Conference PaperPDF Available

Definitional Foundations for Intelligent Systems, Part I: Quality Criteria for Definitions of Intelligence



We posit that the lack of consensus definitions of (machine or artificial) intelligence might be affected by the lack of knowledge of conceptual analysis and other well-investigated theories. Acute contextualization of the concepts that are defined may also be an issue. Accordingly, in this two-part paper, we review some basic concepts from across research fields on how to explicate a definition. In Part I we propose 30 quality criteria for definitions that shall serve as guidelines for well-defined definitions of any concept. The quality criteria may allow for both better insights into definitions and a wider understanding of the current discourse on AI. In Part II we provide basic terminology on definitions and an iterative process to guide the construction of robust definitions by considering the quality criteria introduced in Part I. Our central goal is twofold: we want to facilitate understanding across fields and inform different stakeholders from industry, academia, legal and governments, among others, by contributing to the formal foundations on elucidating "good and robust definitions" for AI.
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
Definitional Foundations for Intelligent Systems,
Part I: Quality Criteria for Definitions of Intelligence
Dagmar Monett1 and Colin W. P. Lewis2
1 Computer Science Dept., Berlin School of Economics and Law, Germany, and
1, 2
We posit that the lack of consensus definitions of (machine or artificial) intelligence might be
affected by the lack of knowledge of conceptual analysis and other well-investigated theories.
Acute contextualization of the concepts that are defined may also be an issue. Accordingly, in
this two-part paper, we review some basic concepts from across research fields on how to
explicate a definition. In Part I we propose 30 quality criteria for definitions that shall serve as
guidelines for well-defined definitions of any concept. The quality criteria may allow for both
better insights into definitions and a wider understanding of the current discourse on AI. In Part
II we provide basic terminology on definitions and an iterative process to guide the construction
of robust definitions by considering the quality criteria introduced in Part I. Our central goal is
twofold: we want to facilitate understanding across fields and inform different stakeholders
from industry, academia, legal and governments, among others, by contributing to the formal
foundations on elucidating “good and robust definitions” for AI.
Keywords: Artificial Intelligence, definitions, intelligence, quality criteria
Main Conference Topic: Robotization, Engineering and Artificial Intelligence
Artificial Intelligence (AI) is a term that has been used more loosely than defined.
Moreover, intelligence is a fundamental concept that does not have a consensus definition in
any of the concerned fields of study. This is true of disciplines such as Computer Science,
Psychology, Education and the relevant Cognitive Science domains, to say nothing of
intersecting or cross domain research fields where a consensus definition may be even more
pressing. This is also true of concepts like fairness, transparency, model interpretability,
privacy, and other related terminology used around intelligent systems. As a direct
consequence, the lack of overarching definitions is having both positive and negative
implications for the design, development, use, and regulation of those systems. If a concept is
defined well, there is ground for a better understanding and further use of the concept. Yet,
laypeople and concerned parties are being presented with an overwhelming variety of
definitions for the same concept that propitiate more confusion than clarification.
Defining a concept might be a very complex problem. Although recent work has tried to
bring clarity on defining the most important concept of the AI field (for instance, see [18]),
still, there is currently no overarching definition of machine or artificial intelligence.
Furthermore, other equally important concepts that are essential for AI systems remain ill-
defined. This is true of the concepts of model interpretability [13], fairness [5, 27], privacy,
transparency, and many others for which multiple definitions have been provided that, far from
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
clarify, add to a plethora of misinterpretations and misunderstandings about their nature.
Whilst, in Pei Wang’s words [30], “[i]t is just normal for every researcher to believe they have
the best idea [on defining a concept], so we cannot expect some consensus to be achieved
soon,” some possible reasons for not having convincing definitions could be traced back to a
lack of knowledge of conceptual analysis and other well-investigated theories. Acute
contextualization of the concepts that are defined may also be a big issue. We want to change
this unfortunate situation. If we truly want to address fundamental questions concerning the
design, development, deployment, use, and even regulation of AI systems, then we need to
consider definitional foundations regarding the concepts these systems deal with.
Most works on defining a definition are limited to a very general understanding of what
a definition is, thereby providing basic criteria it should fulfill in a certain context, but fail to
consider other criteria that might be equally (or more) relevant in other contexts, i.e., some
works focus on very specific fields and shortlist some criteria but don’t refer to other criteria
that might be of importance. Similarly as it has been the case for engineers when writing
definitions of terms [26], there is a general neglect among computer scientists in general and
AI practitioners in particular of a thorough consideration of conceptual analysis for definitions
of intelligence and other related terms. The common stance is either an acute contextualization
of the new definition that is suggested depending on the concrete application at hand, or the
diminishing of the importance to define the concept at all, two extremes that should actually
be avoided, as called out by Aaron Sloman [23] more than three decades ago. More on
conceptual analysis will be introduced in Part II of this paper.
It is the central goal of the authors of this paper to raise awareness of the role well-defined
definitions play in the study and understanding of any concept. This should be of special
interest in government, legal, and social issues that deal with intelligent systems and the
consequences of their use, deployment, and regulation, for instance. In particular, basic
terminology and quality criteria for definitions are addressed here, with the hope they will
finally find the place they should permanently belong to: the AI and other concerned fields and
audiences dedicated to the conceptualization, development, validation, evaluation, and use of
systems that simulate or show any intelligent thinking or behavior. Only after a theoretical
framework of intelligence is thoroughly defined and understood, well-defined definitions
included, will we be able to move into the next phase: in other words, the acceptance that, and
paraphrasing Raworth [21], as humanity’s context, values, and aims continually evolve, so too
might be the way we define AI and advance the field.1
Our thesis is that a determining step in any attempt to understand intelligent systems
better must start with defining intelligence well. After all, how can we build truly artificially
intelligent machines and understand their behavior if we cannot accurately define what
intelligent abilities should be considered, and how they should be implemented and measured?
Furthermore, defining intelligence, especially machine or artificial intelligence, is not an
exclusive task done by AI and other related practitioners anymore: it turns out that even news
agencies are looking for “their” sound definitions of AI [2]. Unfortunately, there is no
widespread agreed consensus on a definition of intelligence nor of AI.
1 To support this, we have also created the Defintly App, available at and with provisional
password agisiorg. In its initial version, it contains most of the content presented in this two-part paper and
assists users when defining concepts in any field.
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
Sometimes researchers and practitioners deal with concepts that have previously been
defined in the literature or that are common knowledge in a certain domain. Other times they
find themselves in situations where they are the ones that define new concepts. However,
questions like “What is a definition?” “How should a definition be defined?” “What should be
considered when defining a definition?” or “What is a good definition?” remain unanswered in
most situations. This is the case in intelligence research in other disciplines: several definitions
of intelligence have been proposed over the years, but, as indicated by Hunt and Jaeggi in [8],
“after 100 years of research, the definition of the field is still inadequate.” Furthermore,
questions such as: “What is a good definition of intelligence?” or “How should a definition of
(machine) intelligence be defined?” are almost never examined in the scientific literature nor a
systematic conceptual analysis has been performed.
We want to bridge topics from non-technical areas that have been neglected by the AI
community for too long, topics that are not taught in Computer Science courses, and that AI
researchers and practitioners do not consider (as they should, in our opinion) before providing
a new definition of a concept, say a working definition, a dictionary definition, or of any other
type. We are living in a time where defining what AI is and what it is not, for example, is
becoming a crucial concern of stakeholders outside the AI community, but part of the AI
community prefers to avoid that discussion. We are of the opinion this must change.
With this two-part paper we intend to give answers to these and other related questions
and to help bridge diverse research fields on defining definitions. Along with previous work
that analyzes cognitive biases in experts’ written opinions about definitions of intelligence [16],
we also seek to understand why finding a “concordant” definition might be difficult to
accomplish although not impossible. Paraphrasing Kate Raworth’s approach [21] when she
calls for rethinking economics as a science, defining machine intelligence is not about finding
the perfect definition,2 because it does not exist yet and maybe will not; it’s about establishing
a definition that reflects the context we face, the values we hold, and the aims we have in the
Quality Criteria for Definitions
The quality criteria that are introduced in what follows focus on the main characteristics
or properties a definition should entail in order to be considered a good definition. The purport
originates from quality criteria that are crucial for both software specifications and individual
software requirements and what they should meet in order to be considered “of high quality”
[22]. Those are very particular to the field of Software Engineering, but similar high-quality
descriptions are desirable for definitions.
Some of the quality criteria for definitions presented in this paper were collected after an
exhaustive examination of the literature on definitions. Legg and Hutter [12] and Wang [29],
for example, analyze the different properties or requirements a definition of intelligence should
have or fulfill. Other quality criteria are developed by the authors of this paper. And others are
the result of a thorough, qualitative analysis of the responses to the AGISI research survey on
defining machine intelligence [17]. Respondents were asked to rate their level of agreement
with definitions of human and machine intelligence, as well as to justify their decisions. This
resulted in a corpus with more than four thousands opinions or arguments for and against
2 However, sloppy definitions mislead the public’s understanding and make cooperation difficult among different
groups of researchers [29], this also being a topic of discussion in other fields [9].
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
definitions.3 Furthermore, respondents were asked to provide their suggested definitions of
human and machine intelligence. Both the particular characteristics of the most agreed upon
definitions from the literature and the respondents’ suggested definitions were also analyzed
and informed the quality criteria presented here.
The majority of the quality criteria that are listed in what follows could be applied to the
definition of any term: with a few exceptions, these quality criteria are not specific to a
definition of intelligence alone; they can be easily adapted to the definition of other concepts.
The quality criteria are intended to serve as guidelines when defining the definition of a
concept. This is why, in searching for a definition of a new concept, or when using or adapting
a definition from other authors to a concrete context, it might be desirable, or even convenient,
to consider the quality criteria presented here. Paraphrasing Chris Rupp [22], if one knows how
a definition is to be defined from the beginning, writing a good definition becomes a more
realistic goal.
Whenever possible, we provide a short explanation of what each criterion means,
together with the corresponding, original literature references that may lead to more
information about that criterion.4 Some important, last clarifications though: not all criteria are
necessary nor sufficient for a definition to be well-defined. In some contexts, contradictions
among criteria might arise or some of the criteria might never be satisfied at all. However, this
doesn’t undermine the importance of establishing sound quality measures for defining
concepts. Furthermore, the fulfillment of some criteria should not be seen as a dichotomy but
as a continuum, a degree of satisfaction that may depend on the context or the field, for
What is a Good Definition?
A good definition of machine, artificial, or human intelligence,5 in the following just
intelligence ...
1. ... is complete. It defines the “what,” the “how,” and the “why,” of intelligence.
2. ... defines the “what,” the thing to be defined. It defines intelligence.
3. ... defines the “how,” the means to reach the thing to be defined. It defines the means
or the ways to reach intelligence.
4. ... defines the “why,” the purpose of the thing to be defined. It defines the purpose of
5. ... is ostensive. It includes (examples of) characteristic properties of intelligence. It
exemplifies the “what,” i.e. it exemplifies cognitive abilities or cognitive functions that
indicate intelligence [7].
6. ... is operational. It includes (examples of) characteristic operations or processes related
to intelligence. It exemplifies the “how,” i.e. it exemplifies behavioral characteristics for
reaching intelligence [7].
7. ... is useful [19]. It has utility [14], e.g. for the society, for the development of systems.
It expresses meaning [10]. It is fruitful [29]. It exemplifies the “why,” i.e. it exemplifies
the purpose of intelligence.
8. ... is unbiased towards any particular culture, species, etc. [12].
3 Other details about the survey, the results obtained, and their analysis are not in the scope of this paper. We
refer the interested reader to the already indicated work [17] for more.
4 Some quality criteria are self-explained. We don’t dive into deep explanations for space reasons.
5 For the purposes of focusing, we refer here to the concept of intelligence for being an essential concept to AI.
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
9. ... is clear, in that it avoids metaphorical, ambiguous language [1], and obscure terms
[15]. It is not defined in terms of other vague concepts [29]. It is clearly written; it is
10. ... is consistent, it includes no contradictory statements.
11. ... is affirmative [1].
12. ... is easy to understand [14]. It is understandable [1] by the general public [19].
13. ... is objective, in that it is a definition that does not include subjective opinions [12].
14. ... is exclusive, in that it includes nothing which is not a part of intelligence [20].
15. ... is comprehensive, in that it omits no essential attribute of the thing to be defined [1];
it omits nothing which is a part of intelligence [20].
16. ... is simple [14, 29]. It is grammatically simple [25].
17. ... is short (perceived and relative to other definitions). It defines the essential with fewer
words as possible.
18. ... is formal. The definition is specified with a high degree of precision and, ideally,
using formal mathematics [12].
19. ... does not repeat the name of the thing to be defined [15]. There is “no vicious circle”
in the definition [25]. I.e. it does not repeat or uses intelligence as part of the definition to
define intelligence.6
20. ... is scientifically valid [19]. It reflects current scientific knowledge [14].
21. ... is original. If googled, for instance, it does not “exist,” it is not a definition “from
someone else” (unless its author explicitly indicates that it is a definition from the
literature). In other words, if it exists, there is no need to re-define it the same way; it
should be cited instead.
22. ... is an intellectually elegant definition [28], a very specially formulated one.
23. ... is measurable [4, 14]. It can be measured.
24. ... is functional, in that the wording used to state what the thing to be defined is is a
function of the language in which it is stated [1, 3]. I.e. it uses the concepts and language
that are common knowledge in the discipline where intelligence is defined.
25. ... is fundamental, in that it does not need to be changed from time to time due to
changing technology and knowledge [12].
26. ... is specific [14] to what it is being defined.
27. ... is general, it can be applied “to everything” [12], e.g. it does not distinguish between
machine, artificial, or human intelligence.
28. ... is widely accepted in the discipline [6].
29. ... has a universal application [1], in that it is not anthropocentric [12] unless its goal is
to define human intelligence explicitly.
30. ... defines intelligence “as a field.”
As stated above, not all these quality criteria are necessary nor sufficient; such a
distinction should be considered and is not addressed in this paper. For example, that a
definition should be intellectually elegant could be thought of as secondary in importance.
However, intellectually elegant definitions shape the common imagination like no others. Why
not aim at an appealing, elegantly written way of using the language to define a concept?
Any importance, or priority, or ranking of some criteria over others is not addressed here
either, for space reasons. Determining their necessity or sufficiency could help. It is probably
better to reflect on these aspects once the context, where the definition of the concept should
6 In the case of the concept of intelligence, though, it could be very difficult to define it withou t generating
definitional circles [24].
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
be applied, is specified. As one respondent writes in her feedback after answering the questions
from the survey on defining intelligence introduced above, “any definition of [artificial
intelligence] is necessarily contextual.” We believe that assessing the quality of a definition is
better done when, ideally, the “what,” the “how,” and the “why” of a definition (see criteria 1
to 4 above) are clear for a particular context. The 30 quality criteria for definitions could guide
the process of searching for a definition in any context.
Some of the quality criteria refer to characteristics or properties that have a common
origin. They could be grouped into more general categories. For example, there are quality
criteria that refer to grammatical aspects of a definition; others, to their meaning or value.
Figure 1 shows a possible hierarchy containing five main categories and their relationships that
group the first 29 quality criteria introduced above.
Figure 1: Quality criteria for definitions of intelligence grouped into five main categories
Other categorizations might be more convenient depending on the concrete application
or context at hand. A single quality criterion could belong to more than one category, as is the
case of a definition being formal (see the category Structure in Figure 1). This quality criterion
could also be grouped under the general category Semantic, when neither the form nor the
structure of the definition is meant but that is formal for example with respect to the
mathematical language that is used. This might be the case of different formal definitions for
the concept of fairness, for instance.
Part I of this paper presented 30 quality criteria for definitions. The list encompasses
different properties that are desirable for a definition to be considered of high quality or well-
defined. The proposed list of quality criteria is not exhaustive. Yet it should serve as a basis for
defining and evaluating definitions of concepts that can support academics and practitioners in
their work.
Quality criteria for definitions allow for a better comparison between definitions. For
example, well-defined definitions of a concept could assist in the creation of tests that measure
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
that concept, help to reach a common understanding about what it means, and, hopefully, avoid
propagating false and misleading information about its relations to other concepts and how do
artificial systems consider it when interacting with humans. It all starts by understanding the
appropriate meaning of the concept and, for doing that, well-defined definitions are essential.
[1] Aristotle (1901). Aristotle’s Posterior Analytics. Oxford: Blackwell.
[2] Beckett, C. (2019). New powers, new responsibilities: A global survey of journalism and
artificial intelligence. Department of Media and Communications, The London School
of Economics and Political Science.
[3] Cassidy, J. R. (1967) Aristotle on definitions. Southern Journal of Philosophy, 5, 110
[4] Chollet, F. (2019). The measure of intelligence. arXiv:1911.01547 [cs.AI].
[5] Chouldechova, A., & Roth A. (2018). The frontiers of fairness in machine learning.
arXiv:1810.08810 [cs.LG].
[6] De Boeck, P. (2013). Intelligence, where to look, where to go? Journal of Intelligence,
1, 524.
[7] Fox, J., Beveridge, M., & Glasspool, D. (2003). Understanding intelligent agents:
analysis and synthesis. AI Communications, 16, 139152.
[8] Hunt, E., & Jaeggi, S. M. (2013). Challenges for research on intelligence. Journal of
Intelligence, 1, 3654.
[9] Knoepfler, P. (2018). Gene editing: sloppy definitions mislead. Nature, 564, 345.
[10] Korzybski, A. (1994). Science and Sanity: An Introduction to Non-Aristotelian Systems
and General Semantics. Fifth edition, Brooklyn, NY: Institute of General Semantics.
[11] Legg, S., & Hutter, M. (2007). A collection of definitions of intelligence. In B. Goertzel
and P. Wang (eds.), Advances in Artificial General Intelligence: Concepts, Architectures
and Algorithms, 157, 1724, UK: IOS Press.
[12] Legg, S., & Hutter, M. (2007). Universal Intelligence: A Definition of Machine
Intelligence. Minds and Machines, 17, 391444.
[13] Lipton, Z. C. (2018). The Mythos of Model Interpretability: In machine learning, the
concept of interpretability is both important and slippery. ACM Queue Machine
Learning, 16, 127.
[14] McPherson, M., Arango, P., Fox, H. Lauver, C., McManus, M., Newacheck, P. W.,
Perrin, J. M., Shonkoff, J. P., & Strickland, B. (1998). A new definition of children with
special health care needs. Pediatrics, 102, 137140.
[15] Mercier, C. (1912). Elements of Logic. Third edition, New York, NY: The Manhattanville
[16] Monett, D., Hoge, L., & Lewis, C. W. P. (2019). Cognitive biases undermine consensus
on definitions of intelligence and limit understanding. In U. Furbach, S. Hölldobler, M.
Ragni, R. Rzepka, C. Schon, J. Vallverdu, and A. Wlodarczyk (eds.), Joint Proceedings
of the Workshops on Linguistic and Cognitive Approaches to Dialog Agents LaCATODA
2019 and on Bridging the Gap Between Human and Automated Reasoning BtG 2019,
2452, 5259, CEUR-WS.
[17] Monett, D., & Lewis, C. W. P. (2018). Getting clarity by defining artificial intelligence
A Survey. In V. C. Müller (Ed.), Philosophy and Theory of Artificial Intelligence 2017,
SAPERE 44, 212214, Berlin: Springer.
The 10th AC 2020 in Prague
Multidisciplinary Academic Conference
[18] Monett, D., Lewis, C. W. P., & Thórisson, K. R. (2020). Special Issue “On Defining
Artificial Intelligence”—Commentaries and Author’s Response. Journal of Artificial
General Intelligence, 11(2), 1100.
[19] Morse, R. M., & Flavin, D. K. (1992). The definition of alcoholism. Journal of the
American Medical Association, 268, 10121014.
[20] Random, R. A. (1958). Intelligence as a science. Studies in Intelligence, 2, 76.
[21] Raworth, K. (2017). Doughnut Economics: Seven Ways to Think Like a 21st-Century
Economist. London, UK: Penguin Random House.
[22] Rupp, Ch. (2014). Requirements-Engineering und -Management: Aus der Praxis von
klassisch bis agil. Sixth edition, Munich: Carl Hanser Verlag.
[23] Sloman, A. (2019). The Computer Revolution In Philosophy: Philosophy, science and
models of mind. Revised, online edition, Sussex: Harvester Press.
[24] Sloman, A. (2019). Family Resemblance vs. Polymorphism. School of Computer
Science, University of Birmingham, UK.
[25] Tarski, A. (1994). Introduction to Logic and to the Methodology of Deductive Sciences.
Fourth edition, New York: Oxford University Press.
[26] Thomas, S. (2005). The importance of being earnest about definitions. In G. Vossen, F.
Leymann, P. Lockemann, & W. Stucky (eds.), Lecture Notes in Informatics, Proceedings
of Datenbanksysteme in Business, Technologie und Web, P65, 560577, Bonn: GI.
[27] Verma, S., & Rubin, J. (2018). Fairness definitions explained. In Proceedings of the 2018
ACM/IEEE International Workshop on Software Fairness, FairWare 2018, pp. 17,
Gothenburg, Sweden.
[28] Walsh, D. (1979). Occam’s razor: A Principle of Intellectual Elegance. American
Philosophical Quarterly, 16, 241244.
[29] Wang, P. (2019). On defining artificial intelligence. Journal of Artificial General
Intelligence, 10, 37.
[30] Wang, P. (2020). On defining artificial intelligenceAuthor’s Response to
commentaries. Journal of Artificial General Intelligence, 11(2), 7386.
Brief biographies of the authors
Dagmar Monett
Prof. Dr. Computer Science (Artificial Intelligence, Software Engineering) with 30+
years of research and teaching experience. Co-founder of the AGI Sentinel Initiative,, dedicated to understanding intelligence in order to build beneficial AI and to
educate society’s stakeholders across disciplines. AI expert at Ms.AI, “Artificial Intelligence
for and with Women.”
Colin W. P. Lewis
Adjunct Professor and international business and policy advisor, specialized in the
social and economic impact of technological change. With specific knowledge of the evolving
conditions of technology for productivity growth, development, labor, (in)equality, and
competitiveness. Part of his work focuses on understanding human intelligence and creating
solutions for Artificial Intelligence.
... Delineating the boundaries of the discourse on intelligence may help in defining and understanding its most discussed concept, as suggested in [11]. Furthermore, better insights into definitions and how to define them has proven to be essential for a better understanding of concepts, intelligence and AI included (see for example [12], [13] and [14] for more on properties of good definitions). Knowing those concepts and related cognitive abilities (like defining, analyzing, understanding, discussing, and comparing definitions of intelligence, among others) is expected for AI researchers and practitioners in the first place. ...
... The annotation scheme referred to in this paper builds upon different works on properties of good definitions some of which were referenced to in Section 1. It uses most of the properties or quality criteria for definitions suggested in [14], which includes a compendium and thorough analysis of the literature on definitions together with their most desirable properties. ...
... From the 30 quality criteria for definitions introduced in [14], 21 were considered for annotating each definition from the Intelligence Corpus. ...
Conference Paper
Full-text available
Delineating the boundaries of the discourse on machine or artificial intelligence (AI) may help in defining and understanding its most discussed concept, the concept of intelligence. Furthermore, better insights into both definitions and how to define them well has proven to be essential for a better understanding of concepts, intelligence included. These and related cognitive abilities (e.g. defining, analyzing, understanding, discussing, and comparing definitions of intelligence, among others) are expected for AI researchers and practitioners in the first place. Yet, they are also central to extending or at least providing the basics of AI literacy to other stakeholders of our society. Intelligent systems are transforming the way we interact with technology, with each other, and with ourselves, and knowing at least what AI or intelligence mean is becoming essential for designing, developing, deploying, using, and even regulating intelligent artefacts. However, defining intelligence has been one of the most controversial and studied challenges of both ancient and modern human thinking. A lack of consensus on what intelligence is has remained almost constant over the centuries. Interested scholars have not come up with a consensus or cross-domain accepted definition of intelligence. Neither in the ancient Eastern nor in the ancient and contemporary Western conceptions of intelligence. Nor in the more recent perspectives from the last 70 years within the field of AI. We argue that a better understanding of contemporary technologies, AI-based but not only, starts with a grounded exposure to their conceptual pillars. These include fundamental concepts like the concept of intelligence, in general, and of AI, in particular. Learners and decision makers at all levels should face them, as well as be able to discuss their importance and limitations critically and in an informed way. For doing that, they must be confronted with definitions of intelligence and understand their meaning well, for instance. If these contents are already part of study programs, the better. It is the main goal of this paper to present how a few hundreds of definitions of intelligence were annotated, i.e. their properties and characteristics systematically analyzed and commented, in order to construct a corpus (i.e. a collection) of definitions of intelligence for further uses in AI and other fields. The work and particular application domain presented here has not yet been considered in the extended work on linguistic annotation. Even though, both the annotation and the data merit special attention, for they deal with the elusive, important concept of intelligence, i.e. with definitions of both human and machine (or artificial) intelligence. Undergraduate Computer Science students carried out the annotation process and several research activities. They were involved in an AI research project led by faculty and included their findings and work as part of their undergraduate student research projects in their last study year. We provide details about how the student research projects were conceived, conducted, and mentored. We also describe the properties or quality criteria that were considered for annotating the definitions from the intelligence corpus.
Conference Paper
Full-text available
Presently, the term "Artificial Intelligence (AI)" is widely used and sometimes loosely in media. Surprisingly, there is not yet any universally accepted definition of AI. To help fill in the gap, this article examines literature and interviews with experts on the meaning of AI. Researchers reviewed 60 papers and interviewed 20 experts in Finland. These experts consisted of professors, chief executive officers, chief technology officers, chief operation officers, and heads of AI. Findings showed that AI is beyond a specific technology, but is rather, a combination of different technologies, systems and devices that use data to perform some complex complicated tasks. Results also revealed that AI has some features that distinguish it from other digital technologies or software. This article paves the way for scholars and practitioners on how AI should be defined and understood.
Full-text available
This article systematically analyzes the problem of defining “artificial intelligence.” It starts by pointing out that a definition influences the path of the research, then establishes four criteria of a good working definition of a notion: being similar to its common usage, drawing a sharp boundary, leading to fruitful research, and as simple as possible. According to these criteria, the representative definitions in the field are analyzed. A new definition is proposed, according to it intelligence means “adaptation with insufficient knowledge and resources.” The implications of this definition are discussed, and it is compared with the other definitions. It is claimed that this definition sheds light on the solution of many existing problems and sets a sound foundation for the field.
Conference Paper
Full-text available
There are several reasons for the lack of a consensus definition of (machine) intelligence. The constantly evolving nature and the interdisciplinarity of the Artificial Intelligence (AI) field, together with a historical polarization around what intelligence means, are among the most widely discussed rationalizations, both within the community and outside it. These factors are aggravated by the presence of cognitive biases in subjective reasoning by experts on the definition of intelligence, as we have found in a recent study of experts' opinions across multiple disciplines. In this paper, we show how different cognitive biases can undermine consensus on defining intelligence, and thus how an understanding of intelligence can be substantially affected by these human traits. We also provide general recommendations for tackling these problems. An understanding of intelligence can be achieved by understanding the limits of both human expressiveness and the current discourse around definitions of intelligence within and across the concerned fields.
Full-text available
Supervised machine-learning models boast remarkable predictive capabilities. But can you trust your model? Will it work in deployment? What else can it tell you about the world? Models should be not only good, but also interpretable, yet the task of interpretation appears underspecified. The academic literature has provided diverse and sometimes non-overlapping motivations for interpretability and has offered myriad techniques for rendering interpretable models. Despite this ambiguity, many authors proclaim their models to be interpretable axiomatically, absent further argument. Problematically, it is not clear what common properties unite these techniques. This article seeks to refine the discourse on interpretability. First it examines the objectives of previous papers addressing interpretability, finding them to be diverse and occasionally discordant. Then, it explores model properties and techniques thought to confer interpretability, identifying transparency to humans and post hoc explanations as competing concepts. Throughout, the feasibility and desirability of different notions of interpretability are discussed. The article questions the oft-made assertions that linear models are interpretable and that deep neural networks are not. © 2018 Association for Computing Machinery. All rights reserved.
Full-text available
After 100 years of research, the definition of the field is still inadequate. The biggest challenge we see is moving away from a de-factor definition of intelligence in terms of test scores, but at the same time making clear what the boundaries of the field are. We then present four challenges for the field, two within a biological and two within a social context. These revolve around the issues of the malleability of intelligence and its display in everyday life, outside of a formal testing context. We conclude that developments in cognitive neuroscience and increases in the feasibility of monitoring behavior outside of the context of a testing session offer considerable hope for expansion of our both the biological and social aspects of individual differences in cognition.
Full-text available
The first issue of the Journal of Intelligence is devoted to a discussion based on the following two questions: 1. What are the most important scientific issues in the domain of human intelligence? 2. What are the most promising new ideas and approaches in the study of human intelligence? [...]
A group of industry, academic, and government experts convene in Philadelphia to explore the roots of algorithmic bias.
Conference Paper
Algorithm fairness has started to attract the attention of researchers in AI, Software Engineering and Law communities, with more than twenty different notions of fairness proposed in the last few years. Yet, there is no clear agreement on which definition to apply in each situation. Moreover, the detailed differences between multiple definitions are difficult to grasp. To address this issue, this paper collects the most prominent definitions of fairness for the algorithmic classification problem, explains the rationale behind these definitions, and demonstrates each of them on a single unifying case-study. Our analysis intuitively explains why the same case can be considered fair according to some definitions and unfair according to others.
THE SUNDAY TIMES BESTSELLER'I see [Raworth] as the John Maynard Keynes of the 21st Century: by reframing the economy, she allows us to change our view of who we are, where we stand, and what we want to be.' George Monbiot, Guardian'This is sharp, significant scholarship . . . Thrilling.' Times Higher Education'[A] really important economic and political thinker.' Andrew MarrEconomics is broken. It has failed to predict, let alone prevent, financial crises that have shaken the foundations of our societies. Its outdated theories have permitted a world in which extreme poverty persists while the wealth of the super-rich grows year on year. And its blind spots have led to policies that are degrading the living world on a scale that threatens all of our futures.Can it be fixed? In Doughnut Economics, Oxford academic Kate Raworth identifies seven critical ways in which mainstream economics has led us astray, and sets out a roadmap for bringing humanity into a sweet spot that meets the needs of all within the means of the planet. En route, she deconstructs the character of ‘rational economic man’ and explains what really makes us tick. She reveals how an obsession with equilibrium has left economists helpless when facing the boom and bust of the real-world economy. She highlights the dangers of ignoring the role of energy and nature’s resources – and the far-reaching implications for economic growth when we take them into account. And in the process, she creates a new, cutting-edge economic model that is fit for the 21st century – one in which a doughnut-shaped compass points the way to human progress.Ambitious, radical and rigorously argued, Doughnut Economics promises to reframe and redraw the future of economics for a new generation.'An innovative vision about how we could refocus away from growth to thriving.' Daily Mail'Doughnut Economics shows how to ensure dignity and prosperity for all people.' Huffington Post