, both alternatives
will have important implications for educa-
tion policy. National policymakers have tra-
ditionally formulated educational policies
that address local needs. This is also re-
flected in the division of competences with-
in the European Union, where the EU has
only supporting competences in education
and vocational training.44 With the increas-
ing use of data-driven AI, more and more
we see that the reliance on global actors,
mainly the EdTech industry, even though
the remit of education remains with the
Member States. This may be shiing power
control from to the public sector to the pri-
vate industry to some extent. The wide use
of digital platforms also means that large
global platform providers probably now
have access to more data about students
and education providers than teachers, ed-
ucational institutions, or policymakers. The
extraordinary dynamics of data-driven AI
therefore generates an industrial and inno-
44. Treaty of the Functioning of the European Union, Article 6.
vation ecosystem whose structure has di-
rect impact on education.
The economic cost of developing state-of-
the-art foundation models is high for small
start-ups, but small when compared with
the costs of education. It is commonly es-
timated that the cost of training of Ope-
nAI’s GPT-3.5 (on which ChatGPT was based
when it was originally released) was be-
tween $2 and $5 million. The training cost
for Google’s PaLM has been estimated to
be $8 million (Maslej et al., 2023, p. 62).
Such estimates typically include only the
compute costs for training the model. These
probably represent only a fraction of the to-
tal costs that include the iterative develop-
ment and training of several related mod-
els, data collection and curation, and other
similar costs. The total cost still remains a
fraction of the total cost of education. Ac-
cordingly, we may consider whether part of
the costs could be shied to governments if
educational benefits are identified from de-
veloping foundation models (or, for exam-
ple, training such models for small regional
European languages).
Machine learning architectures for large lan-
guage models are, however, also becoming
more efficient, and there are several ways
in which high-performance models can be
made smaller. For example, DeepMind’s
Chinchilla (Rae et al., 2022) performs on
par with the earlier 280 billion parameter
Gopher with one fourth of the parameters.
Similarly, with some additional training us-
ing human feedback, OpenAI’s InstructGPT
was able to generate outputs that the us-
ers preferred over the responses of GPT-3
that had 100 times as many parameters
(Ouyang et al., 2022). The benefits of even
larger models are actively being studied,
and the development of increasingly large
models continues. For example, in April
2023 Google switched its Bard from the
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
39
137 billion parameter LaMDA model to its
540 billion parameter PaLM model.45 The
latter was trained using 8192 Google’s TPU
AI processors on a platform that can theo-
retically achieve up to 1.1 exaflops, or 1.1
million trillion calculations per second.
The different variants of Meta’s recently
published instruction-following LLaMA use
a much smaller number of parameters (7B,
13B, or 65B) than PaLM or the 175 billion
parameter GPT-3. LLaMA is an open-source
large language model released in February
2023 that can be fine-tuned for user appli-
cations. Due to its relatively small size, this
is also possible with limited computation-
al resources. For example, researchers at
Stanford University did this with their Alpaca
model, fine-tuning it with just a few hun-
dred dollars of rented computing capacity.
Researchers from UC Berkeley, CMU, Stan-
ford, and UC San Diego further improved on
Alpaca, training their Vicuna model using
70,000 user-shared conversations.46 Using
OpenAI’s GPT-4 to compare the quality of
the conversation outputs, Vicuna was es-
timated to achieve about 90 per cent of
the ChatGPT performance. From the point
of view of practical applications this is re-
markable as the cost of training of the 16
billion parameter LLaMa-based Vicuna was
about $300.47 Recent research (Gudibande
45. The exact model sizes used for Bard have not been pub-
lished, but it is known that Bard was built using LaMDA. The CEO
of Google has characterized the original Bard as a ‘Honda Civic’
among race cars (Roose et al., 2023), which suggests that one
of the smaller LaMDA models could have been used for the first
Bard version.
46. https://github.com/lm-sys/FastChat
47. It should be noted that Alpaca and Vicuna are both fine-
tuned versions of Meta’s LLaMa, and the costs of pre-training
the LLaMa model are not included in these estimates. The
training costs for GPT-4 are not publicly known, but OpenAI’s re-
search papers suggest that GPT-4 was trained with 1000 times
more computations than GPT-3.5. The number of parameters in
GPT-4, however, could be similar to GPT-3.5 as presumably much
more data were used to train GPT-4. ChatGPT has been claimed
to have 20 billion parameters, but it is oen also said to have
175 billion parameters, probably because this number is known
for GPT-3.5. At present, the most recent version of ChatGPT
(ChatGPT Plus) is based on GPT-4, but OpenAI has not published
technical details of the system or its training.
et al., 2023) has, however, also suggested
that these cheap derivatives of large lan-
guage models have important limitations
that can be overcome only by developing
larger models.
Although dialog-based text-to-text mod-
els, such as ChatGPT, have attracted much
attention among educators, generative
AI systems have also been developed for
many other areas. OpenAI’s Codex, which
is also based on the GPT-3 architecture, is
now widely used for computer program-
ming. Amazon launched its CodeWhisper-
er, ‘a real-time coding companion’ based
on a large language model, in April 2023.
Related text-to-image generators such as
OpenAI’s DALL-E 2, Stability AI’s Stable Dif-
fusion, and Midjourney have become im-
portant tools for graphics design and image
production.
OpenAI’s GPT-4 broke new ground in gen-
erative AI as it can use both text and im-
age inputs. Multimodal large language
models that can see and talk are rapidly
becoming available. An example is the mul-
timodal large language model KOSMOS-1,
developed by Microso researchers, which
encodes images as a form of language
and uses a large language model to rea-
son about the image content (Huang et al.,
2023).
Low-cost access to large foundation mod-
els developed by the leading AI developers
has led to very rapid growth in data-driven
AI products aimed for learning and teach-
ing. The dominant industrial actors are try-
ing to establish commercial platforms that
would become central points in the emerg-
ing ecosystem. For example, in April 2023
Amazon launched its Bedrock service that
provides access to several state-of-the-art
foundation models.48 These can be linked
48. https://aws.amazon.com/bedrock/
40
SCIENCE FOR POLICY REPORT
with other Amazon services for large-scale
cloud-based deployment.
Beyond the established knowledge-based
AIED tools, many start-ups now use da-
ta-driven AI technologies. These cover the
full range from early childhood to primary,
secondary, tertiary, and lifelong learning. As
the early excitement about the capabilities
of generative AI recedes, and the limitations
of large language models become better
known, generative AI systems will increas-
ingly be combined with knowledge-based
AI. Examples of such ‘augmented language
models’ are discussed below, but knowl-
edge-based AIED research will have more
visible influence in the development of AIED
systems in the future. For example, when
Khan Academy announced its GPT-4 pow-
ered Khanmigo in March 2023, it noted that
Khanmigo was informed by instruction-
al models developed for the well-known
knowledge-based AutoTutor (Nye et al.,
2014).
For pedagogic uses, generative AI appli-
cations – such as video captioning, trans-
lation, video summarizing and highlight
extraction, text-to-animation, and voice to
synthetic video – provide many new oppor-
tunities. These can be used, for example, in
generating new learning materials from ex-
isting content. For example, voice samples
can be used to clone a human speaker, and
the produced synthetic voice can be used to
transform text to speech, which can then be
used to animate still images or animated
characters.49 This would allow, for example,
a teacher to automatically generate video
lectures in multiple languages using lecture
notes written in the teacher’s native lan-
guage.
49. Low-cost voice cloning services and video generators are
now widely available. For example, ElevenLabs (https://beta.
elevenlabs.io) can clone voices and use these to translate text to
speech. D-ID (https://www.d-id.com) markets its Creative Reality
Studio, for example, as a tool for educators that can convert a
photo with text or cloned audio into an interactive and engaging
video presenter in over 100 languages.
Although the initial reaction to the publi-
cation of the ChatGPT at the end of 2022
was largely focused on potential misuses
of emerging technologies and ‘cheating
with AI’, at present the focus is shiing to
the ways in which generative AI could sup-
port learning and teaching (Sabzalieva &
Valentini, 2023; U.S. Department of Educa-
tion, 2023). In many of the proposed edu-
cational applications, ChatGPT collaborates
with the learner, for example, as a Socratic
opponent, co-designer, motivator, or study
companion.
ChatGPT and other generative AI systems
can clearly be used in education in many
ways. Personalised tutoring systems can
now be developed in days using ChatGPT
prompts, and, in addition to Khanmigo, this
approach has already been used in some
ChatGPT plugins. On a more systemic lev-
el, critical discussions on the potentially
conflicting interests of commercial stake-
holders and educators, as well as the need
to understand the various interests of the
stakeholders continues (e.g., Blikstein et al.,
2022; Selwyn, 2022b; Selwyn, 2023; Wil-
liamson, 2021; Williamson et al., 2022). In
general, the AI ecosystem is evolving very
fast and shaping both AIED research and
the AIED industry. Policy developers would
greatly benefit from a better understanding
of this emerging AIED ecosystem and its
implications for policy.
4.1.2 Abstraction and
generalization in generative AI
A noteworthy aspect of these generative
models is that their capabilities have reg-
ularly surprised their developers. For exam-
ple, the DALL-E text-to-image generator
apparently was able to make combinatorial
generalizations (e.g., ‘a tapir made of ac-
cordion’) and variable binding (e.g., ‘a baby
hedgehog in a Christmas sweater walking a
dog’) (Ramesh et al., 2021). This is surpris-
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
41
ing as the underpinning language models
are trained simply to predict the next word
in a sequence of words, without any explic-
it models of ‘generalization’, ‘abstraction’,
or ‘concept formation’. Some observers of
the GPT-4 development went further, argu-
ing that the system had gained unexpected
high-level thinking skills that could proper-
ly be called ‘sparks of general intelligence’
(Bubeck et al., 2023). As abstraction, gener-
alization, and concept formation are central
topics for learning theories, they are also
highly important for understanding the fu-
ture possibilities and limitations of the use
of data-driven AI in education and learning
(Tuomi, 2018b). Although data-driven AI
systems use purely behaviouristic learning
processes and have been characterized as
‘artificial instincts’ and ‘stochastic parrots’
(Bender et al., 2021; Tuomi, 2018a), a bet-
ter understanding of their emergent capa-
bilities for abstraction and generalization
will be highly important both for theories of
learning and for AIED system design.
4.1.3 The future of writing
Writing is a key transversal competence,
and in many theories of learning it has
been viewed as a key for cognitive devel-
opment. Writing skills are also important
in many professions and occupations. It is,
therefore, not surprising that generative AI
systems that can produce high-quality text
have been viewed as a threat to education
and, more generally, human cognition. In
the words of one commentator from the
Stanford Institute for Human-Centric AI,
automated writing tools are ‘a disaster in
making’ and potentially corroding our cre-
ative abilities. This is because ‘becoming a
good writer is the same thing as becoming
a good thinker’ (HAI, 2023, p. 20).
In the 1920s Vygotsky (1986) emphasized
the fact that writing is an unnatural skill for
a child. It is a tool that allows individuals
to communicate and express their ideas
and thoughts, and as such it requires direct
social communication and interaction to
be replaced by a mediated one. For a child
who has learned that spoken words gener-
ate immediate reactions in its environment,
it looks unnecessary to write words on pa-
per. In contrast to orality, writing is artificial
as Ong (1982) has also pointed out. While
there are immediate benefits from writing
for individuals ‒ as the ability to record
words effectively expands one’s own mem-
ory and consciousness – benefits to society
at large emerge primarily at a cultural and
historical level, where writing makes in-
ter-generational communication possible.
Such communication can further be accu-
mulated in conceptual and theoretical sys-
tems.
For Vygotsky, advanced forms of thinking
develop in a child when communicative
speech becomes an internal tool for linguis-
tic thinking. Culturally developed concep-
tual systems enable increasingly abstract
forms of thought, and the child learns these
conceptual systems through instruction and
education. From such a Vygotskian point of
view, generative AI systems are not inter-
esting because they produce text; instead,
their relevance for education is in their ca-
pability to engage humans in advanced
forms of thought where concepts, concep-
tual systems, and language are the tools for
thought. In such a developmental context, it
is not informative to ask whether paper or
pencil corrodes our creative abilities. More
interestingly, we can ask how pen, paper,
books, archives, typewriters, or AI become
integrated in thought. Similarly, we can
ask how generative AI can become part of
human cognitive processes. As with other
previous technologies, it is likely to disrupt
writing, but it is also likely to transform the
way we think or utilize writing for learning.
42
SCIENCE FOR POLICY REPORT
In general education, large language mod-
els show great promise in text analysis and
writing support. Text summarizing, outlining,
and formative writing support have been
important topics in knowledge-based AIED
for a long time (e.g., Knight et al., 2020;
Strobl et al., 2019). Large language models
are now increasingly used to generate sum-
maries of textbook chapters and, for exam-
ple, academic articles. The same text can
be summarized for different user groups
based on their age, language capabilities,
or preferences. Data-driven AI can also an-
alyse written text from various points of
view, helping the author to reflect upon the
produced text.
Commercial products in this area include,
for example, Rephrasely,50 which converts
user sentences from one style to another
in over 100 languages, and Jasper51 that
generates text for blogs, websites, social
media, and ‒ as suggested by the devel-
opers of Jasper – also love letters, based
on user prompts. Large language models
are also currently integrated in collabora-
tion and knowledge management tools. For
example, the collaboration and note-taking
platform Notion52 now supports the gen-
eration of summaries and insights of user
notes, document rewriting, translation, and
tone change, as well as the creation of doc-
ument dras. Researchers have also ex-
plored more broadly the nature of AI-sup-
ported creativity and writing (e.g., Sharples
& Pérez y Pérez, 2022), and suggested that
AI provides opportunities to experiment
with novel pedagogic models.
As generative AI becomes integrated in var-
ious text- and language-related tasks, we
may need to rethink the nature of writing.
For example, Buckingham Shum (2023) has
50. https://rephrasely.com/
51. https://www.jasper.ai/
52. https://www.notion.so/
proposed that generative AI could transform
writing in ways that resemble the changes
in music production during the last decades.
Since the early 1990s, professional-quality
music production has become widely ac-
cessible as digital audio workstations have
replaced expensive recording studios. Pro-
grammable music synthesizers and audio
effects can now easily be combined with
sampled voices and recorded audio and
arranged into final products using special-
ized music production interfaces, non-linear
editing tools, and AI-supported audio pro-
cessing. Buckingham Shum’s ‘writing synth
hypothesis’ proposes that with the emer-
gence of generative AI, authors will be able
to learn writing in new ways, democratizing
writing just as we saw with music synthe-
sizers.
A special form of text production, oen
claimed to be important in the emerging
digital world, is computer programming.
Although fully automated program gener-
ation is still a challenge, several systems
have recently been used to support both
professional soware development and
more novice programmers. Github Copilot,
powered by OpenAI’s Codex, is now widely
used to improve programmer productivity.
Amazon’s CodeWhisperer can generate pro-
gram code from simple natural language
prompts. Somewhat surprisingly, the gener-
ic ChatGPT has also shown remarkable ca-
pabilities in generating program code from
natural language prompts and in analysing
and explaining the functionality of exist-
ing code fragments in natural language.53
Recent code-generation systems, such as
DeepMind’s AlphaCode (Li et al., 2022),
suggest that AI-based code generation can
already outperform highly skilled human
programmers in some cases. These devel-
53. Because of the increasing capabilities of generic large
language models, OpenAI has now discontinued support for the
Codex API.
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
43
opments have implications for the demand
of advanced digital skills and related poli-
cies, for example.
4.1.4 Prompt engineering and
in-context learning
Generative text-to-text AI systems are de-
veloped by iteratively pre-training the AI
model, oen with thousands of trillions of
computations and hundreds of billions of
words. This process creates ‘baseline mod-
els’ such as the GPT-3. The baseline models
can then further be retrained or fine-tuned
for specific tasks, for example so that they
become better at providing relevant out-
puts (e.g., in legal or medical domains), or
in responding to textual instructions. An im-
portant example of such a fine-tuned mod-
el is the InstructGPT (Ouyang et al., 2022),
which has further been refined into the
well-known dialog-based ChatGPT.
Large language models are trained to
predict the following word given a list of
preceding words. A technical breakthrough
in this area was the ‘transformer’ architec-
ture that effectively uses a broader textual
context for this prediction than earlier mod-
els (Vaswani et al., 2017). When transformer
models are used to generate text, they use
a given ‘prompt text’ and extend this initial
context with the generated word sequence
to predict the next word. In the original
ChatGPT model, this context had approxi-
mately 3000 words. This means that the
output will greatly depend on the prompt,
and the system can be ‘steered’ by modify-
ing the prompt. In contrast to pre-training,
this is called ‘in-context learning’ (Brown et
al., 2020). Parts of the prompt, a ‘system
prompt’, can be defined by the system de-
veloper to guide the system output towards
predefined tasks. The remaining part, a ‘user
prompt’, is typically given by the end user.
The generated output strongly depends on
the prompt, and for example, the order of
sentences and words in the prompt (Gao
et al., 2021; Wei et al., 2023; Zhao et al.,
2021). The skilful design of prompt texts, or
‘prompt engineering’, has therefore quickly
become important for effective use and de-
ployment of generative AI systems.
Large language models can, therefore, be
trained on three very different levels. The
textual training data provided to the base
model defines its behaviour, which can fur-
ther be refined by fine-tuning with task-spe-
cific data. The behaviour of the model can,
however, also be changed by prompts.
Large language models can, therefore, be
‘programmed’ through natural language
prompts and telling the model what it is ex-
pected to do. These instructions can provide
a few examples of the sought-aer behav-
iour, or just a single instruction without fur-
ther guidance. The former is known as ‘few-
shot learning’, and the latter as ‘zero-shot
learning’. Zero-shot learning can also be
used to prompt an image generating model
that translates a textual description into an
image, animation, or video (Ramesh et al.,
2021). In-context training can be done in
real time, and allows the system to adapt
to information that was not available when
the base model was trained. In-context
training also oen leads to better perfor-
mance than traditional model fine-tuning
(Caron et al., 2021).
ChatGPT became an instant success aer
its release because large language mod-
els are surprisingly good few-shot learners.
When some generic instructions are defined
by the system developers ‒ for example
on what a dialogue should look like ‒ the
end users see a zero-shot learner that can
meaningfully answer questions and gen-
44
SCIENCE FOR POLICY REPORT
erate text based on a single instruction.54
In ChatGPT, the system is fine-tuned using
examples on how to respond to instruc-
tions, and its behaviour is further guided by
in-context learning.
The importance of effective domain-spe-
cific prompt design has quickly been noted,
and there is now a rapidly increasing group
of people claiming to be specialists in this
area. At present, good prompt design is very
much an art form that requires a good un-
derstanding of the behaviour of the base
model. Automated methods for refining
optimal prompts are currently being devel-
oped (e.g., Zhang et al., 2022; Zhao et al.,
2021), but most prompt design occurs at
present manually on top of vendor-tailored
end-user applications, such as ChatGPT,
Codex, and Stable Diffusion. There are also
open-source initiatives, such as the Lang-
Chain Hub,55 that aim to develop and share
predefined prompts for various use cases.
As the underlying language models are con-
tinuously changing, it is not clear that deep
expertise can emerge on prompt engineer-
ing on these platforms. There are, howev-
er, some general guidelines for effective
prompt design. For example, using prompts
that clearly define expectations and push
the language model to elaborate its rea-
soning can substantially improve system
performance (Kojima et al. 2023). However,
it is clear that standardized evaluation met-
rics will be important for the measurement
of quality and effectiveness of domain-spe-
cific prompts that orient the language mod-
el, for example, in educational applications.
The evaluation framework that OpenAI
54. In this architecture, the developer-provided prompts can
sometimes be deleted, overwritten, or by-passed, with the result
that the system can operate as the fine-tuned base-model,
without further restrictions defined by the developer. This form
of system cracking is called ‘jail-breaking’ and the resulting
systems are called DANs (Do Anything Now).
55. https://github.com/hwchase17/langchain-hub
published56 together with GPT-4 is therefore
potentially an important element in shaping
the generative AI ecosystem. For example,
it would be possible to develop evaluation
benchmarks that measure how well a given
generative AI system is aligned with the EC
guidelines for the ethical use of AI in edu-
cation and learning (European Commission,
2022a).57
Knowledge-based intelligent tutoring sys-
tems typically had sophisticated models
of user’s current knowledge. The develop-
ment of algorithms for managing the user’s
knowledge model and the development of
domain knowledge models has been a very
labour-intensive task in traditional AIED.
Large language models, in contrast, rely
on user prompts to generate a temporary
snapshot of user intentions and knowledge
state. Agent-based language model archi-
tectures that have access to external mem-
ory, such as AutoGPT, BabyAGI, and other
agents supported by LangChain,58 can com-
bine persistent and evolving learner mod-
els and generic domain models that can be
refined for specific domains with minimal
development effort. The possibility for light-
weight personalization using well-designed
prompts and user interaction could become
important for just-in-time learning and
training, for example. An important exam-
ple of such an approach to personalization
is the Khanmigo system, currently being pi-
loted by Khan Academy.59
56. https://github.com/openai/evals
57. The EC ethical guidelines are based on a capability-based
approach, and it is possible to test generative AI systems using
the defined ethical capabilities. These are operationalised in the
EC guidelines as rubrics.
58. https://blog.langchain.dev/agents-round/
59. https://www.khanacademy.org/khan-labs
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
45
4.1.5 Trustworthiness and align-
ment in generative AI
Pure language models such as ChatGPT are
notorious for their ability to generate con-
vincing text that is factually wrong. Large
language models have also traditionally
lacked mathematical skills (though this has
recently improved), are unable to access
up-to-date information on recent events,
are unaware of the progression of time,
have difficulties in understanding languag-
es that are not widely used on the internet,
and invent facts and hallucinate non-exist-
ing realities.60
To reduce factual errors and ethically
and politically unacceptable output text,
ChatGPT uses a human-assisted learn-
ing process (reinforcement learning from
human feedback, or RLHF) (Glaese et al.,
2022; Stiennon et al., 2022). For the time
being it is not clear whether this approach
can be effectively scaled up. It is, however,
possible to improve the trustworthiness of
large language models by linking them to
existing knowledge sources. This approach
will also be important for education and
learning.
A disruptive recent development in genera-
tive AI has been the introduction of plug-in
and agent-based approaches. Agents use
language models to reason about the next
steps that are needed to perform a given
task. An example is the Toolformer archi-
tecture (Schick et al., 2023). Toolformer is
a language model based on a small GPT
transformer model (GPT-J), but it has been
additionally trained to predict external data
sources and tools that can be used to get
accurate information. A range of tools, such
60. For example, in ‘closed-domain’ tasks, such as summariza-
tion and closed-domain question-answering, where the output
should not contain information that is not present in the input,
GPT-3 had a 41% hallucination rate (Ouyang et al., 2022, p. 3).
In GPT-4 and ChatGPT, hallucinations are reduced by training a
separate model that rewards factually correct responses.
as a calculator, a question-answer system,
a search engine, a translation system, and
a calendar have been integrated in the
Toolformer architecture. A similar approach
is used in WebChatGPT61 that adds a web
search plugin to the browser that can aug-
ment the user’s prompt with data found
from the web. These functionally extended
AI models are also called ‘augmented lan-
guage models’ (Mialon et al., 2023).
Augmented large language models gained
visibility when OpenAI started to roll out
ChatGPT plugins at the end of March
2023.62 An interesting open-source variant
of this approach is the LangChain frame-
work,63 which supports the chaining of lan-
guage models, actions, and agents. This
makes it possible, for example, to augment
the language model dynamically with data
from the internet, from local databases, or
from previous interactions with the system.
In education, one possibility would be to link
language models with knowledge-based
models generated from textbooks and ex-
isting assessment rubrics.64
The general challenge of filtering ethically,
or otherwise unacceptable outputs, in gen-
erative AI is known as the ‘alignment prob-
lem’. This challenge is commonly addressed
by fine-tuning the base model to reduce
harmful outputs and to increase the prob-
ability of useful outputs. Recent research
has also tried to address this problem by
using machine learning instead of human
intervention. An example is Anthropic’s
Claude language model (Bai, Kadavath, et
61. https://github.com/qunash/chatgpt-advanced
62. https://openai.com/blog/chatgpt-plugins
63. https://langchain.com/
64. It is useful to note that large language models erode the
traditional distinction between knowledge-based and data-driv-
en AI. Whereas in the knowledge-based approach domain
models were handcraed by expert ‘knowledge-engineers’, large
language models have internal domain models that emerge in
the training process. Because of this, large language models can
be used as ‘expert’ reasoners in the agentic approach discussed
above.
46
SCIENCE FOR POLICY REPORT
al., 2022). Claude specifically aims to be
a ‘constitutional AI model’ whose outputs
are automatically trained to reduce their
potential ‘harmfulness’ based on externally
defined ‘constitutional principles’. This ap-
proach is similar to the RLHF, which is used
among other things to fine-tune OpenAI’s
ChatGPT, though Anthropic relies on AI-au-
tomated reinforcement learning (RLAIF).
The idea that explicitly defined constitution-
al principles and moral maxims could form
an ethically justified foundation for socie-
ties has a long history in political sciences,
legal theory, and ethics.65 The developers
of Claude have adopted a rather straight-
forward interpretation of such principles. At
present, the ‘constitution’ of Claude consists
of 16 ‘principles’ that the language model
uses to learn to rewrite its responses in less
harmful ways. The ‘harms’, themselves, are
classified using nine categories, including
‘discrimination and injustice’, ‘general un-
ethical behaviour’, ‘bullying or harassment’,
and ‘misinformation or conspiracy theo-
ries’. The interpretation of such principles,
of course, is problematic from ethical and
cultural points of view. It has been argued,
for example, that the ethics of AI cannot be
based on universal principles (Mittelstadt,
2019), that a more socio-developmental
approach is particularly needed in the edu-
cational domain (Tuomi, 2023b), and that in
general ethical principles depend on cogni-
tive metaphors that have many incompat-
ible interpretations (Lakoff, 1996). Align-
ment of generative AI in the educational
domain would therefore imply alignment
with some explicit articulation of ethics of
education.
In general, ‘alignment’ at present is about
‘alignment with developer and evaluator
65. For example, in his Law of Peoples, Rawls (1999) proposed
an ideal social contract theory, or a ‘universal constitution’, that
could and should be accepted by both liberal and non-liberal
societies.
preferences’. Any technical definition of
alignment will eventually need to be based
on specific ethical and political theories
that state how preferences or values are
expressed in the social domain, and how
they are taken into account in collective ac-
tion and social choice. As various alignment
methods are now being developed to make
AI more ethical, trustworthy, and acceptable
for various user groups, including school-
age children, it would be important to gain
better understanding of the appropriate
principles in educational settings, as well as
the limits of such alignment approaches. In
general, alignment research is now viewed
as a central technical and theoretical chal-
lenge in generative AI research (e.g., Ga-
briel, 2020; Kenton et al., 2021; Weidinger
et al., 2021). It is, however, clear that our
understanding in this socially and politically
important area is still quite elementary.
4.2 Open learner models,
agency, and humans-in-
the-loop
Knowledge-based intelligent tutoring sys-
tems require ‘learner models’ that repre-
sent the learner’s knowledge state. In many
influential AIED systems, user models have
been updated based on monitoring user’s
performance while the student progress-
es through the learning tasks. Such learn-
er models are updated by the system, and
the system uses information in the learner
model to steer system behaviour. In tradi-
tional intelligent tutoring systems (ITS) the
learner model serves the needs of the sys-
tem but remains opaque for the learner.
To support reflection and self-regulation in
learning processes, AIED researchers have
since the 1990s experimented with various
types of ‘open learner models’ (Bull et al.,
1995; Bull & Kay, 2010; Brusilovsky, 2023;
Bull, 2020). Open learner models, in effect,
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
47
open up the black box of the learner model
in traditional knowledge-based ITS, and al-
low the learner to understand how learning
proceeds. In contrast to content-oriented
views on instruction and education, research
on metacognition and cognitive develop-
ment highlight learning as a process that
enhances our capacity for thinking. Impor-
tant aspects of this capacity are self-reg-
ulation (Azevedo et al., 2019) and socially
shared regulation (Järvelä et al., 2023) dur-
ing learning processes. Mastery of the con-
tent to be learned plays an important role,
but only an auxiliary one (Tuomi, 2023a).
Metacognition and reflection play a central
role in many influential theories of learning
(e.g., Dewey, 1991; Piaget, 1971; Schön,
1987; Vygotsky, 1986). Vygotsky (1986, p.
166), for example, held that those higher
intellectual functions that are at the fore
of development during the early school age
are characterized by reflective awareness
and deliberate control. Similarly, Piaget em-
phasized the all-important role of self-regu-
lation (e.g., Piaget & Inhelder, 1979, p. 159).
Learning, for Dewey, Vygotsky, and Piaget,
is about the development of intelligence
and the capacity to control it. The increasing
capabilities for self-reflection, equilibration,
and self-regulation are the signposts in this
process.
Such a developmental view on education
becomes important when AI and learn-
ing analytics systems are used to support
learning processes. As domain-specific
knowledge becomes easily available and
its lifetime is oen short, the focus on edu-
cation shis from content transfer towards
meta-cognition, including learning strat-
egies. Open learner models can then be
used, for example, to represent the learner’s
knowledge to the learner, allowing them to
reflect on their learning. Beyond helping to
focus on content areas where attention is
needed, AI can be used to generate mean-
ingful interpretations of the learning expe-
rience that help the learner to understand
learning. Beyond automating instruction,
AI can be used to develop capabilities for
learning.
In such a developmental view on learning, AI
systems become ‘intellectual companions’,
‘learning partners’, and ‘cognitive tools’ that
support the development of thought. Over
the decades, this view has been an impor-
tant thread in the research on computer sup-
ported learning (e.g., Pea, 1985; Salomon
et al., 1991; Salomon, 1993) and human
computer interaction (Kaptelinin & Nardi,
2006), but it is again becoming prominent
because generative AI systems can be used
in this mode in many practical educational
settings. This has implications, for example,
for the ways in which agency is distributed
in learning processes where humans and AI
tools interact. When accumulated learning
is used for work tasks performed jointly with
such intellectual companions, the tradition-
al concept of ‘skill’ also needs to be recon-
sidered.
4.2.1 Distributed agency in
education
The integration of AI systems with human
learning processes makes agency a cen-
tral concept in education. Agency has been
defined in various ways in the research lit-
erature, and it has also gained increasing
visibility in AIED and learning analytics re-
search (e.g., Brod et al., 2023; Buckingham
Shum et al., 2022; Jääskelä et al., 2021).
Agency has oen been understood in a de-
cision-making context, for example, as the
freedom and control in determining action
(Sawyer et al., 2017). In many influential
intelligent tutoring systems, the student
simply had to perform actions determined
by the machine. Editable open learner mod-
els that allow the student to interact with
the system give the user some power over
48
SCIENCE FOR POLICY REPORT
the machine and allocate the student some
agency. More broadly, in emerging AI-ena-
bled learning systems, agency can be dy-
namically allocated among the computa-
tional system and the human.66
Strictly speaking, this distribution of agency
is nothing new. It is a key characteristic in
formal education, where teachers guide stu-
dents on their learning path. In educational
settings, the teacher is in the learning loop,
and this is called instruction. A central idea
in Vygotskian learning theories was that
children learn with the help of more com-
petent adults. Learning occurs in the ‘zone
of proximal development’ where the child
can use advanced form of thinking, but only
when guided by someone who already has
this capability, and who can provide cog-
nitive scaffolding for the developing child
(e.g., Rogoff, 1990; Wertsch, 1985). In the
context of such theories, one can ask how
hybrid AI-human systems can be integrated
in the learning process (Järvelä et al., 2023;
Luckin & Boulay, 2016; Molenaar, 2022;
Scardamalia & Bereiter, 1991). One can
also ask how AI could support new forms of
agentic distribution in educational settings,
for example, by putting the student, teacher,
and the parent in the learning loop with AI.
This requires that the common individual-
istic views on agency and competence are
expanded to include those social and tech-
nical resources that underpin agentic action
(Tuomi, 2022).
Learning analytics and open learner models
can be used to support student self-reflec-
tion and self-regulation. When more agen-
tic AI systems are used for this, dynamic
division of intellectual labour becomes pos-
66. It is perhaps useful to note that one main claim in actor-net-
work theory (Law & Hassard, 1999) was that there is symmetry
between technical and human actors. In contrast, in activity
theory (Kaptelinin & Nardi, 2006), technology does not have its
own motives and the symmetry between humans and functional
artifacts breaks down. It is an interesting question how these
theoretical approaches should be revised when AI systems and
humans share agency (Newman et al., 2019).
sible. In contrast to learning theories rooted
in child psychology, research on knowledge
creation has focused on adult learners and
their social collaboration. In this context,
learning has oen been understood as the
production of new knowledge.67 The devel-
opment of AI suggests that these studies
could be reinterpreted in the emerging tech-
nological contexts. In contrast to the rela-
tively static environments in a classroom,
workplace environments vary with the tasks
performed and tools used, and knowledge
creation oen becomes more important
than assimilation of prescribed knowledge
(Buckingham Shum et al., 2022).
Large language models are especially in-
teresting in the Vygotskian theoretical con-
text. For Vygotsky, language provides the
foundation that makes the transition to ad-
vanced forms of adult thinking possible. For
Piaget, a similar qualitative developmental
transition resulted when a child internalized
egocentric speech. According to Vygotsky,
external tools and symbol systems that
support human cognition are necessary for
the development of a child. In educational
contexts, the importance of large language
models may, therefore, be less in their au-
tonomous capacity to show ‘sparks of gen-
eral intelligence’ than in their capability to
transform human intelligence.
4.2.2 Agency as capability
In a broader developmental and capabili-
ty-based context, agency can be interpreted
67. The term ‘knowledge creation’ was introduced to the Eng-
lish-speaking organizational learning and innovation literature
by Nonaka (1991, 1994). The term suggests that knowledge
is socially constructed, but it also invites making explicit the
epistemological assumptions that underpin models of learning.
Nonaka’s work was informed by the phenomenological episte-
mology developed in the Kyoto school of philosophy, but other
conceptual frameworks – such as autopoietic systems theory
(von Krogh & Roos, 1995), anticipatory systems theory and cate-
gory theory (Ehresmann et al., 2018), activity theory (Engeström,
1999), and semiotics (Paavola & Hakkarainen, 2005) – have
also been used. A more static view on knowledge underpins
management literature on organizational learning (e.g., Argyris &
Schön, 1978; Hedberg, 1981; Levitt & March, 1988).
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
49
as the capability for informed and heedful
action. Informed action requires information
and knowledge. In the ethics of AI this has
oen been understood as the requirement
of system transparency and explainability.
The core concepts of AI system governance:
transparency, explainability, and ‘human-in-
the-loop’, therefore, can be understood in
terms of agency. This makes open learner
models relevant for the practical implemen-
tation of risk-based AI governance models
in education, and beyond it.
A key assumption in open learner models
is that learning requires self-reflection and
self-regulation. For learning to occur, the
learner must be able to participate in the
human-machine loop as an agent. For re-
sponsible action, the learner further must
be able to adjust the allocation of agency
and, if needed, control the technical system.
In contrast to most discussions on AI gov-
ernance, learning theorists would point out
that agency is a developmental outcome.
Whether a human in the loop can act as an
agent depends on a person’s competenc-
es and the possibilities to express these
in given cultural and normative contexts.
These, in turn, largely depend on accumu-
lated learning. Whereas AI engineers oen
approach explainability and transparency
from a technical point of view, as access
to data, a capability-based approach sug-
gests that agency development would be a
more productive starting point. Knowledge
about learning, in other words learning the-
ories, becomes therefore highly relevant
for apparently technical discussions about
explainability, transparency, and the control
and governance of high-risk AI systems.
A key difference between knowledge-based
and data-driven AI systems is the potential
openness of user, domain, and inference
models. Knowledge-based AI systems im-
plement explicitly defined models of rea-
soning and the domain of system’s exper-
tise. Because these knowledge structures
are explicitly defined during system design,
it is possible to represent and trace the rea-
soning process for the user. In data-driven
AI systems the reasoning process is inher-
ently opaque as it depends on the current
state of oen billions of parameter values.
Because of this, data-driven AI systems are
oen regarded as ‘black boxes’.
The distribution of agency between a hu-
man and an AI system requires that both
are able to track the other’s reasoning. Sur-
prisingly, this seems to be something that
can be done with large language models.
When large language models are used in
an agentic mode, they can be prompted to
describe the reasons why they make specif-
ic choices. This is called ‘chain of thought’
prompting (Wei et al., 2023). A chain of
thought can be induced simply by telling the
model to ‘think step-by-step’(Kojima et al.,
2023) or by giving it examples that demon-
strate a series of intermediate natural lan-
guage reasoning steps that lead to a final
output. A sufficiently large language model
can then follow this example and make ex-
plicit the intermediate steps that it uses in
problem solving. For instance, Table 1 illus-
trates how aer a user (i.e. human) inputs a
sequence containing a gap in it (i.e. the sec-
ond answer), the AI assistant (i.e. GPT 3.5)
can infer that what the user wants is for it
to produce an explanation of the situation
presented in the third statement within the
prompt submitted by the user. Whereas
there has been much debate about the im-
possibility of explaining the behaviour and
choices made by data-driven AI systems,
language models can do what humans do:
if asked, they can explain what they are do-
ing and why. Nevertheless, just like humans
too, they might well hallucinate or provide
factually wrong information when providing
such further explanations.
50
SCIENCE FOR POLICY REPORT
An example of chain of though with GPT-
3.5 is shown in Table 1.
Table 1: Chain of thought prompt with explanation
User
Q: A coin is heads up. Dan-
iel flips the coin. Romina
flips the coin. Is the coin
still heads up?
A: The coin was flipped by
Daniel and Romina. So the
coin was flipped 2 times,
which is an even number.
The coin started heads up,
so aer an even number
of flips, it will still be heads
up. So the answer is yes.
Q: A coin is heads up. Ilkka
flips the coin once and then
again.
A:
Assistant
The coin was flipped twice,
which is an even number.
The coin started heads up,
so aer an even number
of flips, it will still be heads
up. So the answer is yes,
the coin is still heads up.
Source: the authors and GPT 3.5.
In the capability-based approach, agency
characterizes what a person is free to do
and achieve in pursuit of goals and val-
ues the person regards as important (Sen,
1993). As Sen has pointed out, the capacity
to transform information and other resourc-
es to action depends on individual charac-
teristics, as well as social relations, the cul-
tural context, and its norms.
The development of agency, therefore, can
be viewed as a central sustainable objec-
tive of education (Tuomi, 2023b). This view
contrasts with common skill- and compe-
tence-oriented views that typically con-
ceptualize skills and competences from an
instrumental point of view. In the instru-
mental approach it is natural to ask why
a specific skill is useful and what it can be
used for. Such questions, however, assume
a closed world where the future unfolds ac-
cording to a given plan, and innovation does
not disrupt social and economic practices.
The capability-based model of agency and
improved understanding of the ways in
which agency can be distributed in AI sup-
ported environments can be argued to have
fundamental importance for educational
theories and policies in the coming years.
Although there has been much research on
these topics over the last century, technical
developments in AI make earlier studies rel-
evant in new ways. When the objective of
education is viewed as the development of
agency, the distribution of agency between
human and AI agents becomes important
also for the development of new pedagogic
approaches.
4.2.3 Future of skills
Although education has many social func-
tions, policy and societal debates oen focus
on the development of skills and knowledge
that are useful for employment. The future
of work is clearly also important for the
development of education policies. Econo-
mists have great difficulties in measuring
the labour market impact of AI (Frank et al.,
2019), but it is widely accepted that AI will
have a profound impact on labour market
skill demand and composition (e.g., Arregui
Pabollet et al, 2019; Brynjolfsson, 2022;
Eloundou et al., 2023; Felten et al., 2023;
Nedelkoska & Quintini, 2018; Servoz, 2019;
Sostero & Tolan, 2022; Tuomi, 2018a).
AI will change occupations and job tasks
in the coming years, perhaps more funda-
mentally than previously predicted. As was
noted in the introduction, researchers from
Goldman Sachs suggest that two-thirds of
current jobs are exposed to AI-based au-
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
51
tomation. Influential earlier studies on the
AI impact ‒ conducted before the recent
breakthroughs in generative AI ‒ were based
on the assumption that tasks that require
communication and social interaction skills
will be difficult to automate using AI (e.g.,
Autor et al., 2003; Frey & Osborne, 2017).
Because of this, researchers have believed
that AI will have a relatively minor impact in
teaching occupations (Nedelkoska & Quin-
tini, 2018).68 Recent developments in large
language models make such estimates of
labour market impact obsolete, as AI is now
widely used to support also non-routine
cognitive tasks that require high levels of
domain expertise.
Whereas most economic studies on the
impact of automation and AI have used
skill-biased and task-biased models of
technological change, the development of
data-driven AI systems requires data. It
has, therefore, been suggested that ‘da-
ta-biased’ models of labour market change
would be necessary to understand future
skill demand (Tuomi, 2018b). Large lan-
guage models can bypass this data bottle-
neck as they rely on human-produced texts
available on the internet. It is clear, how-
ever, that more detailed analysis would be
required to understand, for example, AI-in-
duced changes in teaching practice and skill
structures.
In general, well-established domain-spe-
cific skills become obsolete as a result of
technical change. As new key technologies
and general-purpose technologies emerge,
new types of knowledge, skills, and expe-
rience become economically and socially
important (Freeman & Soete, 1997; Perez,
1985). In practice, many prototypical skills
of the last century were mirror images of
68. For a critical assessment of this literature and a more de-
tailed analysis of AI impact on teacher tasks, see Tuomi (2018a).
the existing production technology, and
their content remained relatively stable as
large-scale industrial production relied on
standardization and automation. In this
sense, the car, for example, generates the
skills of vehicle maintenance engineer, and
a computer creates a computer program-
mer (Tuomi, 2022). More generally, for each
technical device, and the tools that are used
to make it, it is possible to define the ‘skills’
of making and maintaining the device.
Partly because of this, the current Europe-
an Skills, Competences, Qualifications and
Occupations (ESCO)69 classification now in-
cludes over 13,000 skill definitions.
Although skills and competences are key
concepts in policy development, the in-
creasing importance of non-epistemic
competence components (Tuomi, 2022)
suggests that these concepts need to be
reconsidered. There has been much effort
in defining and characterizing 21st centu-
ry competences and skills (e.g., Bellanca,
2010; EC, 2018; OECD, 2019a), but the
links between learning theory and the con-
cept of competence deserve further study.
In many discussions about skills, the under-
lying model of learning and knowing is what
some educational theorists have called the
‘banking model’ of learning (Freire, 1972).
In this model, knowledge is transferred from
the teacher’s head, a book, or an adaptive
computer system to the student’s mind,
with the resulting skill and mastery as the
outcome. In research on workplace learning
analytics, this has been called the ‘knowl-
edge-acquisition metaphor’ (Ruiz-Calleja et
al., 2021). Knowledge is acquired but not
created. In terms of the MATURE knowledge
maturation model (Maier & Schmidt, 2015),
the associated knowledge is structural in-
69. https://esco.ec.europa.eu/
52
SCIENCE FOR POLICY REPORT
stead of emergent.70 This model of learning
that focuses on formalized, standardized,
and culturally established knowledge, leads
to a relatively static view on skills, at the
same time allowing these skills to be de-
fined, categorized, and classified. There are
now many commercial AI-based tools that
aim to support people in reflecting on their
skills, locating skill gaps in the labour mar-
ket, and guiding learners towards socially
and individually useful career goals.71 Skills
data extracted from online job advertise-
ments and other labour market data are
now widely used to generate evidence for
policymakers. As Buckingham Shum et al.
(2022) note, these tools typically rely on
predefined skills vocabularies.
Within constructivist models of learning,
detailed domain-specific skill definitions, in
general, might not always make sense. At
the level of individuals, accommodation to
the external reality plays an important role
in Piaget’s theory of learning, but Piaget
also emphasized that knowledge is a rela-
tion between the knower and the observed
reality. Learning changes both the learn-
er and the observed environment (Furth,
1981). In such a Piagetian constructivist
view, the prototypical concept of ‘skill’ is
just a passive reflection of a given reality, in
this case fixed by external social and indus-
trial interests. The prototypical concept of
skill can then be understood as a static ar-
tefact generated by historically established
methods of collective production. Such a
70. The knowledge maturation model distinguishes emergent
and generally accepted types of knowledge. It suggests that
learning occurs in the continuum between guidance, where
established knowledge is learned by novices, and emergence,
where more expert collaborate in creating knowledge. A similar
distinction underpins the 5-A knowledge creation model in Tuomi
(1999). For a useful review of organizational knowledge creation
initiatives in the EU framework programmes, see Ley (2020).
71. An example of this approach is the CareerBot developed by
HeadAI in collaboration with the Finnish Metropolitan Univer-
sities of Applied Sciences. It uses AI-processed skills data to
help students find personalised study paths (https://headai.com/
careerbot-guides-students-towards-optimal-job-market-fit/)
conceptualization of skills and competenc-
es has limited relevance for learning, un-
derstood as development. As long as work-
place routines and practices do not change,
this conceptualization may be useful. When
innovation and knowledge creation become
important for organizations, or when there
are disruptive technological changes in
the economy, traditional skill-based labour
market policies may become misleading
and less useful than before.
The debate on whether education is for hu-
man development or to produce economi-
cally useful skills is an old one. Large cor-
porations have addressed the proliferation
of skills and related expertise by developing
competence and knowledge management
systems since the mid-90s (Tuomi, 1999),
but it is also well-known that, in parallel, in
empirical surveys business executives state
that they are looking for employees with
generic competences. These include com-
petences in team-working, creativity, com-
munication, problem-solving, and learning.
An important question is how such skills can
be assessed (Martinez-Yarza et al. 2023;
Lucas, 2022) and linked to learning objec-
tives and whether they should be creden-
tialed and certified (Tuomi, 2022).
4.3 Learning analytics and
new forms of assessment
In learning analytics, the use of AI-based
pattern detection will provide new, increas-
ingly advanced ways to support continuous
formative assessment for learners and also
a better understanding of learning process-
es for teachers and researchers. This will
potentially lead to important breakthroughs
in learning sciences. Much of our current
knowledge about learning is based on la-
bour-intensive data collection and self-re-
porting that can only provide limited access
to learning processes. Better understanding
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
53
of learning can be expected to lead to new
architectures for AI-based learning support
systems and new pedagogic and andragog-
ic models.
4.3.1 AI-supported formative
assessment
Assessment plays two quite different roles
in education. Summative assessment is
used to validate and certify learning, and
to compare and sort learners and educa-
tional institutions. Formative assessment,
in contrast, aims to provide feedback to the
learner and improve learning. Formative as-
sessment has been challenging because it
needs to be tailored for the present state
of each learner. While standardized tests
can be used to compare students and the
knowledge they have accumulated, forma-
tive assessment measures individual pro-
gress on oen idiosyncratic learning paths.
Until recently, assessment has been seen as
a natural task for AI-supported automation.
Marking tests and homework represents a
large workload for teachers, and it has com-
monly been expected that AI could reduce
this effort (e.g., Baker & Smith, 2019). Much
of the visible and invisible work of teach-
ers is concerned with formative assess-
ment. Autograders are now widely used for
assessing homework in computer science
classes (e.g., Hsu et al., 2021; Nurminen
et al., 2021), and Automated Short An-
swer Grading (ASAG) systems have gained
increasing attention as data-driven AI has
gained traction (Haller et al., 2022). Com-
puter-based assessment of student es-
says was already expected to be imminent
half-a-century ago (Page, 1966), and, with
some delay, automated essay scoring (AES)
has now become a vast field of research
and a thriving industry (Belgman Klebanov
& Madnani, 2021; Ke & Ng, 2019).
Assessment profoundly shapes pedagogic
practices, the organization of work in educa-
tional institutions, and policy. The data that
are collected in assessment and the types
of learning processes that are assessed in-
fluence both learning and the development
of education. Emerging technologies will
generate new types of data that could be
used in assessment, and both old and new
data sources can be used in novel ways in
the emerging landscape of learning.
Over the last decades, assessment prac-
tices have increasingly been based on
evidence-centred design of assessment
(Mislevy, 2018). This is a data-oriented ap-
proach that has been inspired by object-ori-
ented soware analysis and design meth-
ods (Mislevy et al., 2012). Evidence-centred
design (ECD) starts from domain analysis,
which leads to a domain model that con-
ceptualizes the domain to be assessed. ECD
emphasizes the specification of the logic of
assessment, expressed in an ‘assessment
argument’. The assessment argument de-
scribes what observable behaviours or per-
formances should reveal the constructs of
interest. A ‘conceptual assessment frame-
work’ then makes the linkages between as-
sessment tasks and evidence about profi-
ciency explicit, linking a student model, an
assessment task model, and an evaluation
model. Based on these conceptual struc-
tures, the actual assessment instruments
and test items can then be defined and
delivered to the students. Evidence-centric
assessment therefore defines the types of
evidence and data that are collected to
assess learner proficiency. When learning
occurs on digital platforms, these data can
also be collected on learner activities as a
‘side effect’ (DiCerbo & Behrens, 2014). This
is sometimes called ‘stealth assessment’
(Shute & Ventura, 2013).
In future-oriented visions of education,
adaptive learning systems are oen pre-
54
SCIENCE FOR POLICY REPORT
sented as a technology that can relieve
teachers from the tedious tasks of mark-
ing homework or even at times replace
the teacher making continuous formative
assessment possible. More importantly,
continuous feedback could also improve
learning outcomes. The need to explicitly
design assessment tasks and link them to
progress in learning also makes the objec-
tives of instruction explicit. For example,
knowledge-based intelligent tutoring sys-
tems require explicit models of what counts
as evidence of learning, thus clarifying the
objectives and goals of instruction (Luckin,
2018, p. 121).
A challenge found in this approach is that
system designers oen have a very simpli-
fied understanding of the complexities of
the socio-technical environments where au-
tomation is introduced (Cerratto Pargman
et al., 2023; Selwyn, 2022a). Much of the
research on human-computer interaction
and computer-supported collaborative work
since the late 1980s addressed this chal-
lenge by turning to ethnographic methods
(e.g., Bannon & Bødker, 1991; Nardi, 1995;
Star, 1996; Suchman, 1987). The focus in
these studies was on observing actual work
practices instead of defining them based
on abstract models of what people were
supposed to do. Failures in automating or-
ganizational work processes were frequent-
ly shown to result from inadequate under-
standing of the concrete situations where
people work.
For example, although marking homework
is an onerous task, it is also an important
activity for the teacher to gain insight into
the students’ learning and development.
Homework does not only inform the teacher
about the proficiency of a student; it also
gives the teacher feedback on the stu-
dent’s emotions, attitudes, areas of missing
knowledge and skills, and other contextual
factors. Total automation of grading can
therefore have a negative impact on the
teacher’s capacity to teach. Effective in-
struction requires that also the teacher is
continuously learning, and automation may
inadvertently cut the teacher off from the
learning loop.72
A five-year vision for the future of assess-
ment, outlined by the UK digital technology
and education agency Jisc (2020), suggests
that assessment needs to become more
authentic, accessible and inclusive, appro-
priately automated, continuous, and secure.
Authentic assessment means, for example,
that instead of measuring practical skills
using pen and paper, they are measured
in more realistic contexts, for example by
using simulations with immersive technolo-
gies depending on the subject matter. Sim-
ilarly, project-based assessment can move
away from traditional pen and paper tests
and essay-based assessment and evaluate
the outcomes of the project. Appropriately
automated assessment, according to Jisc,
retains critical elements of student-teach-
er interaction, relieving teachers from the
marking workload and improving the feed-
back students receive. According to Jisc
(2020, p. 17), AI and learning analytics
might be used to provide students with a
personal learning assistant that continu-
ously assesses student progress and helps
in formative development, and it could
make some ‘stop-and-test’ assessment
points redundant.
72. Similarly, the visions of ‘personalised’ learning are oen
excessively individualistic and miss key learning processes.
Personalised learning systems may cut the learner out of social
interactions and relations that can be crucial for educational and
life outcomes. Individualistic learning models oen neglect the
importance of socialization (Biesta, 2015), peer learning (Kim-
brough et al., 2022), and teacher-student relations (Guilherme,
2019). In some countries, education is also important for the
development of social capital, which is reflected, for example,
in the high costs that students and their parents are willing to
pay for access. Nardi et al. (2000), for example, argued that the
motto of the Information Age is: ‘It’s not what you know, but who
you know’.
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
55
A recent critical review of extant AI-sup-
ported assessment approaches suggests
several ways in which AI could change as-
sessment practices (Swiecki et al., 2022).
First, it is possible to reduce the onerous-
ness of assessment by automated as-
sessment construction, AI-assisted peer
assessment, and writing analytics. Second,
AI makes it possible to move from ‘stop-
and-test’ assessment that at best provides
discrete snapshots of task performance,
towards continuous formative assessment.
Third, AI techniques can be used to adjust
assessment tasks to the student’s abilities
and personalize assessment. Fourth, AI can
be used to generate increasingly authentic
assessment situations, for example by us-
ing simulations. AI can also be used to col-
lect and analyse data from authentic and
complex assessments that may be difficult
for humans to assess. Fih, as AI itself is
increasingly being used by the students and
in work tasks, new assessment practices
need to incorporate AI tools in the assess-
ment designs.
Generative AI has been declared to particu-
larly threaten essay-based assessment. As
soon as GPT-3 and ChatGPT were released
to the public several university professors
around the world reported that they were
able to detect AI-produced content in stu-
dent homework. An anonymous poll at
Stanford University, conducted in early Jan-
uary 2023, suggested that almost one-fih
of students had used ChatGPT in their fi-
nal exams (Cu & Hochman, 2023). Several
school districts and universities around the
world quickly banned the use of ChatGPT,
university honour codes and ethical guide-
lines were rapidly revised, professors
changed assessment practices, and tools to
detect and fingerprint AI-produced content
were developed, sparking worries about a
new technological arms race between gen-
erative AI systems and detection tools.
At the same time, it has been pointed out
that ChatGPT makes it urgent for educa-
tors and institu-tions to reimagine their
approaches to assessment, for example
‘replacing exams or other assess-ments
with in-person assessments or altering the
types of questions or exam formats that
are used’. (Sabzalieva & Valentini, 2023,
p. 13). Academics in higher education now
compete for student attention using grades
as both carrot and stick (Carless, 2023).
Against the quick first reaction of banning AI
from schools, it was suggested that teach-
ers should embrace AI, and integrate it into
their teaching. For example, the capability
of ChatGPT to provide fast feedback to the
student while working on projects, instead
of providing conventional feedback aer
the project has been completed, was seen
as an opportunity.
From a learning theory point of view, these
debates on the threats and opportunities of
AI-supported assessment highlight a shi
from summative to formative assessment.
Along the spectrum of assessment, intel-
ligent tutoring systems that rely on fine-
grained learning analytics are in one end,
with almost continuous observation of stu-
dent behaviour and progress. At the other
end is high-stakes testing that aims to re-
cord learning outcomes. In the middle, there
are now many suggestions that generative
AI systems could be used as interactive
learning companions that provide feedback
to the learner (Sabzalieva & Valentini, 2023,
p. 9). Some adaptive learning environments
could also be found in this middle area.
As assessment is a major factor in shap-
ing student learning, the above examples
suggest that AI will have a profound impact
on education in the coming years. To avoid
automating outdated assessment practices
using AI, it will be important to review and
reconsider the assumptions that underpin
existing assessment practices. As Swiecki
56
SCIENCE FOR POLICY REPORT
et al. (2022, p. 7) point out, AI-enabled as-
sessment is not a neutral site where any
form of learning will be detected and as-
sessed. AI-enabled assessment will inev-
itably codify specific cultural, disciplinary,
and individual norms, value systems and
knowledge hierarchies. AI-enabled assess-
ment may remain limited in its capacity to
recognize learning that is creative and in-
novative simply because data-driven AI re-
lies on detecting historical patterns (Tuomi,
2018a). The proposed pedagogical uses
of generative AI, however, suggest that AI
can become an important tool in re-design-
ing assessment. Further research is there-
fore needed to understand what should be
assessed, how assessment data should
be collected and used, and how the new
emerging forms of assessment could bene-
fit learning and education.
From a technical point of view, automated
essay scoring will probably make fast pro-
gress in the near future. This is because re-
inforcement learning from human feedback
(RLHF), which is currently used to train large
language models, can also be used to fine-
tune language models for improved classi-
fication of essays. For example, criteria such
as argument persuasiveness, specificity,
and evidence (Ke & Ng, 2019, p. 6306) can
easily be assessed by competent humans,
and the results can then be used to fine-
tune language models for essay grading.
Beyond scoring, such models can also give
the student detailed feedback on the rea-
sons why the score is what it is and how the
essay could be improved.73 As mentioned
before, policy action will play a key role in
shaping the use of AI, also in assessment,
with the proposed EU AI Act being a pioneer
initiative in the field.
73. In reverse, improved essay grading can, of course, also be
used to train generative AI systems such as ChatGPT, so that
they produce higher-quality outputs. Improved assessment mod-
els, therefore, can also be highly important for the development
of generative AI technologies.
4.3.2 Ethics of data-driven
assessment
Continuous assessment generates large
amounts of personal data. The develop-
ment of new assessment methods there-
fore needs to address existing and emerg-
ing regulations and data policies. Ethics
of learning analytics has extensively been
discussed since the emergence of the field
(e.g., Buckingham Shum, 2019; Buckingham
Shum & Luckin, 2019; Hakimi et al., 2021;
Pargman & McGrath, 2021; Prinsloo &
Slade, 2017; Tzimas & Demetriadis, 2021)
but data-driven assessment deserves fur-
ther attention as multimodal data streams,
data-driven AI, and XR become combined in
the Next Internet.
With the advances of AI, machine learn-
ing and other emerging technologies, so-
cio-emotional human-centred skills, such
as leadership, communication, collabora-
tion, empathy, etc, are becoming increas-
ingly important and essential. LifeComp74,
a European framework developed by the
EC, has contributed to this area of research
by providing a shared understanding of
personal and social development compe-
tencies and has become a framework ref-
erence in supporting the development of
life competencies as part of the curriculum
(Sala, et al., 2020). More research is still
needed to understand how to frame the
learning objectives around these skills and
how they can be supported and assessed
(Martinez-Yarza et al., 2023; Lucas, 2022).
A temporary moratorium on using data on
these skills for machine learning systems
has, therefore, been proposed until the so-
cial, educational, and ethical implications
are clarified (Tuomi, 2022).
74. https://publications.jrc.ec.europa.eu/repository/bitstream/
JRC120911/lcreport_290620-online.pdf
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
57
In technical terms, collecting data on ‘so
skills’ looks like a very similar problem to
collecting data on student knowledge and
other skills. A more detailed look at so
skills, however, also highlights the point
that students are different, and addressing
these differences is most oen not feasi-
ble through simple educational interven-
tions. In some European countries, policies
try to respond to these differences through
classroom-level differentiated education. In
others, educational paths are differentiated
based on expected educational achieve-
ment. AI systems introduce new ways to
address the need to differentiate and per-
sonalize education, but as many socially
and economically important skills and com-
petences are linked to personality charac-
teristics, in the future the appropriate unit
of analysis for research on AIED and other
educational technologies could be a tech-
nology-augmented learner (Tuomi, 2023a).
For example, technology can potentially be
used in support of neurodiversity and to
augment capabilities in areas that are im-
portant for learning and individual well-be-
ing. We can use spell-checkers to support
dyslexic students and text-to-speech tech-
nologies to support vision impaired learners.
In a similar way, AI systems could be used
to augment human learning capabilities
when and where this is needed. It can be
expected that as 21st century competences
become more important, future AIED sys-
tems could benefit from such an augmen-
tation approach.
Beyond the ethics of AI in education, where
developmental considerations are central,
education administrators also need to ad-
dress AI governance from a regulatory point
of view. At present, it is not known how the
developmental and regulatory concerns
should be combined and connected. This
linking will be of special importance in edu-
cational contexts, but it should also inform
more general AI-related policy develop-
ment. As suggested above, the develop-
ment of the broader AI ecosystem will also
shape the emerging educational landscape.
Better understanding of the dynamics of
ecosystem development is therefore also
important for education policy development.
This, in turn, could link industrial policy ‒ un-
derstood here as the intentional structuring
of markets and innovation processes ‒ and
educational policy in novel ways.
58
SCIENCE FOR POLICY REPORT
Learning is at the centre of the ongoing
transition from the Industrial Age towards
what is commonly called the Knowledge
Society. It should not come as a surprise
that emerging technologies that shape
learning, education, and knowledge crea-
tion have broad impacts beyond education
itself. This report anticipates that a diverse
range of technological innovations, which
are currently at different stages of develop-
ment, will complement each other and fuse
digital, material, cognitive, and social reali-
ties in ways that we have not seen before.
This calls for an unprecedented interlinking
of educational, digital, environmental, and
industrial policy, especially the regulation
around the structuring of markets and in-
novation processes. The future of education
itself needs to be viewed in this broader
context of social, economic, and technical
change.
This report looks at technology and soci-
ety mainly from a European perspective,
focusing on an area where access to new
technologies is too oen taken for granted.
However, it should be noted that access is
not universal within Europe and that, in any
case, access itself is not enough to ensure
competence, as we have seen in various
studies during and following the Covid-19
pandemic (Cachia et al. 2021). Equity and
access to high quality education is a prior-
ity for a transitioning Europe as it adapts
education and training systems to the digi-
tal age, as outlined in the Digital Education
Action Plan (2021-2027) (European Com-
mission 2020) and two Council Recom-
mendations published this year aimed at
ensuring universal access to inclusive and
high-quality digital education and training
(European Commission 2023a) as well as
addressing the growing demands for digital
skills triggered by the digital transforma-
tion of society and the economy (European
Commission 2023b).
Digital technologies have a material basis
and a significant societal and environmen-
tal impact that it is too easily overlooked.
Climate change, demographic transitions,
environmental concerns, and, for exam-
ple, the growth of mental health problems
among young people will shape the emerg-
ing landscape of education. Responding to
those challenges, in 2022, the Council of
the European Union (2022c) adopted new
conclusions on the need to support well-be-
ing in digital education for both students
and educators and the European Com-
mission published a European Sustainable
competence framework identifying a set
of competencies that would help learners
develop knowledge, skills and attitudes
that promote learning on environmental
sustainability (Bianchi et al. 2022). The
challenges and opportunities of the twin
digital and green transitions will require ed-
ucational systems to adapt the curriculum
and prioritise the development of new key
competences.
New technical solutions are easy to adopt
when they address pre-existing needs and
support established practices. In this re-
gard, technological innovation is oen a
conservative force, meaning that it is pri-
marily concerned with the problems of the
5 Conclusions and
policy implications
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
59
past. It is therefore important to critically
assess the potential of emerging and new
technologies instead of hard-wiring obso-
lete practices. The future does not exist yet,
and there cannot be facts about yet-to-be
futures that could justify policy choices.
This is one of the reasons why policy de-
velopment needs experimentation and ex-
ploration of imagined futures. These can be
supported by knowledge and argument, but
it should be noted that choices are always
rooted on values. Visions of technological
futures, therefore, should be accompanied
with explicit discussion of ethics, which is a
major priority in Europe. This report noted
that a capability-based approach may be a
useful starting point as it links developmen-
tal considerations central for education with
the idea that individuals have varying ca-
pacities to translate resources to well-be-
ing. Technology plays a central role in this
translation.
This report has suggested many areas where
policy-related research and initiatives could
be useful. Some key observations and re-
sults that also summarize potential areas
for future work are highlighted below.
1. The Next Internet
The Next Internet will profoundly change
the social and cognitive infrastructures of
knowing, learning, and action. Over the com-
ing decade, technological developments,
including 6G networks, immersive technolo-
gies, and new distributed data and process-
ing architectures will fuse digital, material,
cognitive, and social realities in ways that
we have not seen before. This creates op-
portunities to transform educational prac-
tices and institutions. The consequences of
this transition are only very superficially un-
derstood today. Future-oriented policy-re-
lated research and development of edu-
cational use cases for this emerging world
would be needed to understand the impli-
cations of such developments. The ‘sen-
sorization’of wireless networks and their
emerging capability to sense humans, ac-
tion, and material objects in physical space
links digital networks with the material
world in radically new ways. Decentralized
identifiers and distributed data structures
will be important for the Next Internet, and
the impact of these apparently technical
developments on learning, teaching, and
education deserves further research for a
better understanding of what policy action
would be needed.
2. Data
The amount of data on education and
learning is growing fast. This provides op-
portunities for new pedagogic and an-
dragogic approaches, as well as for the
governance and management of education.
Datafication of education comes with ma-
jor ethical challenges and the development
of evidence-based learning analytics and AI
systems for education and learning requires
careful balancing of opportunities and risks.
Partly due to historical reasons, data on
learning is now very fragmented across the
Member States, and fine-grained data on
learning processes have mainly been pro-
duced in isolated research projects and by
large global platform providers, with limited
visibility to educational stakeholders.
Although learning and education are cen-
tral to the ongoing social, technical, and
economic transition, data on learning and
education has not yet received the level of
attention it deserves, despite data being
crucial assets that may enhance our un-
derstanding of the emerging landscape of
education and for policy development more
generally. The European Data Spaces could
play a key role in this regard, by addressing
learning and education as strategically im-
60
SCIENCE FOR POLICY REPORT
portant elements. Moreover, regular moni-
toring, evaluation, and assessment of en-
abling factors of digital education will also
produce significant data on what is working
and what is not, given most Member States
still lack adequate structures to gather such
data in some places (Cosgrove et al, 2023).
In any case, while benefitting from learning
and education data, it is essential to ensure
the protection of children’s rights and well-
being (European Commission, 2021d).
3. Human-AI interaction in
learning processes
AI is already influencing many educational
processes and practices, with important im-
plications for teaching, learning and assess-
ment. Recent developments in generative
AI suggest that agency can be distributed
between AI-systems and human learners.
Dynamic distribution of agency between
learners and teachers is a key element in
many influential learning theories, but it
should be reinterpreted in the emerging
technological context. Beyond theoretical
interest, new models of technology-ena-
bled learning and teaching could also form
the basis for educational innovation and in-
novative designs of digital learning technol-
ogies. While it is important that the sector
is able to allow innovation to take place,
design and development of such models
should also be regulated to ensure and pro-
tect children’s safety, security, privacy and
ownership (Cachia, et al, 2020).
4. Skills
As many as two-thirds of existing jobs will
be exposed to AI-based automation in the
coming years. This will generate a huge
impact in education and vocational train-
ing. The resulting rapid changes in the la-
bour-market will make the traditional con-
cept of skill inadequate, both in education
and in labour-market contexts. This makes
us question whether a broader understand-
ing of competences is needed for curricu-
lum development. Generative AI will refine
digital skills, prompting policy makers to
reconsider what is meant by digital skills.
Transversal skills and human-centred com-
petences, such as creativity, problem-solv-
ing, and socio-emotional skills are a result
from the interaction between individual and
contextual factors and not always easy to
be replaced by technology. AI systems could
play a vital role as an assistive technology,
in support of a shi in education towards
the development of skills that has been
traditionally out of the formal curriculum.
This makes us question whether a redefini-
tion of digital skills is required, especially if
we are to reach the target set by the Digital
Compass (European Commission 2021a)
and the European Pillar of Social rights Ac-
tion Plan (European Commission 2021b) to
have 80% of adults with basic digital skills
and 20 million ICT specialists employed in
the EU, with more participation by women.
5. Assessment
Assessment practices shape education
and learning in fundamental ways. AI and
learning analytics potentially enable new
authentic, continuous, and learning-orient-
ed methods of assessment. The emerging
technologies discussed in this report could
contribute to a decrease in the importance
of summative assessment and ‘teaching
for test’,as new digitally-mediated forma-
tive and continuous assessment approach-
es become more widely used. The devel-
opment of these formative assessment
methods, however, will greatly depend
on existing and proposed regulations and
available ICT infrastructures. The current
technological landscape has also prompt-
ed debates on the relevance of assessment
and the need for a better understanding of
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
61
how it will influence assessment in differ-
ent disciplines, as the impact will probably
vary widely across different subject areas.
Developers in this area have themselves
highlighted the need for regulation. In this
respect, more research is needed to under-
stand how AI, specifically generative AI, is
affecting assessment and how it will affect
transversal skills like creativity, critical think-
ing and problem solving. Beyond how edu-
cational organizations and systems assess
students’ achievement of intended learning
outcomes, there are also relevant changes
in relation to how learners can evidence the
expertise they have developed with a high
level of granularity. Digital credentials are
not only replacing traditional paper-based
certificates, but they also enable learners
to gather evidences of their learning as
they progress through their own learning
paths. Micro-credentials may support life-
long learning by certifying the learning out-
comes of short-term learning experiences,
but in order to reach their full potential it
is essential to implement common stand-
ards to ensure their quality, transparency,
cross-border comparability, recognition and
portability; as established by the recom-
mendation defining a European approach
to micro-credentials adopted by the Council
of the European Union (2022a).
6. The new trivium
Trivium and quadrivium formed the basis
for education in the medieval Europe. The
associated seven liberal arts ‒ grammar,
logic, rhetoric, arithmetic, geometry, music,
and astronomy ‒ were considered to be
the foundational thinking skills. In the age
of generative AI, core skills such as writing,
mathematics, communication, and knowl-
edge about the world, need to be rethought.
For example, the ‘writing synthesizer’hy-
pothesis discussed in this report suggests
that writing could soon involve genera-
tive AI systems in novel ways in the writ-
ing process, re- and deconstructing writing,
and redefining what we mean by it. AI sys-
tems are already used in problem-solving,
mathematics, writing human language and
computer programs, and in visual arts and
in music analysis and production. To under-
stand the impact on educational practices,
well-elaborated use cases are needed that
show how generative AI could be used in
various educational settings, and what poli-
cy implications such uses would have.
62
SCIENCE FOR POLICY REPORT
Akyildiz, I. F., Su, W., Sankarasubramaniam,
Y., & Cayirci, E. (2002). Wireless sensor net-
works: A survey. Computer Networks, 38(4),
393–422. https://doi.org/10.1016/S1389-
1286(01)00302-4
Alammary, A., Alhazmi, S., Almasri, M., &
Gillani, S. (2019). Blockchain-based appli-
cations in education: A systematic review.
Applied Sciences, 9(12), Article 12. https://
doi.org/10.3390/app9122400
Argyris, C., & Schön, D. (1978). Organiza-
tional learning. Addison-Wesley.
Autor, D. H., Levy, F., & Murnane, R. J.
(2003). The Skill Content of Recent Tech-
nological Change: An Empirical Explo-
ration. The Quarterly Journal of Eco-
nomics, 118(4), 1279–1333. https://doi.
org/10.1162/003355303322552801
Azevedo, R., Mudrick, N. V., Taub, M.,
& Bradbury, A. E. (2019). Self-regula-
tion in computer-assisted learning sys-
tems. In The Cambridge Handbook of
Cognition and Education (pp. 587–618).
Cambridge University Press. https://doi.
org/10.1017/9781108235631.024
Bai, J. Y. H., Zawacki-Richter, O., & Müskens,
W. (2022). Developing strategic scenar-
ios for artificial intelligence applications
in higher education. 47–70. https://doi.
org/10.5281/zenodo.7488343
Bai, Y., Kadavath, S., Kundu, S., Askell,
A., … Kaplan, J. (2022). Constitution-
al AI: Harmlessness from AI Feedback
(arXiv:2212.08073). arXiv. https://doi.
org/10.48550/arXiv.2212.08073
Bailenson, J. (2018). Experience on De-
mand: What Virtual Reality Is, How It Works,
and What It Can Do. W.W. Norton & Com-
pany. https://wwnorton.com/books/experi-
ence-on-demand/
Baker, T., & Smith, L. (2019). Educ-AI-tion
Rebooted? Exploring the future of artificial
intelligence in schools and colleges. NESTA.
Ball, J., & Savin-Baden, M. (2022). Post-
digital Learning for a Changing Higher Ed-
ucation. Postdigital Science and Education,
4(3), 753–771. https://doi.org/10.1007/
s42438-022-00307-2
Ball, M. (2023, January 23). Why VR/AR
Gets Farther Away as It Comes Into Focus.
MatthewBall.Vc.w.matthewball.vc/all/why-
vrar-gets-farther-away-as-it-comes-into-
focus
Bannon, L., & Bødker, S. (1991). Beyond the
interface: Encountering artifacts in use. In J.
M. Carroll (Ed.), Designing Interaction: Psy-
chology at the Human-Computer Interface
(pp. 227–253). Cambridge University Press.
Barrows, H.S., & Tamblyn, R.M. (1980). Prob-
lem-based learning: An approach to medi-
cal education. New York: Springer
Belgman Klebanov, B., & Madnani, N.
(2021). Automated Essay Scoring. Springer.
Bellanca, J. A. (2010). 21st Century Skills:
Rethinking How Students Learn. Solution
Tree Press.
Bender, E. M., Gebru, T., McMillan-Major, A.,
& Shmitchell, S. (2021). On the dangers of
stochastic parrots: Can language models
be too big? Proceedings of the 2021 ACM
Conference on Fairness, Accountability,
and Transparency, 610–623. https://doi.
org/10.1145/3442188.3445922
References
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
63
Beullens, W., D’Anvers, J.-P., Hülsing, A.,
Lange, T., … Smart, N. P. (2021). Post-Quan-
tum Cryptography: Current state and
quantum mitigation. ENISA. https://doi.
org/10.2824/92307
Bianchi, G., Pisiotis, U. and Cabrera Giraldez,
M. (2022). GreenComp The European sus-
tainability competence framework, Punie,
Y. and Bacigalupo, M. editor(s), EUR 30955
EN, Publications Office of the European
Union, Luxembourg. doi:10.2760/821058,
JRC128040.
Biesta, G. (2010). Good Education in an Age
of Measurement: Ethics, Politics, Democra-
cy. Routledge.
Biesta, G. (2015). What is education for?
On good education, teacher judgement,
and educational professionalism. European
Journal of Education, 50(1), 75–87. https://
doi.org/10.1111/ejed.12109
Blaskó, Z., Costa, P. da, & Schnepf, S. V.
(2022). Learning losses and education-
al inequalities in Europe: Mapping the
potential consequences of the COV-
ID-19 crisis. Journal of European So-
cial Policy, 32(4), 361–375. https://doi.
org/10.1177/09589287221091687
Blikstein, P., Zheng, Y., & Zhou, K. Z. (2022).
Ceci n’est pas une école: The discourses of
artificial intelligence in education through
the lens of semiotic analytics. European
Journal of Education, 57(4), 571–583.
Bosch, J. M., Tangi, L., & Burian, P. (2022).
European Landscape on the Use of Block-
chain Technology in the Public Sector (JRC
Technical Report). European Commission.
https://doi.org/10.2760/18007
Bowker, G., & Star, S. L. (1999). Sorting
Things Out: Classification and its Conse-
quences. The MIT Press.
Braghieri, L., Levy, R. & Makarin, A. So-
cial Media and Mental Health (July 28,
2022). Available at SSRN: https://ssrn.
com/abstract=3919760 or http://dx.doi.
org/10.2139/ssrn.3919760
Briggs, J., & Kodnani, D. (2023). The Poten-
tially Large Effects of Artificial Intelligence
on Economic Growth [Global Economic Ana-
lyst]. Goldman Sachs.
Brod, G., Kucirkova, N., Shepherd, J., Jolles, D.,
& Molenaar, I. (2023). Agency in education-
al technology: Interdisciplinary perspectives
and implications for learning design. Educa-
tional Psychology Review, 35(1), 25. https://
doi.org/10.1007/s10648-023-09749-x
Brown, A. L., Ash, D., Rutherford, M., Nagaga-
wa, K., … Campione, J. C. (1993). Distributed
expertise in the classroom. In Gavriel Salo-
mon (Ed.), Distributed Cognitions: Psycho-
logical and Educational Considerations (pp.
188–228). Cambridge University Press.
Brown, J.S., Collins, A., & Duguid, P. (1989).
Situated cognition and the culture of learn-
ing. Educational Researcher, 18(1), 32-42.
Brown, J. S., & Duguid, P. (2001). Knowledge
and organization: A social-practice perspec-
tive. Organization Science, 12(2), 198–213.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M.,
… Amodei, D. (2020). Language models are
few-shot learners. ArXiv:2005.14165 [Cs].
http://arxiv.org/abs/2005.14165
Bruner, J. (1986). Actual Minds, Possible
Worlds. Harvard University Press.
Brusilovsky, P. (2023). AI in education, learn-
er control, and human-AI collaboration. In-
ternational Journal of Artificial Intelligence
in Education, 33(in press).
Brynjolfsson, E. (2022). The Turing Trap: The
promise & peril of human-like artificial in-
telligence. Daedalus, 151(2), 272–287.
64
SCIENCE FOR POLICY REPORT
Bubeck, S., Chandrasekaran, V., Eldan, R.,
Gehrke, J., … Zhang, Y. (2023). Sparks of
Artificial General Intelligence: Early exper-
iments with GPT-4 (arXiv:2303.12712).
arXiv. https://doi.org/10.48550/arX-
iv.2303.12712
Buckingham, D. (2007). Beyond technology:
Children’s learning in the age of digital cul-
ture. Polity.
Buckingham Shum, S. (2019). Critical data
studies, abstraction & learning analytics: Ed-
itorial to Selwyn’s LAK keynote and invited
commentaries. Journal of Learning Analyt-
ics, 6(3), Article 3. https://doi.org/10.18608/
jla.2019.63.2
Buckingham Shum, S. (2023, March 24).
The Writing Synth Hypothesis. https://simon.
buckinghamshum.net/2023/03/the-writ-
ing-synth-hypothesis/
Buckingham Shum, S., Littlejohn, A., Kitto,
K., & Crick, R. (2022). Framing Professional
Learning Analytics as Reframing Oneself.
IEEE Transactions on Learning Technologies,
15(5), 634–649. https://doi.org/10.1109/
TLT.2022.3190055
Buckingham Shum, S., & Luckin, R. (2019).
Learning analytics and AI: Politics, pedagogy
and practices. British Journal of Educational
Technology, 50(6), 2785–2793. https://doi.
org/10.1111/bjet.12880
Bull, S. (2020). There are open learner mod-
els about! IEEE Transactions on Learning
Technologies, 13(2), 425–448. https://doi.
org/10.1109/TLT.2020.2978473
Bull, S., Brna, P., & Pain, H. (1995). Ex-
tending the scope of the student model.
User Modeling and User-Adapted Interac-
tion, 5(1), 45–65. https://doi.org/10.1007/
BF01101801
Bull, S., & Kay, J. (2010). Open Learner Mod-
els. In Roger Nkambou, Jacqueline Bour-
deau, & Riichiro Mizoguchi (Eds.), Advances
in Intelligent Tutoring Systems (pp. 301–
322). Springer. https://doi.org/10.1007/978-
3-642-14363-2_15
Cachia, R., Vuorikari, R., Velicu, A., Di Gioia, R.,
& Chaudron, S. (2021). Emergency remote
schooling during COVID-19: A closer look at
European families. European Commission.
Joint Research Centre. https://data.europa.
eu/doi/10.2760/613798
Camilleri, A., Inamorato dos Santos, A., Grech,
A. (2017). Blockchain in education, Euro-
pean Commission, Joint Research Centre.
https://data.europa.eu/doi/10.2760/60649
Carless, D. (2023, April 19). How ChatGPT
can help disrupt assessment overload
[Times Higher Education]. THE Cam-
pus Learn, Share, Connect. https://www.
timeshighereducation.com/campus/how-
chatgpt-can-help-disrupt-assessment-
overload
Caron, M., Touvron, H., Misra, I., Jégou,
H., … Joulin, A. (2021). Emerging proper-
ties in self-supervised vision transform-
ers (arXiv:2104.14294). arXiv. https://doi.
org/10.48550/arXiv.2104.14294
Cascone, Kim. ‘The Aesthetics of Failure:
‘Post-Digital’Tendencies in Contemporary
Computer Music’. Computer Music Journal
24, no. 4 (2000): 12–18.
Castells, M. (1989). The Informational City:
Information Technology, Economic Restruc-
turing, and the Urban-Regional Process.
Basil Blackwell.
Castells, M. (1996). The Information Age:
Economy, Society and Culture: Volume I:
The Rise of the Network Society. Blackwell
Publishers.
Castells, M. (2001). The Internet Galaxy:
Reflections on the Internet, Business, and
Society. Oxford University Press.
On the Futures of Technology in Education:
Emerging Trends and Policy Implications
65
Cerratto Pargman, T., Lindberg, Y., & Buch,
A. (2023). Automation Is coming! Exploring
future(s)-oriented methods in education.
Postdigital Science and Education, 5(1),
171–194 https://doi.org/10.1007/s42438-
022-00349-6
Chang, H.-Y., Binali, T., Liang, J.-C., Chiou,
G.-L., … Tsai, C.-C. (2022). Ten years of
augmented reality in education: A me-
ta-analysis of (quasi-) experimental stud-
ies to investigate the impact. Computers
& Education, 191, 104641. https://doi.
org/10.1016/j.compedu.2022.104641
Chernyshenko, O. S., Kankaraš, M., & Dras-
gow, F. (2018). Social and emotional skills
for student success and well-being: Con-
ceptual framework for the OECD study on
social and emotional skills. OECD. https://
doi.org/10.1787/db1d8e59-en
Christopoulos, A., Mystakidis, S., Pellas, N., &
Laakso, M.-J. (2021). ARLEAN: An augment-
ed reality learning analytics ethical frame-
work. Computers, 10(8), Article 8. https://
doi.org/10.3390/computers10080092
Coghlan, S., Miller, T., & Paterson, J. (2021).
Good Proctor or ‘Big Brother”? Ethics of on-
line exam supervision technologies. Philos-
ophy & Technology. https://doi.org/10.1007/
s13347-021-00476-1
Cohen, J. E. (2019). Between Truth and
Power: The Legal Constructions of In-
formational Capitalism. Oxford Uni-
versity Press. https://doi.org/10.1093/
oso/9780190246693.001.0001
Cole, M. (1996). Cultural Psychology: A Once
and Future Discipline. The Belknap Press of
Harvard University Press.
Cosgrove, J., Tsotsou, I., Cachia, R., Centeno,
C., Sala, A., & Punie, Y. (2023). Themat-
ic analysis in support of the Staff Working
Document for Council Recommendations:
Improving the provision of digital skills in
education and training, and Digital educa-
tion – enabling factors for success. In Eu-
ropean Commission (2023, April), Commis-
sion Staff Working Document (SWD(2023)
205), Annex 3: Thematic analysis in sup-
port of the Staff Working Document (pp.
113-197). Strasbourg: European Commis-
sion. Retrieved June 2023 from EUR-Lex -
52023SC0205 - EN - EUR-Lex (europa.eu)
Council of the European Union. (2022a).
Council Recommendation of 19 June 2022
on a European approach to micro-creden-
tials for lifelong learning and employability.
Official Journal of the European Union, C
243, 10–25.
Council of the European Union. (2022b).
Proposal for a Regulation of the European
Parliament and of the Council laying down
harmonised rules on artificial intelligence
(Artificial Intelligence Act) and amending
certain Union legislative acts. https://data.
consilium.europa.eu/doc/document/ST-
14954-2022-INIT/en/pdf
Council of the European Union (2022c).
Council of the European Union. (2018).
Council Recommendation on Key Com-
petences for Lifelong Learning. Re-
trieved from https://ec.europa.eu/
education/education-in-the-eu/coun-
cil-recommendation-on-key-competenc-
es-for-lifelong-learning_en
Cox, A. M. (2021). Exploring the impact of
Artificial Intelligence and robots on higher
education through literature-based design
fictions. International Journal of Education-
al Technology in Higher Education, 18(1),
3. https://doi.org/10.1186/s41239-020-
00237-8
Crang, Mike, & Thri, Nigel (Eds.). (2000).
Thinking Space. Routledge.Crawford, K.
(2021). The Atlas of AI: Power, Politics, and
the Planetary Costs of Artificial Intelligence.
Yale University Press.