Content uploaded by Lucas Memmert
Author content
All content in this area was uploaded by Lucas Memmert on Jan 03, 2022
Content may be subject to copyright.
Complex Problem Solving through Human-AI Collaboration:
Literature Review on Research Contexts
Lucas Memmert
University of Hamburg
lucas.memmert@uni-hamburg.de
Eva A. C. Bittner
University of Hamburg
eva.bittner@uni-hamburg.de
Abstract
Solving complex problems has been proclaimed as
one major challenge for hybrid teams of humans and ar-
tificial intelligence (AI) systems. Human-AI collabora-
tion brings immense opportunities in these complex
tasks, in which humans struggle, but full automation is
also impossible. Understanding and designing human-
AI collaboration for complex problem solving is a
wicked and multifaceted research problem itself. We
contribute to this emergent field by reviewing to what
extent existing research on instantiated human-AI col-
laboration already addresses this challenge. After clar-
ifying the two key concepts (complex problem solving
and human-AI collaboration), we perform a systematic
literature review. We extract research contexts and as-
sess them considering different complexity features. We
thereby provide an overview of existing and guidance
for designing new, suitable research contexts for study-
ing complex problem solving through human-AI collab-
oration and present an outlook for further work on this
research challenge.
1. Introduction
Solving complex problems has been proclaimed as
one major challenge for hybrid teams of humans and ar-
tificial intelligent (AI) systems [1]. The nature of work
changes towards higher knowledge intensity and com-
plexity, often exceeding the abilities of individual hu-
mans [2]. In parallel, AI, in particular machine learning,
has improved over the last years even outperforming hu-
mans in some areas. However, those AI systems are lim-
ited to specific, narrowly defined tasks, far from reach-
ing a general level of intelligence [3].
[4] describe a combination of human and AI –
named “hybrid intelligence” by different authors [4, 5]
– as a likely paradigm for the next years. The underlying
reasoning is that the strengths and weaknesses of hu-
mans and AI systems are complementary, resulting in a
mutually beneficial relationship when combined, ena-
bling the team to jointly achieve more than individually
possible [4]. According to [6] for AI-enabled machines
to be “effective teammates” they need to “engage in at
least some of the steps in a complex problem solving
process”. However, how humans and AI systems collab-
orate to solve problems is still unclear [7].
Though the challenge of complex problem solving
(CPS) through human-AI collaboration was posed com-
paratively recently [1], there is already an extensive
body of knowledge available around the broader topic
of human-AI collaboration. However, it is unclear, to
what extent this research deals with CPS.
Addressing this open point, we review extant liter-
ature on instantiated human-AI collaboration and
thereby answer the following research questions (RQ):
RQ 1: In which instantiated research contexts is re-
search on human-AI collaboration conducted?
RQ 2: To what extent can these research contexts be
considered to deal with complex problem solving?
[6] advocate for taking a socio-technical perspec-
tive on human-AI collaboration. Figure 1 illustrates a
schematic view of socio-technical systems, consisting
of user(s) interacting with technology to work on a task
within a specific context. To what extent a study can be
considered to deal with CPS depends on certain charac-
teristics of the task that the human-AI team performs
and the context this task is performed in. Therefore, the
combination of task and context will be the focus of this
study. It will be referred to as research contexts for the
remainder of this paper (some authors use scenario as a
term, e.g., [8, 9]). For the analysis, the research contexts
will be extracted from the relevant publications, de-
scribed briefly (RQ 1) and assessed against literature-
based characteristics for CPS (RQ 2).
For researchers wanting to study CPS through hu-
man-AI collaboration it is crucial to select an appropri-
ate research context, as the choice of the research con-
text will define the class of the problem the derived in-
sights can be generalized to. This paper provides re-
searchers with an overview of potential research con-
texts and relevant underlying characteristics.
Proceedings of the 55th Hawaii International Conference on System Sciences | 2022
Page 378
URI: https://hdl.handle.net/10125/79376
978-0-9981331-5-7
(CC BY-NC-ND 4.0)
Figure 1. Schematic depiction of a socio-tech-
nical system (according to [10]); focus added
The analysis is limited to studies going beyond only
providing theoretical descriptions of potential research
contexts, focusing on studies with instantiated research
contexts, i.e., actual implementations of research con-
texts within which user studies or experiments were
conducted. This focus ensures that only research con-
texts are included which have already proven to be suit-
able for studying CPS through human-AI collaboration.
They are therefore thought to be most helpful for schol-
ars as a starting point, both for conducting studies within
these already established research contexts and for in-
spiring the construction of new research contexts.
Additionally, this focus on instantiations was set in
the light of the design science research tradition in IS
research [11] and the call for design knowledge [1, 6],
with instantiations playing a critical role in theory build-
ing (e.g., for developing cause-effect-theories). As [11]
note, “[m]id-range theories or grand design theories” are
usually not simply invited, but are built through abstrac-
tion from a number of specific instantiations.
The remainder of this paper is structured as follows:
As part of the background section, “human-AI collabo-
ration” and its adjacent concepts as well as “CPS” are
explained. Within the methodology section the ap-
proach for reviewing the literature is described. The
overview regarding research contexts and their com-
plexities along with additional explanatory details is
provided in the succeeding section. Subsequently, limi-
tations are described and the paper closes with a conclu-
sion and an outlook on relevant future research.
2. Background
Both concepts relevant in the context of this paper
(human-AI collaboration, CPS) have a long research
history outside Information System (IS) research and
are described in detail in the following section.
2.1. (Human-AI) collaboration
Human-AI collaboration is the idea of humans and
AI systems working together. The underlying reasoning
is that humans and AI systems are complementary in
their strengths and weaknesses [1, 5]. AI systems can
perform clearly defined tasks, process large amounts of
data, recognize patterns and make coherent predictions
[1]. Already today AI systems can outperform humans
who can only process a limited amount of information
and are subject to cognitive bias [5]. Humans, on the
other hand, are creative, empathic and have common
sense [1]. They comparatively easily navigate dynamic
environments, put unexpected events into context and
react flexibly to changing circumstances [1, 5]. They
can deal with fragmented information and solve abstract
problems [10]. [1] provide an in-depth discussion and
conceptualization of “hybrid intelligence”.
While there are already many contributions around
this topic, there is some uncertainty regarding the termi-
nology. Several papers use the term collaboration with-
out a shared definition [e.g., 4, 5]. Additionally, the idea
of combining humans and AI systems is researched un-
der a variety of labels, such as human-AI collaboration,
human-AI teaming or hybrid intelligence. However, the
concepts are not clearly distinguished, which might hin-
der the progress of research. This terminological chal-
lenge is not only faced by IS research but exists cross-
disciplinarily [12]. [12] point out that “without shared
understanding of the collaboration construct, it is diffi-
cult to theorize and empirically test […] outcomes asso-
ciated with collaboration” and therefore state that “it is
necessary to reduce construct confusion about collabo-
ration to systematically and collectively advance
knowledge”.
[12] provide a comprehensive, multi-discipline lit-
erature review identifying the relevant aspects of collab-
oration, going beyond the idea of working together and
defining collaboration as an “evolving process whereby
two or more social entities actively and reciprocally en-
gage in joint activities aimed at achieving at least one
shared goal” [12]. This widely cited review is used as a
foundation for the understanding of “collaboration” in
this study. The relevant characteristics of the definition
are briefly explained according to [12]:
Evolving process. Collaboration is neither struc-
ture nor outcome, but a process leading to a (collabora-
tive) outcome. This process consists of emergent states
(e.g., motivations or values of team members) and col-
laborative behaviors (e.g., adaptation or sense making)
and is influenced by contextual factors (e.g., task or en-
vironment). These influences feed back into the collab-
orative process making it dynamic or evolving in nature.
Two or more social entities. Collaboration or
“working together” requires at least two entities such as
individuals, teams or organizations.
Reciprocal. Collaboration requires mutual engage-
ment of the involved entities, contributing to the collab-
orative effort, i.e., collaboration is neither one-sided nor
simply a delegation of work, but is a “back-and-forth”
reciprocal process. The requirement for a “back-and-
forth” reciprocal process is interpreted in this paper as
Page 379
follows: a simple way of working together with an AI
system, such as the AI making a recommendation and a
human taking the decision (i.e., decision support sys-
tem) does not qualify as collaboration. The reciprocity
characteristic here is in alignment with the understand-
ing of hybrid intelligence where “activities conducted
by each agent are conditionally dependent” [1].
Participation in joint activities. Collaboration fo-
cusses on performing joint activities such as solving
problems, performing actions, or executing tasks.
Achieving a shared goal. Entities participating in
shared work might have a variety of goals, which might
be conflicting. However, what sets collaboration apart,
is, that all entities have at least one shared or mutually
agreed upon goal.
According to the analysis of [12], teaming and col-
laboration differ to the extent, that collaboration might
occur on different levels (e.g., individuals, teams or or-
ganizations), whereas teaming usually applies only on
team-level (see Figure 2). Given this relationship, both
concepts will be considered in this paper as part of the
literature search. Coordination and cooperation lack
some of the before mentioned features and are therefore
not included (though considered during initial search).
Figure 2. “Shared criterion space among col-
laboration and related constructs” according
to [12]; focus added
A widely used means to analyze team effectiveness
is the Input-Mediator-Output-Model (I-M-O-Model)
[13]. According to [13], inputs can be considered “ante-
cedent factors that enable and constrain members’ inter-
actions”, mediators are processes and emergent states
impacting outputs, and outputs are “results and by-prod-
ucts of team activity that are valued by one or more con-
stituencies” [13]. While an in-depth analysis of team ef-
fectiveness is beyond the scope of this paper, we will
use the model to point our relevant aspects of the studies
with regards to CPS through human-AI collaboration.
2.2. Complex problem solving
Solving complex problems was proclaimed as the
next challenge [1] for the collaboration of humans and
AI systems. [6] ask for AI systems in order to be “effec-
tive teammates” to “engage in at least some of the steps
in a complex problem solving process [emphasis
added]” and [1] outline their understanding of complex
problems as being “time variant, dynamic, requir[ing]
much domain knowledge and hav[ing] no specific
ground truth”. They state that “these highly uncertain
contexts require intuitive and analytic abilities, as well
as human strengths such as creativity and empathy”.
These descriptions provide a first impression of
what could be studied under the label “CPS”. In order to
systematically review literature, however, a more de-
tailed understanding of CPS rooted in theory is required.
CPS has a long research history in the psychology
literature dating back to the mid-1970s [14]. It started
with researching relatively simple problems in labora-
tory environments [9]. However, researchers realized
that insights derived from those studies did not general-
ize well to more complex, real-life problems [9]. In re-
action, European and North American scholars pursued
different avenues, with European scholars focusing on
increasingly complex, computerized laboratory tasks
(so called “microworlds”) whereas Northern American
researchers focused on separate, natural knowledge do-
mains (e.g., reading, writing, calculation) [9]. As a re-
sult, key terms such as “complexity” or “CPS” itself are
not defined consistently in the field [8, 9].
Moreover, the transition from simple problem solv-
ing to complex problem solving is not clearly defined
either [14]. [9] state, that these two concepts are quali-
tatively different. While simple problem solving re-
quires to “overcome a single barrier; [complex problem
solving], in contrast, deals with a large number of barri-
ers that coexist simultaneously. Because there are mul-
tiple barriers, a single cognitive or behavioral activity
may not be sufficient to reach the goal state”. Problems
solvable by varying one thing at a time (VOTAT)
should not be considered complex problem solving [14].
Reflecting on this discourse, [8] provide a taxon-
omy of CPS tasks in order to enable an assessment of
“the similarity between tasks and to generalize results
from one type of task to another”. The taxonomy con-
tains both “formal” and “psychological” features. Using
psychological features for the analysis would require
“[postulating] processes and representations […] in par-
ticipants facing the task” [8]. Given the novelty of the
research challenge and the lack of understanding on
both processes and representations of joint human-AI
CPS in research, only “formal” features are used for the
analysis.
The taxonomy consists of three feature-groups
(time related, variable related, and system behavior re-
lated), each containing three features. Each feature is
shortly explained below according to [8]:
Time related aspects are:
Time variance. While static systems only change
when participants intervene, dynamic or time vari-
ant systems can change independently of the partic-
ipant, increasing complexity.
Page 380
Time continuity. Progression of time might be con-
tinuous, e.g., as in most real-life settings or might
occur stepwise, e.g., as in games were participants
take turns marking discrete “steps”.
Degree of time pressure. Time pressure refers to
whether the research contexts require rapid deci-
sions and actions, which increases complexity.
Variable related aspects are:
Number and type (discrete/continuous) of varia-
bles. Provides a first indication for the size of the
research context or problem, with more variables
increasing complexity.
Number and pattern of relationships between vari-
ables. Refers to the number of connections between
variables. The more interconnected variables are,
i.e., the more variables change as a consequence of
a manipulation of a variable, the more complex.
Certain patterns make the detection of causalities
more difficult and therefore increase complexity.
Non-linearity. Refers to the type of connection be-
tween variables. As humans do not perform well at
estimating non-linear relations, those are consid-
ered to increase complexity.
System behavior related aspects are:
Opacity. Refers to whether the complete state of the
problem is visible or transparent to the participant.
A system containing variables not directly observ-
able (opaque) is considered more complex.
Stochasticity. Refers to the changes within the sys-
tem resulting from an action. If the same action is
taken in the same state always results in the same
succeeding state, the system is deterministic. If the
resulting state might vary (stochastic), this in-
creases complexity. Some opaque research contexts
might appear to be stochastic to participants (be-
cause the participants do not see the “full picture”),
however, during this literature review the research
contexts are assessed with an outside view.
Delayed feedback. Refers to the possibility of relat-
ing feedback to actions. If feedback (e.g., success-
ful completion or increased score) cannot be related
directly to an action, this increases complexity.
In CPS literature, typically a human is assessed re-
garding their problem solving competency and strate-
gies. For this paper, we considered a team of at least one
human and an AI system. Considering multiple (includ-
ing non-human) stakeholders makes a clarification re-
garding opacity and stochasticity necessary.
Crucial for determining opacity for this study is the
combined perception of the participant team, i.e., if the
team members can jointly perceive the complete state of
the research context, it is not considered to be opaque.
Additionally, stochasticity here is considered a
problem characteristic. The fact that one of the team
members (e.g., a human that is part of the human-AI
team) might act stochastically, does not make the prob-
lem or research context stochastic. Simply speaking, a
problem should not be considered complex because a
human and AI work on it together but because the prob-
lem is inherently complex. Complexity introduced
through working with AI systems is out-of-scope for
this study as it was discussed already by [15].
This conceptualization of complexity is actor-inde-
pendent, which is in alignment with CPS literature: “In
the literature on CPS, it is mostly the structure of the
external problem representation that is considered com-
plex. So a problem usually is considered being of a cer-
tain complexity, even if it might seem less complex to
problem solvers with more expertise” [16].
It should be noted, that none of the features are nec-
essary or sufficient to qualify a research context as CPS
[8]. However, they can be used to compare research con-
texts and help to understand, which findings might gen-
eralize across which research contexts [8].
3. Methodology
In order to ensure transparency and reproducibility,
a systematic literature review following [17] is con-
ducted. The search string used is "hybrid intelligence"
OR ("human" AND ("AI" OR "artificial intelligence")
AND ("team*" OR "collaboration" OR “coordination”
OR “cooperation”)). The asterisk works as a wildcard
for any number of characters providing flexibility. For
databases without wildcard-support the terms team,
teams, and teaming where used.
In alignment with [1] hybrid intelligence is used to
find contributions describing combinations of humans
and AI systems. However, as this is a relatively new
term for describing human-AI collaboration, contribu-
tions containing human and artificial intelligence (or
abbreviated) either in a collaboration- or team-setting
are included additionally. These two concepts (collabo-
ration, team*) are chosen due to their similarity [12].
We presumed that some authors might use the concepts
interchangeably and after initial screening we included
the terms coordination and cooperation into our search.
The search string does not contain any reference to
CPS as problems or tasks considered to qualify as CPS
vary widely [8] and including specific aspects might
lead to premature exclusion of relevant search results.
The search string is applied to title, abstracts and
keywords. Where possible, the search is restricted to
peer-reviewed contributions. Relevant publications for
further analysis are identified in two review cycles.
1st review: By examining title, abstract and key-
words, the following type of contributions are excluded:
“hybrid intelligence” or “AI” have a different meaning
than intended in the search string (e.g., “hybrid intelli-
gence” refers to different types of AI algorithms being
Page 381
used together), purely conceptual/ non-instantiated
(e.g., theoretical frameworks, use case descriptions), fo-
cus on technical aspects of improving machine learning,
non-collaborative (in alignment with the understanding
outlined earlier), no usage of AI (e.g., suggestion of AI
usage only as an outlook), and certain formats (e.g.,
workshops, talks).
2nd review: By examining the (user) study and re-
search context descriptions within the contributions,
only those providing sufficient information to assess the
“CPS” characteristics outlined previously are included.
This requires a specific description of the problem
sought to be solved, not some (more generic) outlook as
to which research contexts could potentially benefit
from the research. It also requires an instantiation of the
research context, e.g., accompanied by a user study in
order to maintain the focus on suitable research contexts
and provide a strong foundation for theorizing. Publica-
tions with duplicate research contexts (e.g., same con-
text investigated in different publications) and primarily
physical tasks (i.e., non-intellectual [1]) were excluded.
The searched domain-relevant databases, the num-
ber of results during the initial search, as well as after
each of the review cycles are shown in Table 1. We per-
formed a forward and backward search to identify addi-
tional relevant articles.
Table 1. Search results per review stage and
database, and forward/backward (f/b) search
Database
Initial
hits
After 1st
review
After 2nd re-
view + f/b
AISeL
19
1
-
ACM DL
201
53
16 + 3
ScienceDirect
265
23
2 + 1
Total
485
77
18 + 4
4. Results
A total of 22 relevant publications were identified.
They were analyzed in order to extract the research con-
texts used to study human-AI collaboration (RQ 1).
Similar research contexts were grouped into research
context groups for further analysis. In order to under-
stand to what extent these contexts can be classified as
CPS, the CPS taxonomy of [12] was applied (RQ 2).
4.1. Research contexts of human-AI collab-
oration (RQ 1)
Each of the identified publications describes an in-
stantiation of human-AI collaboration in a specific re-
search context. Similar research contexts are grouped
for reasons of clarity. Table 2 provides an overview re-
garding the research context groups, the identified ex-
amples and the references. We also report exemplary
key inputs, mediators and outcomes (IMO) as well as
the key findings for the studies, in order to provide read-
ers with inspiration and to facilitate the analysis.
The following describes the research context
groups along with information regarding the reasoning
for using AI in this research context.
Creative content creation. Several studies deal
with the creation of creative content (e.g., composing
music or drawing a picture). Usually, a human performs
an initial creative action (e.g., draws first lines of an idea
onto a digital canvas) and can then evoke the AI system
to add to the composition. Authors root their argument
for choosing AI systems in their research context in lit-
erature citing several benefits from previous studies,
e.g., [18] cite the effective provision of inspiration, stim-
ulation of divergent thinking and support for innovation.
Turn-based cooperative games. There is a variety
of cooperative games (e.g., card game or puzzles). Usu-
ally it is two players taking actions in alternating order.
While the puzzles game [19] allows simultaneous action
by both players, it does not necessitate simultaneous ac-
tion-taking and is thus categorized as turn-based here.
The reasoning for choosing AI systems is less related to
their abilities within this context but more to the re-
search interest for mediators like social perception [20],
effective communication [21] or explanations [19].
Real-time cooperative games & simula-
tions. There are several real-time (video) games requir-
ing players to move around and perform actions swiftly.
As will be described in the next section, these two types
of games (i.e., turn-based, real-time) differ in their com-
plexity and are thus sorted into different research con-
text groups. Similarly to turn-based games, research in
the real-time research contexts did primarily focus on
collaboration aspects e.g., on how beliefs about team-
mates manifest in actions [22] or behavior [23]. Due to
the similarity, simulations are included in this group.
Design. Five studies are related to design teams
supported by AI systems, e.g., focusing on the creation
of mood boards (“visual collages composed of images,
text, and objects, that express concepts, ideas and emo-
tions” [24]) or on engineers designing bridges [25] or
drones [26]. While the humans are in the lead, the AI
system provides suggestion for further ideas or im-
provements. AI systems were chosen to support in those
research contexts due to their high task-specific perfor-
mance expected to be effective when combined with hu-
man creativity and agility [25] and their ability to sup-
port the entire creative process including inspiration and
reflection [24].
Negotiation. In one study, two sellers (two AI sys-
tems) competitively negotiate with a human buyer re-
garding a transaction. The human buyer and AI system
seller might be seen as collaborators with the shared
goal of reaching an agreement, making them better off.
Page 382
With each of them aiming to maximize their own utility,
they have conflicting goals. This, however, is common
within collaborative settings [12]. The research context
was chosen due to potential benefits for “retail, e-com-
merce, legal, business, and industrial sectors” [27].
Investigation. In one study, an AI system jointly
with a team of conservation scientists, law enforcement
and criminologists iteratively analyze data to identify in-
stances of illegal trade of plants. AI is used to improve
the effectiveness of search and analysis [28].
In summary, most of the human-AI collaboration
research contexts are either real-life creative content
creation or design tasks, or well-defined game environ-
ments (both turn-based and real-time). For the former,
the AI systems generative ability is supposed to provide
inspiration to humans supporting the creative process.
In many cases, performance with and without AI is in-
vestigated (input – team composition). For the latter, the
focus is usually not set on game-specific aspects but on
researching aspects of collaboration such as effective
communication, focusing more on mediating aspects.
Reflecting on the characteristics of collaboration
described earlier, an evolving process and reciprocity
are essential. All these research context require a certain
degree of sense-making, i.e., trying to understand the
situation as well as actions and intentions of the other(s)
and adjusting accordingly (e.g., match the intentioned
drawing style [29]). They are also inherently iterative
and interdependent, requiring a back-and-forth between
agents. Games and simulations allow researchers more
fine-grained control over the environment enabling
them to enforce those characteristics, e.g., by creating
hard dependencies through providing agents with
unique capabilities or restricting access to information,
making communication necessary to achieve the joint
goal [30]. This is more difficult for real-world tasks
(e.g., writing, drawing, composing).
Measuring team effectiveness is a way to assess and
compare designs options. While for all research contexts
some performance outcome metric is used (see table 2),
researchers should keep in mind, that for game environ-
ments measuring performance is usually comparatively
easy, as a scoring system is (in most cases) inherently
available. For real-world tasks an external (expert) judg-
ment is required, increasing study effort. Self-reported
metrics on (agent) performance (commonly used in the
included studies) should be used with caution as these
might not align with actual performance (e.g., [20]). For
virtual research environments an in-depth analysis of the
log data might be fruitful (e.g., [31]).
In the included studies outcomes were measured on
team, but not on role or organizational level, which
might be due to the novelty of the research field.
4.2. Complex problem solving (RQ 2)
In order to understand to what extent CPS is already
addressed in the identified studies on human-AI collab-
oration, an assessment against the CPS taxonomy [8] in-
troduced earlier is performed.
Each of the research contexts identified is evaluated
according to the nine characteristics of CPS. Results are
aggregated towards research context groups and feature
groups (time related, variable related, system behavior
related) for clarity. As a result, for each research context
group the complexity along the three feature groups is
displayed in Table 3. Complexity is assessed and
marked as follows:
Complex – marked with ‘X’: assigned to a feature
group, if the majority of research contexts fulfilled
all three features of complexity
Somewhat complex – marked with ‘(X)’: assigned,
if the majority of research contexts fulfilled at least
one feature of complexity
In the following, the reasoning for each assessment
is explained briefly. The assessments only apply to the
research contexts analyzed as part of this literature re-
view without a claim to generalize to other research con-
texts that could be considered under these labels. Fea-
tures not mentioned explicitly within the explanations
do not fulfill the complexity features for the respective
research context groups marked somewhat complex.
Creative content creation. Typically, humans
work together with AI systems on the “drawing board”
in these research contexts with no time related complex-
ity. The exception among the assessed creative content
creation research contexts is the live improvisation of
music [32]. While drawing, many interrelated variables
(e.g., position, shape, coloring, size) need to be consid-
ered, making it complex in this regard. With regards to
the system behavior, the research contexts are somewhat
complex due to the delayed feedback, however, without
being opaque or stochastic.
Turn-based cooperative games. The turn-based
games within the analysis follow discrete steps without
changes to the environment (which are not player in-
duced) and are thus not complex with regards to time.
They require users to reason considering a variety of in-
terrelated factors and can thus be considered somewhat
complex. Given the adjusted understanding of opacity
described earlier, the games analyzed here are not con-
sidered opaque as the team can perceive the entire game
state. However, due to the delayed feedback they can be
considered somewhat complex.
Real-time cooperative games & simula-
tions. With time passing and the environment changing
continuously, participants are constantly required to de-
cide, which action to perform. Waiting is also consid-
ered an (implicit) action. Thus, real-time cooperative
Page 383
Table 2. Research context groups with examples and exemplary study details
Research context
groups/Example(s)
Key Outcome(s)
Key Input(s)/ Mediator(s)
Key Finding(s)
Creative content crea-
tion
Sketching [18, 29], com-
posing [32, 33], writing
[34, 35]
PO: Novelty, integrity, interest-
ingness & balance [18, 32],
scariness [34]
PB: Social & collaboration dy-
namics, flow [29, 32, 33]
AP: Match writing style, novelty,
creativity [35]
I-TC: AI vs. WoZ [29]
I-TC: AI vs. no-AI [18, 33]
I-TC: AI vs. human-AI vs.
human [34]
M-TP: Communication: inner
state [32], suggestion style
[35]
Positive effect: Visualization of machine confidence on flow &
composition [32]
Positive effect: AI on novelty [18] and social dynamics [33]
Positive effect: Hybrid approach on scariness [34]
Writer prefer more fine-grained control over outputs and edita-
bility [35]
Inconclusive: Group comparison; focus on framework [29]
Turn-based cooperative
games
Card game (Hanabi),
puzzle, word guessing
[19–21]
PO: Score, win/loose, efficiency
[20, 21]
AP: Helpfulness, intelligence, so-
ciability, humanness, likability,
creativity, trustworthiness [19–
21]
I-TC: AI vs. presumed-human
[20]
M-TP: Communication: im-
plicature & explanations
[19, 21]
Positive effect: Implicature on score and perceived humanness
[21]
Positive effect: Explanation style on helpfulness, trustworthi-
ness and overall experience [19]
Negative attitude towards AI despite equal performance [20]
Real-time cooperative
games & simulations
Dearth, capture the gun-
ner, defend the pass,
don’t starve together, re-
connaissance missions
overcooked [22, 23, 31,
36–38]
PO: Score [36], efficiency [37],
mission success [38]
PB: Protecting or sacrificing team-
mate [22, 23]
AP: Player identity (humanness)
[31], perceived helpfulness, ef-
ficiency [37], trust, under-
standability [38]
I-TC: AI vs. human [31, 36]
I-TC: AI vs. presumed-human
[22, 23]
I-E: Score visible [23]
M-TP: Communication: dif-
ferent modalities [36]/ ex-
planation types [37, 38]/
skill, reliability [38]
Dependability: Visibility of score on treatment of AI teammate
[23]
Humans act differently in presumed social context [22]
Online players value co-presence [36]
Positive effect: Staying close to other player on correct identi-
fication (AI or human) [31]
Positive effect: explanation on non-experts’ performance
[37]/on performance if facilitating decision-making [38]
Design
Mood board-, bridge-,
drone-, accessory-, game
level- design [24–26, 39,
40]
PO: Strength mass ratio [25],
quality, novelty [40], range, ve-
locity, payload, cost [26]
PB: Design effort, search strategy,
mental workload [26], solution
exploration [24]
AP: Usefulness, inspirational [24],
fun, frustration, aid, creativity,
adaptation [39], friendliness,
engagement [40]
I-TC: AI vs. no-AI [25, 26]
I-E: Task complexity [26]
M-TP: task-related algorithms
[39], embodiment [40],
agency variants (different
features) [24]
Positive effect: design on low performing teams (negative on
high performing teams) [25]
Dependability Agency preference of designer and design step;
AI algorithm for outcome [24]
Positive effect: embodiment on satisfaction, friendliness, en-
gagement & helpfulness [40]
Dependability Algorithm preference on designer [39]
Positive effect: AI on efficiency of solution space exploration
[26]
Impact: Task complexity on AI systems effects [26]
Negotiation
Baking items [27]
PO: Profit
PB: Likeability/trust
M-TP: Agents variants –
negotiation tactics
Positive effect: Specfic negotiation tactics on outcomes and
likeability
Investigation
Illegal trade [28]
PO: Efficiency
I-TC: AI vs. no-AI
Presumed positive effect: Hybrid approach on effiecency
PO = performance outcome, PB = performance behavior, AP = agent performance,
I-TC = input – team composition, I-E: input – environment, M-TP = mediator – team process; WoZ = Wizard of Oz
games are considered complex with regards to time.
Similar to turn-based games, they also require partici-
pants to consider a variety of interconnected variables
(somewhat complex). Lastly, besides delayed feedback,
some research contexts are stochastic or opaque making
them somewhat complex regarding system behavior.
Design. Similar to the creative content generation
group, humans create their designs iteratively and the AI
provides feedback or suggestions. The research contexts
are set in a dedicated design program with no time re-
lated complexity. However, a plethora of different, non-
linearly related variables need to be considered (e.g., de-
termining if a constructed bridge can carry a certain
weight) to successfully achieve the goal, making it com-
plex in this regard. The system behavior can be de-
scribed as somewhat complex, usually deterministic and
transparent but with delayed feedback.
Negotiation. The negotiation is performed in dis-
crete steps (offer, counteroffer, etc.) and is therefore not
considered to be complex with relation to time. It re-
quires participants to keep track of several interrelated
variables, however, mostly linearly connected, thus con-
sidered somewhat complex in this regard. With another
AI based system negotiating and the underlying policy
of the agent being unknown to the human-AI team, the
system behavior can be considered complex.
Investigation. Though time passes continuously
and (bad) actors might continue to trade plants during
participants deliberating, the system evolves slowly
(i.e., low degree of time pressure) and is thus only some-
what time-complex. The team needs to keep track of a
large number of interconnected variables and is there-
fore considered complex in this regard. Lastly, with
other people acting within the system, the behavior is
opaque, stochastic and provides only delayed feedback
(e.g., after offline investigation), making it complex.
In summary, the research contexts are mostly sim-
ulations of problems, which are not complex with re-
gards to time as they are usually static, meaning
“changes occur only whether participant inter-
venes” [8]. The real-time cooperative games & simula-
tions are an exception. In this research context group,
while participants deliberate, the environment can
change (“move on”). Many research contexts are not
carefully constructed “microworlds” but model real
world problems with many, non-trivially interrelated
variables and are therefore at least somewhat complex
with regards to variables. Similarly, the system behavior
Page 384
is at least somewhat complex for all research contexts
due to the delayed feedback, with some research con-
texts even being stochastic or opaque.
Table 3. Research context groups and com-
plexity assessments
Research context group
T
V
SB
Creative content creation
X
(X)
Turn-based cooperative games
(X)
(X)
Real-time cooperative games
& simulations
X
(X)
(X)
Design
X
(X)
Negotiation
(X)
X
Investigation
(X)
X
X
T = time related, V = variable related, SB = system behavior re-
lated, X = complex; (X) = somewhat complex
While in AI research reporting on the environment
characteristics is common, only few authors explicitly
report on the characteristics in the included studies. An
exceptions is [26], reporting characteristics in detail and
pointing out the before mentioned advantage of game
and simulation environments: fine-grained control.
[26] study is particularly relevant in the context of
this paper, as [26] investigate complexity as one of their
input factors and find significant changes in the AI sys-
tems effect on performance: “AI assistance is most ben-
eficial when addressing moderately complex objectives
but exhibits a reduced advantage in addressing highly
complex objectives”. Only one other study investigates
environment changes, similarly reporting changes in
strategy [23]. This sensitivity of results to complexity is
well-documented in CPS literature [9]. We therefore en-
courage researchers on CPS through human-AI collab-
oration to investigate the robustness of their findings
with regards to complexity, e.g., by using systematic
variation of system properties [9]. This will enable re-
searchers to more precisely describe to which contexts
their findings generalize to. While for real-world task
practical value might be easier to demonstrate, for re-
search results in games and simulations the argument
must be made convincingly, as results on CPS in the la-
boratory might not necessarily transfer into practice [9].
While none of the research contexts are considered
complex according to all features they might still be de-
fined as complex as “[o]f all the features studied, none
is necessary and sufficient to define a task as ‘complex’”
[8]. Thus, there is already human-AI collaboration re-
search dealing with (aspects of) CPS that can be used as
a foundation to address this recently posed challenge.
5. Limitations
Several limitations need to be considered. The search
term was informed by the theoretical concepts surround-
ing the broad term “working together” and focused on
“collaboration” and “teaming” due to their similarity
and fit to what was described as a research challenge by
[1]. However, joint work of humans and AI is studied
under a variety of labels and even though we addition-
ally considered “cooperation” and “coordination”, po-
tentially not all relevant works are covered. Addition-
ally, this paper builds on a cross-disciplinary literature-
based definition of collaboration from a widely cited ar-
ticle in reviewing and excluding search results. Using a
different understanding of collaboration might lead to a
different set of papers and potentially different results.
In combination with the novelty of the research chal-
lenge only 22 articles were included. Therefore, we do
not claim conclusiveness. On the contrary, our work
should be complemented and extended, as the breadth
of research contexts in the scientific literature evolves.
Additionally, the conceptualization of “CPS” is,
even though having a long research history [14], still
fuzzy [8]. There are many definitions and potential cri-
teria inspired by different understandings of what
should be studied under the label of “CPS” [9]. Here, the
formal features of a literature-informed, widely cited
taxonomy of CPS tasks was used. In favor of an inte-
grated framework, frequently used criteria such as the
necessity to use “creativity as opposed to routine behav-
ior” [14] are neglected. Furthermore, some of those for-
mal features of the taxonomy, while widely used for de-
termining complexity (e.g., number of variables [8, 41])
lack clearly defined decision boundaries and are diffi-
cult to apply ex post, without insight into the research
context instantiation process.
6. Conclusion and outlook
The nature of work is evolving and is increasingly
characterized by complexity, requiring among other
things solving complex problems [2]. Therefore, [1] de-
fine CPS through human-AI teams as a major challenge.
The goal of this study was to understand to what extent
this challenge is already addressed by extant research.
We offer a literature-informed clarification of the
two key concepts: human-AI collaboration and CPS.
Based on that, we perform a structured literature review
to provide an overview regarding the instantiated re-
search contexts used to perform research on CPS
through human-AI collaboration. Most research con-
texts assessed are either creative content generation/ de-
sign settings or cooperative games. Applying the fea-
tures of CPS from literature, most assessed research
contexts are not complex with regards to time, but some-
what complex with regards to variables and system be-
havior. While none fulfill all complexity features, some
might still be considered complex. Thus, extant human-
AI collaboration literature can be leveraged as a founda-
tion for the new challenge of complex problem solving.
Page 385
Besides the identified research contexts, we provide
insight into the suitability of research context for study-
ing CPS through human-AI collaboration. Games and
simulations offer the advantage of flexible adjustments,
which is helpful for enforcing characteristics of collab-
oration (e.g., reciprocity) and exploring sensitivity to
complexity variations (e.g., opacity). Real-world tasks
(e.g., drawing) are more difficult to control and results
more difficult to measure. However, justification of
practical value might be easier for the latter as compared
to lab environments (see CPS literature [9]).
We contribute to research in multiple ways. By de-
scribing the two key concepts and explicating their de-
fining characteristics, we provide foundational termi-
nology and frame the research challenge more clearly.
For researchers interested in human-AI collabora-
tion through CPS (who are not committed to a specific
use case) we provide inspiration by pointing out suitable
research context examples. We provide guidance for re-
search context selection or construction by explaining
important aspects (e.g., flexibility, measurability) and
study design by highlighting the sensitivity to complex-
ity and recommending systematic variation of system
properties to achieve more robust results [9].
For researchers who have conducted a study we
provide a structure along which they can describe their
contribution more precisely. Assuming findings gener-
alize better to more similar situations, the taxonomy of-
fers a way to compare research contexts and more pre-
cisely describe the relevant class of problems. Given the
sensitivity of results to complexity (documented in CPS
literature; also shown e.g., by [26]), this is important.
Conceptually, going forward we encourage the
identification of additional suitable research contexts,
e.g., through the evaluation of existing testbeds from re-
inforcement learning research and proven contexts from
CPS literature regarding their suitability for CPS
through human-AI-collaboration research. We also ad-
vocate to consider additional dimensions of complexity
(e.g., technological or organizational) in future work.
This should be done with the type of problems in mind
that [1] mentioned when posing this research challenge:
e.g., managerial, political or science problems – “tasks
that so far seem to be at the core of human intellect” [1].
Acknowledgement. This research was funded by the
German Federal Ministry of Education and Research
(BMBF) in the context of the project HyMeKI (reference
number: 01IS20057).
7. References
[1] Dellermann, D., P. Ebel, M. Söllner, and J.M. Leimeister,
"Hybrid Intelligence", Business & Information Systems
Engineering, 61(5), 2019, pp. 637–643.
[2] Jacobs, R.L., "Knowledge Work and Human Resource De-
velopment", Human Resource Development Review,
16(2), 2017, pp. 176–202.
[3] Grace, K., J. Salvatier, A. Dafoe, B. Zhang, and O. Evans,
"Viewpoint: When Will AI Exceed Human Performance?
Evidence from AI Experts", Journal of Artificial Intelli-
gence Research, 62, 2018, pp. 729–754.
[4] Dellermann, D., A. Calma, N. Lipusch, T. Weber, S. Wei-
gel, and P. Ebel, "The Future of Human-AI Collabora-
tion: A Taxonomy of Design Knowledge for Hybrid In-
telligence Systems", in Proceedings of the 52nd Hawaii
International Conference on System Sciences, T. Bui, Ed-
itor. 2019.
[5] Akata, Z., D. Balliet, M. de Rijke, F. Dignum, V. Dignum,
G. Eiben, A. Fokkens, D. Grossi, K. Hindriks, H. Hoos,
H. Hung, C. Jonker, C. Monz, M. Neerincx, F. Oliehoek,
H. Prakken, S. Schlobach, L. van der Gaag, F. van Har-
melen, H. van Hoof, B. van Riemsdijk, A. van
Wynsberghe, R. Verbrugge, B. Verheij, P. Vossen, and
M. Welling, "A Research Agenda for Hybrid Intelli-
gence: Augmenting Human Intellect With Collaborative,
Adaptive, Responsible, and Explainable Artificial Intelli-
gence", Computer, 53(8), 2020, pp. 18–28.
[6] Seeber, I., E. Bittner, R.O. Briggs, T. de Vreede, G.-J. de
Vreede, A. Elkins, R. Maier, A.B. Merz, S. Oeste-Reiß,
N. Randrup, G. Schwabe, and M. Söllner, "Machines as
teammates: A research agenda on AI in team collabora-
tion", Information & Management, 57(2), 2020.
[7] Benbya, H., S. Pachidi, and S.L. Jarvenpaa, "Special Issue
Editorial: Artificial Intelligence in Organizations: Impli-
cations for Information Systems Research", Journal of
the Association for Information Systems, 22(2), 2021.
[8] Quesada, J., W. Kintsch, and E. Gomez, "Complex prob-
lem-solving: a field in search of a definition?", Theoreti-
cal Issues in Ergonomics Science, 6(1), 2005, pp. 5–33.
[9] Frensch, P.A. and J. Funke, Complex Problem Solving:
The European Perspective, Taylor and Francis, Hoboken,
1995.
[10] Maedche, A., C. Legner, A. Benlian, B. Berger, H.
Gimpel, T. Hess, O. Hinz, S. Morana, and M. Söllner,
"AI-Based Digital Assistants", Business & Information
Systems Engineering, 61(4), 2019, pp. 535–544.
[11] Gregor, S. and A.R. Hevner, "Positioning and Presenting
Design Science Research for Maximum Impact", MIS
Quarterly, 37(2), 2013, pp. 337–355.
[12] Bedwell, W.L., J.L. Wildman, D. DiazGranados, M. Sal-
azar, W.S. Kramer, and E. Salas, "Collaboration at work:
An integrative multilevel conceptualization", Human Re-
source Management Review, 22(2), 2012, pp. 128–145.
[13] Mathieu, J., M.T. Maynard, T. Rapp, and L. Gilson,
"Team Effectiveness 1997-2007: A Review of Recent
Advancements and a Glimpse Into the Future", Journal of
Management, 34(3), 2008, pp. 410–476.
[14] Dörner, D. and J. Funke, "Complex Problem Solving:
What It Is and What It Is Not", Frontiers in psychology,
8, 2017, p. 1153.
[15] Yang, Q., A. Steinfeld, C. Rosé, and J. Zimmerman, "Re-
examining Whether, Why, and How Human-AI Interac-
tion Is Uniquely Difficult to Design", in Proceedings of
the 2020 CHI Conference on Human Factors in Compu-
ting Systems. 2020. ACM: NY, USA.
Page 386
[16] Fischer, A., S. Greiff, and J. Funke, "The Process of Solv-
ing Complex Problems", The Journal of Problem Solv-
ing, 4(1), 2012.
[17] vom Brocke, J., A. Simons, B. Niehaves, K. Riemer, and
A. Cleven, "Reconstructing the Giant: On the Importance
of Rigour in Documenting the Literature Search Pro-
cess", 17th European Conference on Information Sys-
tems (ECIS). 2009.
[18] Zhang, C., C. Yao, J. Liu, Z. Zhou, W. Zhang, L. Liu, F.
Ying, Y. Zhao, and G. Wang, "StoryDrawer: A Co-Crea-
tive Agent Supporting Children’s Storytelling through
Collaborative Drawing", in Extended Abstracts of the
2021 CHI Conference on Human Factors in Computing
Systems. 2021. ACM: NY, USA.
[19] Tabrez, A., S. Agrawal, and B. Hayes, "Explanation-
Based Reward Coaching to Improve Human Performance
via Reinforcement Learning", in HRI ’19: Proceedings of
the 14th ACM/IEEE International Conference on Hu-
man-Robot Interaction. 2019. IEEE Press.
[20] Ashktorab, Z., Q.V. Liao, C. Dugan, J. Johnson, Q. Pan,
W. Zhang, S. Kumaravel, and M. Campbell, "Human-AI
Collaboration in a Cooperative Game Setting: Measuring
Social Perception and Outcomes", Proc. ACM Hum.-
Comput. Interact., 4(CSCW2), 2020.
[21] Liang, C., J. Proft, E. Andersen, and R.A. Knepper, "Im-
plicit Communication of Actionable Information in Hu-
man-AI Teams", in Proceedings of the 2019 CHI Confer-
ence on Human Factors in Computing Systems. 2019.
ACM: NY, USA.
[22] Merritt, T. and K. McGee, "Protecting Artificial Team-
Mates: More Seems like Less", in Proceedings of the
SIGCHI Conference on Human Factors in Computing
Systems. 2012. ACM: NY, USA.
[23] Ong, C., K. McGee, and T.L. Chuah, "Closing the Hu-
man-AI Team-Mate Gap: How Changes to Displayed In-
formation Impact Player Behavior towards Computer
Teammates", in Proceedings of the 24th Australian Com-
puter-Human Interaction Conference. 2012. ACM: NY,
USA.
[24] Koch, J., N. Taffin, M. Beaudouin-Lafon, M. Laine, A.
Lucero, and W.E. Mackay, "ImageSense: An Intelligent
Collaborative Ideation Tool to Support Diverse Human-
Computer Partnerships", Proc. ACM Hum.-Comput. In-
teract., 4(CSCW1), 2020.
[25] Zhang, G., A. Raina, J. Cagan, and C. McComb, "A cau-
tionary tale about the impact of AI on human design
teams", Design Studies, 72, 2021, p. 100990.
[26] Song, B., N.F. Soria Zurita, H. Nolte, H. Singh, J. Cagan,
and C. McComb, "When faced with increasing complex-
ity: The effectiveness of AI assistance for drone design",
Journal of Mechanical Design, 2021, pp. 1–38.
[27] Geraghty, R., J. Hale, S. Sen, and T.S. Kroecker, "FUN-
Agent: A 2020 HUMAINE Competition Entrant", in Pro-
ceedings of the 1st International Workshop on Multi-
modal Conversational AI. 2020. ACM: NY, USA.
[28] D. Whitehead, C.R. Cowell, A. Lavorgna, and S.E. Mid-
dleton, "Countering plant crime online: Cross-discipli-
nary collaboration in the FloraGuard study", Forensic
Science International: Animals and Environments, 1,
2021, p. 100007.
[29] Davis, N., C. Hsiao, K.Y. Singh, B. Lin, and B. Magerko,
"Quantifying Collaboration with a Co-Creative Drawing
Agent", ACM Trans. Interact. Intell. Syst., 7(4), 2017.
[30] van den Bosch, K., T. Schoonderwoerd, R. Blankendaal,
and M. Neerincx, "Six Challenges for Human-AI Co-
learning", in Adaptive Instructional Systems, R.A. Sot-
tilare and J. Schwarz, Editors. 2019. Springer Interna-
tional Publishing: Cham.
[31] Tulk, S., R. Cumings, T. Zafar, and E. Wiese, "Better
Know Who You Are Starving with: Judging Humanness
in a Multiplayer Videogame", in Proceedings of the
Technology, Mind, and Society. 2018. ACM: NY, USA.
[32] McCormack, J., T. Gifford, P. Hutchings, M.T. Llano Ro-
driguez, M. Yee-King, and M. d’Inverno, "In a Silent
Way: Communication Between AI and Improvising Mu-
sicians Beyond Sound", in Proceedings of the 2019 CHI
Conference on Human Factors in Computing Systems.
2019. ACM: NY, USA.
[33] Suh, M., E. Youngblom, M. Terry, and C.J. Cai, "AI as
Social Glue: Uncovering the Roles of Deep Generative
AI during Social Music Composition", in Proceedings of
the 2021 CHI Conference on Human Factors in Compu-
ting Systems. 2021. ACM: NY, USA.
[34] Yanardag, P., M. Cebrian, and I. Rahwan, "Shelley: A
Crowd-Sourced Collaborative Horror Writer", in Crea-
tivity and Cognition. 2021. ACM: NY, USA.
[35] Calderwood, A., V. Qiu, K.I. Gero, and L.B. Chilton,
"How Novellists use Generative Language Models: An
Exploratory User Study", IUI ’20 Workshops, 2020.
[36] McGee, K., T. Merritt, and C. Ong, "What We Have Here
is a Failure of Companionship: Communication in Goal-
Oriented Team-Mate Games", in Proceedings of the 23rd
Australian Computer-Human Interaction Conference.
2011. ACM: NY, USA.
[37] Gao, X., R. Gong, Y. Zhao, S. Wang, T. Shu, and S.-C.
Zhu, "Joint Mind Modeling for Explanation Generation
in Complex Human-Robot Collaborative Tasks", in 29th
IEEE International Conference on Robot and Human In-
teractive Communication. 2020: NY, USA.
[38] Wang, N., D.V. Pynadath, and S.G. Hill, "The Impact of
POMDP-Generated Explanations on Trust and Perfor-
mance in Human-Robot Teams", in Proceedings of the
2016 International Conference on Autonomous Agents &
Multiagent Systems. 2016. ACM: New York, USA.
[39] Guzdial, M., N. Liao, J. Chen, S.-Y. Chen, S. Shah, V.
Shah, J. Reno, G. Smith, and M.O. Riedl, "Friend, Col-
laborator, Student, Manager: How Design of an AI-
Driven Game Level Editor Affects Creators", in Proceed-
ings of the 2019 CHI Conference on Human Factors in
Computing Systems. 2019. ACM: NY, USA.
[40] Lin, Y., J. Guo, Y. Chen, C. Yao, and F. Ying, "It Is Your
Turn: Collaborative Ideation With a Co-Creative Robot
through Sketch", in Proceedings of the 2020 CHI Confer-
ence on Human Factors in Computing Systems. 2020.
ACM: NY, USA.
[41] Funke, J., "Complex Problem Solving", in Encyclopedia
of the Sciences of Learning, N.M. Seel, Editor. 2012.
Springer US: Boston, MA.
Page 387
