Content uploaded by Jeba Rezwana
Author content
All content in this area was uploaded by Jeba Rezwana on Apr 26, 2022
Content may be subject to copyright.
Designing Creative AI Partners with COFI: A Framework for Modeling
Interaction in Human-AI Co-Creative Systems
JEBA REZWANA and MARY LOU MAHER, University of North Carolina at Charlotte, USA
Human-AI co-creativity involves both humans and AI collaborating on a shared creative product as partners. In a creative collaboration,
interaction dynamics, such as turn-taking, contribution type, and communication, are the driving forces of the co-creative process.
Therefore the interaction model is a critical and essential component for eective co-creative systems. There is relatively little research
about interaction design in the co-creativity eld, which is reected in a lack of focus on interaction design in many existing co-creative
systems. The primary focus of co-creativity research has been on the abilities of the AI. This paper focuses on the importance of
interaction design in co-creative systems with the development of the Co-Creative Framework for Interaction design (COFI) that
describes the broad scope of possibilities for interaction design in co-creative systems. Researchers can use COFI for modeling
interaction in co-creative systems by exploring alternatives in this design space of interaction. COFI can also be benecial while
investigating and interpreting the interaction design of existing co-creative systems. We coded a dataset of existing 92 co-creative
systems using COFI and analyzed the data to show how COFI provides a basis to categorize the interaction models of existing
co-creative systems. We identify opportunities to shift the focus of interaction models in co-creativity to enable more communication
between the user and AI leading to human-AI partnerships.
CCS Concepts: •Human-centered computing →Interaction design process and methods;Collaborative interaction.
Additional Key Words and Phrases: Human-AI Co-Creativity, Co-Creativity, Interaction Design, Framework
ACM Reference Format:
Jeba Rezwana and Mary Lou Maher. 2022. Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in
Human-AI Co-Creative Systems. ACM Trans. Comput.-Hum. Interact. 1, 1, Article 1 (January 2022), 27 pages.
1 INTRODUCTION
Computational creativity is an interdisciplinary eld that applies articial intelligence to develop computational systems
capable of producing creative artifacts, ideas and performances [
33
]. Research in computational creativity has lead
to dierent types of creative systems that can be categorized based on their purposes: systems that generate novel
and valuable creative products, systems that support human creativity, and systems that collaborate with the user on
a shared creative product combining the creative ability of both the user and the AI [
39
]. Davis introduces the term
human-computer co-creativity, where humans and computers can collaborate in a creative process as colleagues [42].
In human-computer co-creativity, both humans and AI agents are viewed as one system through which creativity
emerges. The creativity that emerges from a collaboration is dierent from creativity emerging from an individual
as creative collaboration involves interaction among collaborators and the shared creative product is more creative
than each individual could achieve alone [
137
]. Stephen Sonnenburg demonstrated that communication is the driving
force of collaborative creativity [
145
]. Interaction is a basic and essential component of co-creative systems as both
Authors’ address: Jeba Rezwana, jrezwana@uncc.edu; Mary Lou Maher, m.maher@uncc.edu, University of North Carolina at Charlotte, USA, .
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not
made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to
redistribute to lists, requires prior specic permission and/or a fee. Request permissions from permissions@acm.org.
©2022 Association for Computing Machinery.
1
2 Rezwana and Maher
the human and the AI actively participate and interact in the co-creation, unlike autonomous creative systems that
generate creative artifacts alone and creativity support tools that support human creativity.
Designing and evaluating co-creative systems has many challenges due to the open-ended and improvisational
nature of the interaction between the human and the AI agent [
41
,
78
]. Humans utilize many dierent creative strategies
and reasoning processes throughout the creative process, and ideas and the creative product develop dynamically
through time. This continual progression of ideas requires adaptability on the agent’s part. Additionally, it is not always
clear how the co-creative AI should contribute and interact during the course of the co-creative process. For example,
sometimes the human may want to lead and have the AI assist with some tasks, whereas other times the human may
want the AI to lead to help nd inspiration or to work independently. Understanding the mechanics of co-creation is
still very much open questions in the young eld of human-computer co-creativity. Bown asserted that the success of a
creative system’s collaborative role should be further investigated as interaction plays a key role in the creative process
of co-creative systems [
17
]. AI ability alone does not ensure a positive collaborative experience of users with the AI
[
100
] and interaction is more critical than algorithms where interaction with the users is essential [
152
]. In this paper
we focus on the interaction design space as an essential aspect of eective co-creative systems.
Interaction design is the creation of a dialogue between users and the system [
87
]. Recently, interaction design in
co-creative systems is being addressed as a signicant aspect of computational creativity. Kantosalo et al. said that
interaction design, specically, interaction modality should be the ground zero for designing co-creative systems [
77
].
However, the interaction designs of many existing co-creative systems provide only one-way interaction, where humans
can interact with the AI but the system is not designed for the AI to communicate back to humans. For example,
Collabdraw [
54
] is a co-creative sketching environment where users draw with an AI. The user interface includes only
one button that users click to submit their artwork and indicate that their turn is complete. Other than the button, there
is no way for the user to communicate with the AI and vice versa to provide information, suggestions, or feedback.
Although the AI algorithm is capable of providing intriguing contributions to the creative process, the interaction
design is inadequate for collaboration between human and AI. Another example, Image to Image [
68
] is a co-creative
system that converts a line drawing of a particular object from the user into a photo-realistic image. The user interface
has only one button that users click to tell the AI to convert the drawing. Interaction design can provide more than
the transfer of instructions from a user to an AI agent to generate a creative artifact and can lead to a more engaging
user experience. A recent study showed increased user satisfaction with text-based instructions from the AI rather
than button-based instructions in a co-creation [
118
]. A starting point to investigate interaction models is the study
of collaboration among humans [
39
]. Understanding the factors in human collaboration can build the foundation for
the development of human-AI collaboration in co-creative systems [
107
]. Interaction models developed for computer
supported collaborative work is an important source for identifying interaction models related to co-creative systems.
In this paper, we present Co-Creative Framework for Interaction design (COFI) that describes interaction components
as a space of possibilities for interaction design in co-creative systems. These interaction components represent various
aspects of a co-creation, such as participation style, contribution type, and communication between humans and the AI.
COFI is informed by the literature on human collaboration, CSCW, computational creativity, and human-computer
co-creativity. We adopted interaction components based on a literature review and adapted the components to concepts
relevant to co-creativity. COFI can be used as a guide when designing the interaction models in co-creative systems.
COFI can also be benecial for investigating and interpreting the interaction design of existing co-creative systems.
We coded and analyzed the interaction models of a dataset of 92 co-creative systems using COFI to evaluate the value
and analytical power of the framework. Three distinct interaction models for co-creative systems emerged from this
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
3
analysis: generative pleasing AI agents that follow along with the user, improvisational AI agents that work alongside
users on a shared product spontaneously, and advisory AI agents that both generate and evaluate the creative product.
The analysis reveals that the co-creative systems in this dataset lack communication channels between the user and AI
agent. Finally, this paper discusses the limitations in the existing interaction models in co-creative systems, potential
areas for further development, and the importance of extending the scope of human-AI communication in co-creative
systems.
2 RELATED WORK
2.1 Co-creative Systems
Creativity is dened as the exploration and production of novel and useful ideas [
47
,
70
,
72
]. Wiggins dened creative
systems as "A collection of processes, natural or automatic, which are capable of achieving or simulating behavior which
in humans would be deemed creative [
156
]." Davis et al. discussed the three main categories of creative systems based
on their working processes and purposes [
39
]: standalone generative systems, creativity support tools, and co-creative
systems. Standalone generative systems refer to fully autonomous intelligent systems that work independently without
any interaction with humans in the creative process. Creative systems that support the user’s creativity without
contributing to the creative process are considered creativity support tools (CST). In co-creative systems, humans and
computer both contribute as creative colleagues in the creative process[
42
]. Co-creative systems originated from the
concept of combining standalone generative systems with creativity support tools as computers and humans both take
the initiative in the creative process and interact as co-creators [
76
]. Mixed initiative creative systems are often used as
a substitute term for co-creative systems in the literature [161].
In a co-creative system, interaction between the human and AI agent make the creative process complex and emergent.
Maher explores issues related to who is being creative when humans and AI collaborate in a co-creative system [
106
].
Antonios Liapis et al. argued that when creativity emerges from human-computer interaction, it cannot be credited
either to the human or to the computer alone, and surpasses both contributors’ original intentions as novel ideas arise
in the process [
95
]. Designing interaction in co-creative systems has unique challenges due to the spontaneity of the
interaction between the human and the AI [
41
]. A co-creative AI agent needs continual adjustment and adaptation to
cope with human strategies. A good starting point to investigate questions about modeling an eective interaction
design for co-creative systems can be studying creative collaboration in humans [
39
]. Mamykina et al. argued that by
understanding the factors of human collaborative creativity, methods can be devised to build the foundation for the
development of computer-based systems that can augment or enhance collaborative creativity [107].
2.2 Interaction Design in Co-creative Systems
Regarding interaction design in interactive artifacts, Fallman stated: "interaction design takes a holistic view of the
relationship between designed artifacts, those that are exposed to these artifacts, and the socio-cultural context in
which the meeting takes place" [
53
]. In the eld of co-creativity, interaction design includes various parts and pieces of
the interaction dynamics between the human and the AI, for example - participation style, communication between
collaborators, and contribution type. Now the question is how researchers and designers can explore the possible
spaces of interaction design in co-creative systems. For instance, turn-taking is the ability for agents to lead or follow in
the process of interaction [
158
]. While designing a co-creative system, should the designer consider turn-taking or
concurrent participation style? Turn-taking models work well in many co-creative systems but may not t well for all
co-creative systems. Lauren and Magerko investigated whether the user experience is improved with a turn-taking
model applied to Lumin AI, a co-creative dance partner, through an empirical study [
158
]. However, their results showed
4 Rezwana and Maher
negative user experience with a turn-taking model compared to a non-turn taking model. The negative user experience
resulted from the dislike for the AI agent to take the lead.
Bown argued that the most practiced form of evaluating articial creative systems is mostly theoretical and is not
empirically well-grounded and suggested interaction design as a way to ground empirical evaluations of computational
creativity [
16
]. Yee-King and d’Inverno also argued for a stronger focus on the user experiences of creative systems,
suggesting a need for further integration of interaction design practice into co-creativity research [
162
]. There is a lack
of a holistic framework for interaction design in co-creative systems. A framework for interaction design is necessary
to explain and explore the possible interaction spaces and compare and evaluate the interaction design of existing
co-creative systems for improving the practice of interaction modeling in co-creative systems.
There are recent developments in frameworks and strategies for interaction in co-creative systems. Kantosalo et al.
proposed a framework to describe three aspects of interaction, interaction modalities, interaction style and interaction
strategies, in co-creative systems [
77
]. They analyzed nine co-creative systems with their framework to compare
dierent systems’ creativity approaches even if they are within the same creative domain [
77
]. Bown and Brown
identied three interaction strategies - operation-based interaction, request-based interaction and ambient interaction
in metacreation, the automation of creative tasks with machines [
18
]. Bown et al. explored the role of dialogue between
the human and the user in co-creation and argued that both linguistic and non-linguistic dialogues of concepts and
artifacts are essential to maintain the quality of co-creation [
19
]. Guzdial and Riedl proposed an interaction framework
for turn-based co-creative AI agents to better understand the space of possible designs of co-creative systems [
62
]. Their
framework is limited to turn-based co-creative agents and has a focus on contributions and turn-taking. In this paper
we present COFI, a description of a design space of possibilities for interaction in co-creative systems that includes and
extends these existing frameworks and strategies.
2.3 Creative Collaboration among Humans
Sawyer asserted that the creativity that emerges from collaboration is dierent from the creativity emerging from an
individual where interaction among the group is a vital component of creativity [
137
]. He investigated the process
of creativity when emerging from a group by observing and analyzing improvisational theater performances by a
theater group [
137
] and argued that the shared product of collaborative creativity is more creative than each individual
alone could achieve. Sonnenburg introduced a theoretical model for creative collaboration, and this model presents
communication among the group as the driving force of collaborative creativity [
145
]. Interaction among the individuals
in a collaboration makes the process emergent and complex. For investigating human collaboration, many researchers
stressed the importance of understanding the process of interaction. Fantasia et al. proposed an embodied approach of
collaboration which considers collaboration as a property and intrinsic part of interaction processes [
55
]. They claimed
that interaction dynamics help in understanding and fostering our knowledge of dierent ways of engaging with others
and argued that it is crucial to investigate the interaction context, the environment, and how collaborators make sense
of the whole process for gaining more knowledge and understanding more about collaboration. In COFI, we include
components that address the interaction we observe in human to human collaboration as possibilities for human-AI
co-creativity.
Computer supported cooperative work (CSCW) is computer assisted coordinated activity carried out by a group of
collaborating individuals [
9
]. K Schmidt dened CSCW as an endeavor to understand the nature and characteristics of
collaborative work to design adequate computer-based technologies [
139
]. A foundation of CSCW is sense-making and
understanding the nature of collaborative work for designing adequate computer based technology to support human
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
5
collaboration. CSCW systems are designed to improve group communication while alleviating negative interactions
that reduce collaboration quality [
75
]. For building eective CSCW systems for collaborative creative work, many
CSCW researchers investigated creative collaboration among humans to understand the mechanics of collaboration.
For this reason, the design space of CSCW is relevant for interaction design in co-creative systems.
2.4 Sense-making in Collaboration
Sense-making is motivated by a continual urge to understand connections among people, places, and events to anticipate
their trajectories [
86
]. Russel et al. discussed the processes involved in a sense-making: searching for representations,
instantiate representations by encoding the representations, and utilizing the encoding in task-specic information
[
133
]. Davis argued that participatory sense-making is useful to analyze, understand and model creative collaboration
[
39
]. De Jaegher and Di Paolo also proposed participatory sense-making as a starting point of understanding social
interaction [
44
]. To understand participatory sense-making, the denition of sense-making from cognitive theory
is crucial. Sense-making is the way cognitive agents meaningfully connect with their world, based on their needs
and goals as self-organizing, self-maintaining, embodied agents [
43
]. Introducing multiple agents in the environment
makes the dynamics of sense-making more complex and emergent as each agent is interacting with the environment
as well as with each other. Participatory sense-making evolves from this complex, mutually interactive process [
39
].
Participatory sense-making occurs where "A co-regulated coupling exists between at least two autonomous agents
where the regulation itself is aimed at the aspects of the coupling itself so that the domain of relational dynamics
constitutes an emergent autonomous organization without destroying the autonomy of the agents involved." [
44
]. In
this quote, De Jaegher and Di Paolo outlines the process of participatory sense-making where meaning-making of
relational interaction dynamics such as the rhythm of turn-taking, manner of action, interaction style, etc is necessary
[41].
Shared
Product
Interaction with
the Shared
Product
Interaction
between
Collaborators
Fig. 1. Interactional Sense-making in a Co-creation.
To understand interaction dynamics in an open-ended improvisational collaboration, Kellas and Trees present a model
of interactional sense-making [
81
]. They describe two types of interaction in the sense-making process: interaction
between collaborators and interaction with the shared product (Figure 1). We adapt and extend this model for COFI to
ground our space of possibilities for interaction design on the concept of interactional sense-making. Interaction with the
shared product, in the context of a co-creative system, describes the ways in which the co-creator can sense, contribute,
and edit content of the emerging creative product. Interaction between collaborators explains how the interaction
between the co-creators is unfolding through time which includes turn-taking, timing of initiative, communication
etc. Participatory sense-making occurs when there is a mutual co-regulation of these two interactional sense-making
6 Rezwana and Maher
processes between the co-creators []. For example, when both participants are adapting their responses based on each
other’s contribution while maintaining an engaging interaction dynamic, participatory sense-making occurs.
3 CO-CREATIVE FRAMEWORK FOR INTERACTION DESIGN (COFI)
We develop and present Co-Creative Framework for Interaction Design (COFI) as a space of possibilities for interaction
design in co-creative systems. COFI also provides a framework for analyzing the interaction design trends of existing
co-creative systems. This framework describes various aspects involved in the interaction between the human and
the AI. COFI is informed by research on human collaboration, CSCW, computational creativity, and human-computer
co-creativity.
The primary categories of COFI are based on two types of interactional sense-making of collaboration as described by
Kellas and Trees (Figure 2) [
81
]: interaction between collaborators and interaction with the shared product. Interaction
with the shared product, in the context of co-creative systems, describes interaction aspects related to the creation of
the creative content. Interaction between collaborators explains how the interaction between the human and the AI is
unfolding through time which includes turn-taking, timing of initiative, communication, etc. Thus, COFI characterizes
relational interaction dynamics between the collaborators (human and AI) as well as functional aspects of interacting
with the shared creative product. Kellas and Trees’ framework was used for explaining and evaluating the interaction
dynamics in human creative collaboration in joint storytelling. Understanding collaborative creativity among humans
can be the basis for designing eective co-creative systems where the AI agent acts as a creative partner.
Each of the two categories of interaction is further divided into two subcategories. Interaction between collaborators
is divided into collaboration style and communication style. On the other hand, interaction with the shared product
is divided into the creative process and creative product. CSCW literature discusses collaboration mechanics among
the collaborators to make eective CSCW systems. Many frameworks about groupware and CSCW systems discuss
and emphasize both collaboration components and communication components among collaborators. For example,
Baker et al. proposed an evaluation technique based on collaboration mechanics for groupware and emphasized both
coordination and communication components in a collaboration [
11
]. Creativity literature focuses more on creativity
emergence, which includes creative processes and the creative product. For example, Rhodes’s famous 4P, which is
one of the most acknowledged model, includes creative process and product [
130
]. Therefore, in COFI, the literature
regarding human collaboration and CSCW literature informs the category ’interaction between the collaborators’,
while the creativity and co-creativity literature provides descriptions of the ’interaction with the shared product’. In
human-AI co-creativity, the focus should be on both creativity and collaboration. As a result, both the CSCW and
creativity literature provide the basis for dening the interaction components of COFI under the four subcategories.
We performed a literature review to identify the components of COFI. We identied a list of search databases for
relevant academic publications: ACM Library, arXiv, Elsevier, Springer, and ScienceDirect, and google scholar. We used
keywords based on the 4 Cs in COFI: Collaboration style, Communication style, Creative process, Creative product. The
total list of keywords are: ’human collaboration mechanics,’ ’creative collaboration among humans,’ ’communication in
collaboration,’ ’cooperation mechanics,’ ’Interaction in joint action,’ ’groupware communication,’ ’interaction design in
computational creativity,’ ’interaction in co-creativity,’ ’creative process,’ ’group interaction in computational creativity,’
’interaction in human-computer co-creation’. We considered documents published from 1990 until 2021. We did not
include papers that are a tutorial or poster, papers that are not in English, papers that by title or abstract are outside the
scope of the research, and papers that do not describe the collaboration mechanics or group interaction. We included
papers describing strategies, mechanisms and components of interaction in a natural collaboration, computer-mediated
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
7
Same Task
Collaboration
Style
Participation
Style
Timing of
Initiative
Task
Distribution
Human to AI
Intentional
Communication
Turn-taking
Parallel
Planned
Spontaneous
Task
Divided
Generate
Define
Embodied
Gaze
Facial
Expression
Embodied
Direct
Manipulation
Evaluate
Human to AI
Consequential
Communication
AI to Human
Communication
Speech
Embodied
Text
Haptic
Visual
Voice Text
Contribution
Similarity
Contribution
Type
Create
New
Extend Transform
Refine
High Low
(a) (b)
Interaction in
Co-creative Systems
Interaction
between
Collaborators
(human & AI)
Interaction
with the
Shared Product
Creative
Process
Creative
Product
Mimicry
Mimic Non-
mimic
Communication
Style
Biometric
Fig. 2. Co-Creative Framework for Interaction Design (COFI): On the le (a) Components of Interaction between the collaborators,
On the right (b) Components of Interaction with the Shared Product.
collaboration and human-AI collaboration. COFI was developed in an iterative process of adding, merging, and removing
components based on the interaction components dened in the literature. We refer to the specic publications that
contributed to each component of COFI in the sections below: for each interaction component the rst paragraph
denes the component, and the second paragraph references the relevant publications that provided the basis for that
component.
3.1 Interaction between Collaborators (Human and AI)
This section presents components related to the relational interaction dynamics between the human and the AI as
co-creators. As shown in Figure 1(a), interaction between collaborators is divided into two subcategories which are
collaboration style and communication style.
3.1.1 Collaboration Style. Collaboration style is the manner of working together in a co-creation. In COFI, collaboration
style comprises participation style, task distribution, timing of initiative and mimicry as interaction components. The
following subsections describe each interaction component in this category.
Participation Style: Participation style in COFI refers to whether the collaborators can participate and contribute
simultaneously, or one collaborator has to wait until the partner nishes a turn. Therefore, participation style in COFI is
8 Rezwana and Maher
categorized as parallel and turn-taking. For example, in a human-AI drawing co-creation, collaborators can take turns
to contribute to the nal drawing or they can draw simultaneously.
Participation style in COFI is based on the categorization of interpersonal interaction into two types: concurrent
interaction and turn-based interaction [
97
]. In concurrent interaction, continuous parallel participation from the
collaborators occurs and in turn-based interaction, participants take turns in contributing. In a parallel participation
style, both collaborators can contribute and interact simultaneously [
124
]. In a turn-taking setting, simultaneous
contribution can not occur [
124
]. In CSCW research, there is a concept for interaction referred to as synchronous and
asynchronous. Synchronous interaction requires the real time interaction where the presence of all collaborators is
required. Whereas asynchronous cooperation does not require simultaneous interaction of all collaborators [
22
,
128
,
132
].
In CSCW, the distinction between synchronous and asynchronous interaction is information exchange in terms of time.
In COFI, participation style describes the way collaborators participate when all are present at the same time.
Task Distribution: Task distribution refers to the distribution of tasks among the collaborators in a co-creative
system. In COFI, there are two types of task distribution, same task and task divided. When it is same task, there is no
division of tasks between collaborators and all the collaborators take part in the same task. For example, in a human-AI
co-creative drawing, both co-creators do the same task, i.e. generating the drawing. In a task-divided distribution, the
main task is divided into specic sub-tasks and the sub-tasks are distributed among the collaborators. For example, in
co-creative poetry, the user can dene the conceptual space for the poetry and generate a poem while the AI agent can
evaluate the poetry.
Cahan and Fewell asserted that division of task is a key factor in the success of social groups [
23
]. According to
Fischer and Mandl, task division should be addressed for co-ordination in a computer-mediated collaboration [
56
]. This
component of COFI emerged from discussions of the two interaction modes presented by Kantosalo and Toivonen:
alternating co-creativity and task divided co-creativity [
79
]. In alternating co-creativity, each party contributes to the
shared artifact while doing the same task by taking turns. Kantosalo and Toivonen emphasized the turn-taking in
alternating interaction mode. In COFI, we renamed alternating co-creativity to be same task as we want to emphasize
the task distribution. Task divided in COFI is the same term used in Kantosalo and Toivenen [79].
Timing of Initiative: In a co-creative setting, the timing of collaborators’ initiative can be scheduled beforehand,
or it can be spontaneous. If the timing of the initiative is planned or xed in advance, in COFI it will be addressed
as planned. If both agents initiate their contribution without any prior plan or xed rules, it will be addressed as
spontaneous. Timing of the initiative should be chosen based on the motivation behind designing a co-creative system.
Spontaneous timing is suitable for increased emergent results, whereas planned timing is more suitable for systems
where users want inspiration or help in a specic way for a particular aspect of the creative process.
Salvador et al. discussed timing of initiative in their framework for evaluating groupware for supporting collaboration
[
134
]. They dened two types of timing of initiative: spontaneous initiatives, where participants take initiatives
spontaneously and pre-planned initiatives, where group interactions are scheduled in advance. Alam et. al divided
interaction among groups into planned and impromptu [
5
]. For COFI, we merged these ways of describing the timing
of initiative into spontaneous and planned.
Mimicry: COFI includes mimicry as a subcategory of collaboration style which is used in co-creative systems as an
intentional strategy for collaboration. When mimicry is a strategy for the AI contribution, the co-creative AI mimics
the human user.
Drawing Apprentice [
40
] is a co-creative web-based drawing system that collaborates with users in real-time abstract
drawing while mimicking users. The authors demonstrated with their ndings that even if the Drawing Apprentice
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
9
mimics the user in the creative process, the system engaged users in the creative process that resulted in generating
novel ideas. An example of a non-mimic co-creative system is Viewpoints AI. Viewpoints AI is a co-creative system
where a human can engage in collaborative dance movement as the system reads and interprets the movement for
responding with an improvised movement [69].
3.1.2 Communication Style. In COFI, communication style refers to the ways humans and AI can communicate.
Communication is an essential component in any collaboration for the co-regulation between the collaborators and
helps the AI agent make decisions in a creative process [
19
]. Communication is critical for achieving understanding
and coordination between collaborators. A signicant challenge in human-AI collaboration is the development of
common ground for communication between humans and machines [
37
]. Collaborators communicate in dierent ways
in a co-creation such as, communication through the shared product and contributions, and communication through
dierent communication channels or modalities. In co-creative systems, collaborators contribute to the shared product
through the creative process and sense-making of each others’ contributions during the process and act accordingly.
Communicating through the shared product is a prerequisite in a co-creation or any collaborative system [
18
]. Hence,
COFI does not include interaction through the shared product under communication style. In COFI, communication style
includes dierent channels or modalities designed to convey intentional and unintentional information between users
and the AI. Human to AI communication channels carry information from users to the AI. On the other hand, AI to
human communication channels carry information from the AI to users.
Human to AI Intentional Communication: Human to AI intentional communication channels represent the
possible ways a human agent can intentionally and purposefully communicate to the AI agent to provide feedback and
convey important information. In COFI, human to AI communication channel includes direct manipulation, voice, text
and embodied communication. The human agent can directly manipulate the co-creative system by clicking buttons
for giving instructions, feedback, or input. It can also provide user preferences by selecting from AI provided options.
Using the whole body or gestures for communicating with the computer will be referred to as embodied. Voice and text
can be also used as intentional communication channels from human to AI.
Gutwin and Greenburg proposed a framework that discusses the mechanics of collaboration for groupware [
61
].
Their framework includes seven major elements and one of them is explicit or intentional communication. Bard dened
intentional communication as the ability to coordinate behavior involving agents [12]. Brink argued that the primary
goal of intentional communication is to establish joint attention [
20
]. In the eld of human-computer interaction, the
communication channel between humans and computers is described as a modality. The modalities for intentional
communication from human to AI include direct manipulation, embodied/gesture, text, and voice [117].
Human to AI Consequential Communication: In COFI, human to AI consequential communication channels
represent the ways the human user unintentionally or unconsciously gives o information to the AI agent. In other
words, this channel represents the ways a co-creative AI agent can track and collect unintentional or consequential
information from the human user such as eye tracking, facial expression tracking, biometric data tracking and embodied
movements. AI agents can track and collect various consequential details from the human to perceive user preference,
user agency and engagement. For example, a posture or facial expression can indicate boredom or lack of interest.
Gutwin and Greenburg reported consequential or unintentional communication as a major element of collabora-
tion mechanics, in addition to intentional communication [
61
]. Collaborators pick up important information that is
unintentionally "given o” by others, which is considered as consequential communication in a human collaboration.
Unintentional communication, such as embodied communication, gaze, biometric measurement and facial expression
10 Rezwana and Maher
are consequential communication [
61
]. Revealing the internal state of an individual is termed ’Nonverbal leakage’
by Ekman and Freisen [
52
]. Mutlu et al. argued that in a human-AI interaction, unintentional cues have a signicant
impact on user experience [114].
AI to Human Communication: AI to human communication represents the channels through which AI can
communicate to humans. Humans expect feedback, critique and evaluation of our contribution from collaborators in
teamwork. If the AI agent could communicate their status, opinion, critique and feedback for a specic contribution, it
would make the co-creation more balanced as the computational agent will be perceived as an intelligent entity and a
co-equal creative partner rather than a mere tool. This communication involves intentional information from the AI to
human. Because the interaction abilities of a co-creative AI agent are programmed, all of the communication from the
AI is intentional. However, one may ask, can AI do anything unintentional or unconscious beyond the programmed
interaction? A co-creative AI can have a body and can make a facial expression of boredom. However, can we call it
unintentional or it is also an intentional information designed to be similar a human’s consequential communication? It
can be an interesting question to ask if consequential communication from the AI to the user is even possible to design.
Mutlu et al. investigated the impact of ’nonverbal leakage’ in robots on human collaborators [
114
], however the leakage
was designed intentionally as part of the interaction design.
In a co-creative setting, the modalities for AI initiated communication can include text, voice, visuals (icons, image,
animation), haptic and embodied communication [
117
]. There are some communication channels that work for both
human to AI and AI to human communication, such as text, voice, and embodied communication. These communication
channels are under both categories to identify the possibilities based on the direction of information ow.
3.2 Interaction with the Shared Product
Interaction components related to the shared creative product in a co-creative setting are discussed in this section and
illustrated in Figure 1(b). Interaction with the shared product is divided into two subcategories, creative contribution to
the product and creative process.
3.2.1 Creative Process. Creative process characterizes the sequence of actions that lead to a novel and creative
production [
101
]. In COFI, there are three types of creative processes that describe the interaction with the shared
product: generate, evaluate, and dene. A co-creative AI can play the role of a generator, evaluator or a dener depending
on the creative process. In the generation creative process, the co-creative AI generates creative ideas or artifacts. For
example, a co-creative AI can generate a poem along with the user or produce music with users. Co-creative AI agents
evaluate the creative contributions made by the user in a creative evaluation process. An example of creative evaluation
will be analyzing and assessing a creative story generated by a user. And in a creative denition process, the AI agent
will dene the creative concept or explore dierent creative concepts along with the user. For example, a co-creative
agent can dene the attributes of a ctional character before a writer starts to write about the character.
The basis of this categorization is the work of Kantosalo et al. that denes the roles of the AI as generator, evaluator,
and concept dener [
79
]. COFI adopts the categorization of Kantosalo et al. as a basis for understanding the range of
potential creative processes: The generator generates artifacts in a specic conceptual description, the evaluator evaluates
these concepts, and the concept dener denes the conceptual space [
79
]. In the recent work of Kantosalo and Jordanous,
they compared their dened roles with the apprentice framework of Negrete-Yankelevich’s and Morales-Zaragoza,
where the roles are generator, apprentice and master [115].
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
11
3.2.2 Creative Product. The creative product is the idea or concept that is being created. Creative product has two
interaction components, contribution type and contribution similarity. We identied these specic components as we
focused on various aspects of contribution making to the shared product as meaning emerges through the contributions
in a collaboration. These components are identied from the literature and discussed in the following subsections.
Contribution Type: In a co-creation, an individual can contribute in dierent ways to the shared product. Co-
creators can generate new elements for the shared product, extend the existing contribution, and modify or rene
the existing contribution. How a co-creator is contributing depends on their interaction with the shared product and
their interpretation of the interaction. The primary contribution types according to COFI are: ’create new’, ’extend’,
’transform’ and ’rene’. ’Extend’ refers to extending or adding on to a previous contribution made by any of the
collaborators. Generating something new or creating new objects is represented by ’create new’, whereas ’transform’
conveys turning a contribution into something totally dierent. ’Rene’ is evaluating and correcting a contribution
with similar type of contribution. For example, in a co-creative drawing, drawing a tree will be considered ’create new’.
Extend is when the collaborator adds a branch to the tree or extends the roots of the tree. Turning a tree branch into
something else, such as a ower, will be considered a ’transformation’, dierent from ’create new’ as it is performed on
a previous contribution to turn it into a new object. ’Rene’ is when the collaborator polishes the branch of the tree to
give more detail.
Contribution types are adopted and adapted from Boden’s categories of computational creativity based on dierent
types of contribution: combinatorial, exploratory, and transformational [
15
]. Combinatorial creativity involves novel
(improbable) combinations of similar ideas to the existing ideas. We adapted ’expand’ and ’rene’ from combinatorial
creativity as ’expand’ is extending the existing contribution and ’rene’ is about correcting or emphasizing the
contribution with similar ideas. Exploratory creativity involves the generation of novel ideas by the exploration of
dened conceptual spaces and ’creating new’ is adapted from this as users use explores the conceptual space when
creating something new. Transformational creativity involves the transformation of some dimension of the space so
that new structures can be generated, which could not have arisen before and ’transform’ is adapted from this.
Contribution Similarity: In COFI, similarity refers to the degree of similarity or association between a new
contribution compared to the contribution of the partner. Near refers to high similarity with the partner’s contribution
and far means less similarity with the partner’s contribution. In this paper, AI agents that use ‘near’ will be referred to
as pleasing agents, and agents that use ‘far’ will be referred to as provoking agents.
Miura and Hida demonstrated that high similarity and low similarity in contributions and ideas among collaborators
are both essential for greater gains in creative performance [
112
]. Both convergent and divergent exploration have
their own value in a creative process. Divergent thinking is "thinking that moves away in diverging directions to
involve a variety of aspects", whereas convergent thinking is demarcated as "thinking that brings together information
focused on something specic" [
1
]. Basadur et al. asserted that divergent thinking is related to the ideation phase and
convergent thinking is related to the evaluation phase [
13
]. Kantosalo et al. dened pleasing and provoking AI agents,
based on how similar their contributions are [
79
]. A pleasing computational agent follows the human user and complies
with the human contribution and preference. Provoking computational agents provoke the human by challenging the
human-provided concepts with divergent ideas and dissimilar contribution.
12 Rezwana and Maher
4 ANALYSIS OF INTERACTION MODELS IN CO-CREATIVE SYSTEMS USING COFI
4.1 Data
We used COFI to analyze a corpus of co-creative systems to demonstrate COFI’s value in describing the interaction
designs of co-creative systems. We initiated our corpus of co-creative systems using the archival website called the
“Library of Mixed-Initiative Creative Interfaces” (LMICI), which archives many of the existing co-creative systems from
the literature [
2
]. Mixed initiative creative systems are often used as an alternative term for co-creative systems [
161
].
Angie Spoto and Natalia Oleynik created this archive after a workshop on mixed-initiative creative interfaces led by
Deterding et al. in 2017 [
2
,
46
]. The archive provides the corresponding literature and other relevant information for
each of the systems. LMICI archive consists of 74 co-creative systems from 1996 to 2017. However, we used 73 systems
from the LMICI archive due to the lack of information regarding one system. We added 19 co-creative systems to our
dataset to include recent co-creative systems (after 2017). We used the keywords ’co-creativity’ and ’human-AI creative
collaboration’ to search for existing co-creative systems from 2017 to 2021 in the ACM digital library and Google scholar.
Thus, we have 92 co-creative systems in the corpus that we used to analyze the interaction designs using COFI. Table 1
shows all the co-creative systems that we analyzed with corresponding years and references. Figure
??
shows the count
of the co-creative systems in our dataset each year.
1 1 1 1
2
1
5
4
8
6 6 6
4 4
16
8
3
4
7
4
0
2
4
6
8
10
12
14
16
18
1996
1999
2000
2001
2003
2005
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
Fig. 3. Counts of Co-creative Systems in the Dataset per Year.
We grouped the systems into 13 categories describing their creative domains. The categories are Painting/Drawing/Art,
Culinary, Dance, Music, Storytelling/Narrative/Writing, Game Design, Theatre/Performance, Video/Animation, Pho-
tography, Poetry, Industrial and Product Design, Graphic Design and Humor/Comic. In Figure 4, the count of the
systems in each category is provided. We see the most common creative domains in the corpus are music, story-
telling/narrative/writing, Game design and Painting/Drawing/art. The distribution shows that some creative domains
are are not well represented in this dataset or rarely used in developing co-creative systems, for example, culinary,
humor, and graphic design.
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
13
Table 1. List of Co-creative Systems in the Dataset Sorted by Year.
Year Co-creative Systems
1996 Improv [125]
1999 GeNotator [149]
2000 NEvAr [103]
2001 Metasynth [38]
2003 Facade [109], continuator [122]
2005 LOGTELL [30]
2008 CombinFormation[83], REQUEST [131], miCollage[159], BeatBender [91], WEVVA [116]
2009 Terrain Sketching [57], JNETIC [14], Synthetic Audience [120], The Poetry Machine [155]
2010 SKETCHAWORLD [142], Tanagra [143], Realtime Generation of Harmonic Progressions [50],
JamBot. [21], Filter Trouve [32], Clap-along [164], EDME [99], LEMu [99],
2011 Shimon [64], Stella [90], Party Quirks [105], Generation of Tracks in a High-end Racing Game [26],
ELVIRA [31], Creating Choreography with Interactive Evolutionary Algorithms [51]
2012 Spaceship Generator [93], MaestroGenesis [148], PINTER [59], Co-PoeTryMe [119],
A formal Architecture of Shared Mental Models [63], Impro-Visor [82]
2013 Sentient Sketchbook [94], Dysphagia [141], Viewpoints AI [69],
Ropossum [48], COCO Sketch[42], Sentient World[94]
2014 Chef Watson [126], Kill the Dragon and Rescue the Princess [89], Nehovah [144], Autodesk Dreamcatcher[119]
2015 CAHOOTS [153], Funky Ikebana [34], StyleMachine [3], Drawing Apprentice [40]
2016 Improvised Ensemble Music Making on Touch Screen [108], AceTalk [150], Chor-rnn [36], Cochoreo [27],
Evolutionary Procedural 2D Map Generation [138], Danesh [35], Plecto [66],
Image-to-Image [68], Robodanza [67], SpeakeSystem [162], TaleBox[28],
ChordRipple [135], Robovie [73], Creative Assistant for Harmonic Blending [74],
Writing Buddy [135], Recommender for Game Mechanics [104]
2017 TOPOSKETCH [154], Trussfab[88], Chimney [113], FabMachine [85],
LuminAI [98], GAIA [60], 3Buddy [102], Deeptingle [84]
2018 The Image Artist [166], DuetDraw [118], Robocinni [4]
2019 In a silent way [110], Metaphoria [58], collabDraw [54], DrawMyPhoto [157]
2020 Shimon the Rapper [136], ALYSIA [29], Cobbie [96], WeMonet [92],
Co-cuild [45], IEC [123], Creative Sketching Partner [80]
2021 BunCho [121], CharacterChat [140], StoryDrawer [165], FashionQ [71]
4.2 Coding Scheme
To analyze the interaction design of the existing co-creative systems, we coded the interaction designs of 92 systems
using COFI. Two coders from our research team independently coded 25% of the systems following COFI. They
then achieved consensus through discussing the disagreements in the codes (Kappa Inter-rater reliability 0.79). The
rest of the systems were coded by a single coder according to the consensus. For each system, the coding shows all
interaction design components according to COFI. All the interaction components of the systems were coded according
to the information provided in the corresponding literature. For a specic interaction component, when none of the
subcategories are present in the interaction design, we coded it as ‘None’.
4.3 Interaction Design Models among Co-creative Systems
For identifying dierent interaction models utilized by the co-creative systems in the dataset, we clustered all the
systems using their interaction components. We used K-modes clustering [
25
,
65
] for identifying clusters as the K-modes
14 Rezwana and Maher
Music
Storytelling/Narrative
Game Design
Painting/Drawing/Art
Dance
Industrial and Product Design
Photography/Digital Art
Theater/Performance
Graphic Design
Poetry
Video/Animation
Culinary
Humor/Comic
0
5
10
15
20
25
22
15
14
13
6 6
5
3
2 2 2
1 1
Fig. 4. Count of Co-Creative Systems in Dierent Creative Domains.
algorithm is suitable for categorical data. K-modes clustering is an extension of K-means, but instead of means, this
algorithm uses modes. For demonstrating the cluster centroids, this algorithm uses modes of all the features. We
used all the interaction components according to COFI as features. We found three clusters of the systems based on
their interaction design (Figure 5). The rst cluster includes 67 co-creative systems and thus indicating a dominant
interaction model. The second cluster includes 9 systems and the third one includes 16 systems. We used chi-square
for determining interaction components that contribute signicantly to forming the clusters and found that all of the
interaction components are signicant factors for the clusters (all P values < 0.05). Figure 5 shows the three major
interaction models, including all the interaction components (cluster centroids represented by feature modes).
4.3.1 Cluster 1 - Interaction Design Model for Generative Pleasing AI Agents. The interaction model of the rst cluster is
the most prevalent as there are 67 systems in this cluster sharing the same or similar model. This dominant interaction
model shows that most of the co-creative systems in the dataset utilize turn-taking as the participation style. Therefore,
each of the collaborators must wait until the partner nishes their turn. This interaction model uses ’planned’ timing
of initiative which is an indication of non-improvisational co-creativity. Hence, most of the systems in the dataset do
not support improvisational creativity. This interaction model uses direct manipulation for human to AI intentional
communication. However, this model does not incorporate any human to AI consequential communication or AI to
human communication. The main task is divided between the collaborators, and the AI agent uses generation as the
creative process in most of the systems in this cluster and creates something new without mimicking the user. The degree
of similarity in contribution is high. In other words, the AI agent pleases the human by generating contributions that
follow along with the contributions made by the human. Mostly, this interaction model is used by non-improvisational
systems that generate creative products to please the users.
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
15
Fig. 5. Interaction Designs for the Three Clusters of Co-creative Systems.
An example of a system that uses this interaction design is Emotion Driven Music Engine (EDME) [
99
]. EDME
generates music based on the emotions of the user. The user selects an emotion, and EDME plays music to match that
emotion. This system works in a turn-taking way with the user. The timing of initiative-taking is planned as the system
will always respond after the human nishes selecting their emotion. The task is divided between the collaborators
as the user denes the conceptual space by choosing an emotion from the interface and the system generates the
music according to that emotion. The system contributes to the collaboration by creating something new and without
mimicking the user. The system creates music that is associated with and similar to the user-dened emotion. The
biggest challenge here is the human can not give any feedback or communicate with the system regarding the generated
music. The system can not track any consequential information from the human such as facial expression, eye gaze
and embodied gestures. Also, the system can not communicate any relevant information to the user such as providing
additional information regarding the contribution or visual cues.
4.3.2 Cluster 2 - Interaction Design Model for Improvisational AI Agents. The interaction design for the systems in
cluster 2 uses parallel participation style where both agents can contribute simultaneously. The task distribution for
16 Rezwana and Maher
these systems is usually ‘same task’ and most of the systems contribute by generating in the creative process. Most
of the systems in this cluster contribute to the collaboration by creating something new and these systems can do
both mimicry and non-mimicry. The degree of similarity in terms of users’ contribution can be both high and low.
This interaction model employs spontaneous initiative-taking while both co-creators contribute to the same task with
parallel participation style, indicating improvisational co-creativity. Systems in this cluster do not have any way of
communication between the user and the system, and a lack of communication in improvisational co-creativity can
reduce the collaboration quality and engagement [64].
An example system for this cluster is LuminAI, where human users improvise with virtual AI agents in real time to
create a dance performance [
98
]. Users move their body and the AI agent will respond with an improvised movement of
its own. Both the AI agent and users can dance simultaneously and they take initiatives spontaneously. Collaborators
contribute to only a single task, generating dance movements. The AI can create new movements and transform user
movements while it can do both mimicry and non-mimicry. The dance movements can be similar or dierent from the
user. There is no way the user can deliberately communicate with the system or the system can communicate with
the user. Here, the creative product itself is an embodied product but the system can not collect any consequential
information from the user such as eye gaze, facial expression or additional gestures other than dance moves.
4.3.3 Cluster 3 - Interaction Design Model for Advisory AI Agents. The third cluster includes systems that work in a
turn taking manner and the task is divided into subtasks between the collaborators. The initiative taking is planned
prior to the collaboration. Users can communicate to the system through direct manipulation, but there is no human
to AI consequential communication channel or AI to human communication channel. The most notable attribute for
this interaction model is both the generation and evaluation ability of the AI agent unlike the other two interaction
models where the AI agent can only contribute by generating. Systems with this interaction model can act as an adviser
to the user by evaluating the contribution of the user. Most of these systems in this cluster contribute by rening
the contribution of the user. These systems do not mimic the contribution of the user and the degree of contribution
similarity can be both high and low.
An example of co-creative systems that utilize this model is Sentient World which assists video game designers in
creating maps [
94
]. The designer creates a rough terrain sketch, and Sentient World evaluates the map created by the
designer and then generates several rened maps as suggestions. This system works in a turn-taking manner with the
user, and the initiative taking is planned. The AI agent uses both generation and evaluation as creative processes by
generating maps and evaluating maps created by the user. The user can communicate with the system minimally with
direct manipulation (clicking buttons) for providing user preference for the maps. The AI agent can not communicate
any explicit information to the human and can not collect any consequential information from the user such as facial
expression, eye gaze and embodied information. Sentient World can both create new maps and rene the map created
by the user. The system does not mimic the user contribution and the similarity with user contribution is high.
4.4 Adoption Rate of the Interaction Components used in the Systems
Figure 6 shows the adoption rate of each of the interaction components in COFI used in the systems. The rst section
of the table comprises interaction components under collaboration style. Turn-taking is the most common participation
style in the dataset (89.1%), while just 10.9% of the systems use parallel participation. Parallel participation is used by
the systems that engage in performative co-creation. Most of the co-creative systems in the dataset use task-divided
distribution of tasks (75%) as they work on separate creative subtasks. 25% systems use same task as their task distribution
as both the user and the AI work on the same creative task/s. Timing of initiative is planned in 86.8% of the systems and
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
17
the rest of the systems take spontaneous initiatives without any xed plan. For mimicry, 90.2% of the systems employ
non-mimicry, 8.7% systems use both mimicry and non-mimicry, and only one system (1.1%) uses mimicry.
The second category, communication style, is concerned with dierent communication channels used by the co-
creative systems. 69.6% systems use direct manipulation as the human to AI communication channel. Voice, embodied
and text is used rarely by the systems. 3.3% of the systems use embodied communication as human to AI consequential
communication and most of the systems (95.7%) do not track and collect any consequential information from the user.
For AI to human communication, most systems do not have any channels. In the next section, we talk about the trend
in communication channels in co-creative systems.
In the creative process category, it is noticeable that the majority of the systems (79.3%) employ generation as the
creative process and 15.2% of the systems use both generation and evaluation as the creative processes. Denition as a
creative process is rarely used in the co-creative systems.
In the creative product category, contribution type is the rst interaction component and most co-creative systems
use create new (59.8%). 10.9% of the systems use both create new and rene as the contribution type. 8.7% of the systems
use both create new and extend as the contribution type.
Collaboration Style
Participation Style
Parallel
Turn-Taking
10.90%
89.10%
Task Distribution
Same Task
Task Divided
25%
75%
Timing of Initiative
Spontaneous
Planned
13.20%
86.80%
Mimicry
Mimic
Non-Mimic
Both
1.10%
90.20%
8.70%
Communication Style
Human to AI
Intentional
Communication
Voice
Direct
Manipulation
Embodied
Text
Direct Manipulation +
Embodied
Voice + Direct
Manipulation
None
1.10%
69.60%
3.30%
2.20%
1.10%
1.10%
21.70%
Human to AI
Consequential
Communication
Gaze
Facial Expression
Biometric
Embodied
None
0%
0%
1.10%
3.30%
95.70%
AI to Human
Communication
Speech
Text
Embodied
Haptic
Visual
Embodied +
Voice
Embodied +
Voice + text
None
1.10%
4.30%
3.30%
0%
5.50%
2.10%
1.10%
82.60%
Creative
Process
Generate
Evaluate
Define
Generate + Define
Generate + Evaluate
79.30%
2.20%
1.10%
2.20%
15.20%
Creative Product
Contribute Type
Create New
Extend
Transform
Refine
Create New
+ Refine
Create New
+ Extend
Create New
+Transform
Transform
+ Refine
59.80%
4.30%
2.20%
5.40%
10.90%
8.70%
7.60%
1.10%
Contribution
Similarity
Low
High
Both
None
2.20%
69.60%
27.10%
1.10%
Fig. 6. Adoption Rate of Each Interaction Component used in the Co-creative Systems in the Dataset.
4.5 Communication in Interaction Models
Our analysis identies a signicant gap in the use of the components of interaction in the co-creative systems in
this dataset: a lack of communication channels between humans and AI (Figure 3). In co-creative systems, subtle
communication happens during the creative process through contributions. For example, in a collaborative drawing co-
creative system where no communication channel exists between the user and the AI, subtle interaction happens through
the shared product as co-creators make sense of each other’s contribution and then make a new contribution. Designing
18 Rezwana and Maher
dierent modalities for communication between the user and the AI has the potential to improve the coordination and
quality of collaboration. However, 82.6% of the systems cannot communicate any feedback or information directly to
the human collaborator other than communicating through the shared product. The rest of the systems communicate
with the users through text, embodied communication, voice, or visuals (image and animation). For human to AI
consequential communication, 95.7% of the systems can not capture any consequential information from the human
user such as facial expression, biometric data, gaze and postures. However, consequential communication can increase
user engagement in collaboration. For the intentional communication from human to AI, most of the systems use
direct manipulation (clicking buttons or selecting options) to communicate (69.6%). In other words, in most of the
systems, users can only minimally communicate with the AI or provide instructions to the AI directly, for example,
through clicking buttons or using sliders. 21.7% of the systems have no way for the user to communicate with the AI
intentionally. The rest of the systems use other intentional communication methods, like embodied communication or
voice or text.
Fig. 7. Distribution of Dierent Kinds of Communication between Humans and AI in the Co-Creative Systems in the Dataset.
Some of the systems in our dataset utilize multiple communication channels. Shimon is a robot that plays the marimba
alongside a human musician [
64
]. Using embodied gestures as visual cues to anticipate each other’s musical input,
Shimon and the musician play an improvised song, responding to each other in real-time. The robot and the human
both use intentional embodied gestures as visual cues to communicate turn-taking and musical beats. Therefore, this
system includes human to AI intentional communication and AI to human communication. Findings from a user study
using Shimon demonstrate that visual cues aid synchronization during improvisational co-creativity. Another system
with interesting communication channels is Robodanza, a humanoid robot that dances with humans [
67
]. Human
dancers use intentional communication by intentionally touching the robot’s head in order to awaken it and the robot
tracks human faces to detect consequential information. The robot is able to detect the noise and rhythm of hands
clapping and tapping on a table. The robot can move its head in the direction of the perceived rhythms and move its
hand following the perceived tempo for communicating its status to the human users.
5 DISCUSSION
We develop and describe COFI to provide a framework for analyzing, comparing, and designing interaction in co-creative
systems. Researchers can use COFI to explore the possible spaces of interaction for choosing an appropriate interaction
design for a specic system. COFI can be benecial while investigating and interpreting the interaction design of
existing co-creative systems. As a framework, COFI is expandable as other interaction components are added in the
future. We analyzed the interaction models of 92 existing co-creative systems using COFI to demonstrate its value in
investigating the trends and gaps in the existing interaction designs in co-creative systems. We identied three major
clusters of interaction models utilized by these systems. In the following paragraphs, we explain the interaction models
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
19
and discuss the potential for further research in specic interaction components. These interaction models can be useful
when designing a co-creative system since they can help identify appropriate interaction components and determine if
interaction components should be modied for the corresponding type of co-creative AI agent.
The most common interaction model in our dataset is suitable for generative co-creative AI agents that follow and
comply with human contributions and ideas by generating similar contributions. Provoking agents are rare in the
literature, and in fact, such a stance seems to be opposed by some in the literature. For example, Tanagra’s creators
ensured "that Tanagra does not push its own agenda on the designer" [
143
]. However, both pleasing and provoking
agents have use-cases within co-creative systems [
79
]. For example, if a user is trying to produce concepts or ideas
that convey their specic style, a pleasing agent that contributes similar ideas is more desirable. However, if a user is
searching for varied ideas, a provoking agent with dierent contributions is an ideal creative partner as it will provide
more divergent ideas. This model can be improved with consequential communication tracking from users and AI to
human communication.
The second interaction model is suitable for improvisational AI agents as it uses spontaneous initiative-taking
and both agents work on the same task in parallel. Additionally, this model includes both mimicry and non-mimicry,
unlike the other models which direct the AI to take proper action in an improvisational performance. This model
can be utilized as a guide while designing interaction in an improvisational co-creative system. However, this model
does not include any intentional or consequential communication channels from humans to AI or AI to humans,
which can negatively impact the collaboration quality and user experience, especially in improvisational co-creativity
where communication is the key. Homan et al. asserted that communication aids synchronization and coordination
in improvisational co-creativity [
64
]. Further research can extend this model by including or extending human-AI
communication channels.
The third interaction model is used by co-creative AI agents that work as an advisor by evaluating user’s contributions
and contributing to the shared product as a generator. In product-based co-creation, AI agents that can both generate
and evaluate help the user generate precise creative ideas and artifacts. For example, in industrial design, the co-creative
AI agent can help in creative ideation by evaluating the user-provided concept for a robust and error-free design and
also help in the generation of the artifact with divergent or convergent ideas [
88
]. AI agents that use this model can
rene the user’s contributions in contrast to the other models. The limitations of this model include the absence of
human to AI consequential communication and AI to human communication.
A notable nding from the analysis of this dataset is the lack of AI agents dening the conceptual space as the
creative process (only 4 out of 92). Most of the systems in the corpus contribute by generating and some contribute by
evaluating the human contributions. In the context of co-creativity, dening the conceptual space is an essential task.
An AI agent can dene the conceptual space without any guidance from the user. For example, the Poetry Machine
is a poetry generator that prompts the user with images that users respond to with a line of poetry [
2
,
155
] and then
organizes the lines of poetry into a poem. An AI agent can also suggest multiple ideas for the conceptual space while
the user can select their preferred one. TopoSketch [
154
] generates animations based on a photo of a face provided by
the human and displays various facial expressions as ideas for the nal animation. CharacterChat inspires writers to
create ctional characters through a conversation. The bot converses with the user to guide the user to dene dierent
attributes of the ctional character. Humans may desire inspiration for creative concepts and ideas at the beginning of a
creative journey. Creative brainstorming and dening creative concepts can be potential research areas for co-creative
systems. There is potential for designing new co-creative systems that both dene the creative conceptual space and
explore it with the user.
20 Rezwana and Maher
The most signicant area of improvement in all of the interaction models identied is communication, the key to
coordination between two agents. Providing feedback, instructions or conveying information about the contribution
is essential for creative collaboration. Without any communication channel between the co-creators, the creation
becomes a silent game [
146
,
147
] as collaborators can not express any concerns and provide feedback about their
contributions. Communication through the creative product is subtle communication and may not be enough to
maintain the coordination and collaboration quality. Most of the existing co-creative systems in our dataset have minimal
communication channels, and this hinders the collaboration ability of the AI agent and the interactive experience. Most
of the systems in the dataset utilize only direct manipulation for communicating intentional information from the users.
Direct manipulations include clicking buttons and using sliders for rating AI contribution, providing simple instructions
and collecting user preferences. For most systems, direct manipulation provides a way for minimal communication
and does not provide users with a way to communicate more broadly. Very few systems in the dataset use other
communication channels other than direct manipulation for human to AI intentional communication. For example,
AFAOSMM (2012) [
63
] is a theatre-based system that uses gestures as intentional communication and Robodanza
(2016) [
67
] uses embodied movements along with direct manipulation for intentional communication. Human to AI
consequential communication is rarely used in the systems but an eective way to improve creative collaboration.
It has been demonstrated that humans, during an interaction, can reason about others’ ideas, goals, intentions and
predict partners’ behaviors, a capability called Theory of Mind (ToM) [
10
,
127
,
163
]. Having a Theory of Mind allows
us to infer the mental states of others that are not directly observable, enabling us to engage in daily interaction. The
ability to intuit what others think or want from brief nonverbal interactions is crucial to our social lives as we see
others’ behavior not just as motions but as an intentional action. In a collaboration, Theory of Mind is essential to
observe and interpret the behavior of a partner, maintain coordination and act accordingly. Collecting unintentional
information from the human partner has the potential to improve the collaboration and user experience in a human-AI
co-creation, and may lead to enabling AI to mimic the Theory of Mind ability of humans. The technology for collecting
consequential information from the user includes eye trackers, facial expression trackers, gesture recognition devices,
and cognitive signal tracking devices.
AI to human communication channels are also rarely utilized in the identied interaction models. However, it is
essential to understand the AI partner by the users to build an engaging and trustworthy partnership. Many intelligent
systems lack the core interaction design principles such as transparency and explainability and it makes them hard to
understand and use [
49
]. To address the challenge of transparency of AI interaction should be designed to support
users in understanding and dealing with intelligent systems despite their complex black-box nature. When AI can
communicate its decision making process to users and explain its contribution, the system becomes more comprehensible
and transparent to build a partnership. So, AI to human communication is critical for interaction design in co-creative
systems. Visuals, text, voice, embodied, and haptic feedback can be used to convey information, suggestions, and
feedback to the users. There is a distinction between AI to human communication and AI steerability. For example,
LuminAI is a co-creative AI that dances with humans [
98
]. Here the generated creative product is dance, an embodied
product created by gestures and embodied movements. However, AI can only communicate by contributing to the
product and does not directly communicate to humans. Humans can steer the AI by contributing dierent embodied
contributions to the nal product and the AI generates contributions based on the user movements. This is dierent
from embodied communication that intentionally communicates that the collaborator is doing great with a thumbs
up. The gap in interaction design in terms of communication is an area of future research for the eld of co-creativity.
User experiments with dierent interaction models can help identify eective interaction design for dierent types
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
21
of co-creative systems [
129
]. COFI provides a common framework for analyzing the interaction designs in existing
co-creative systems to identify trends and gaps in existing interaction designs for designing improved interaction in a
co-creative system.
AI is being used increasingly in collaborative spaces, for example, recommender systems, self-driving vehicles, and
health care. Much AI research has focused on improving the intelligence or ability of agents and algorithms [
8
]. As AI
technology shifts from computers to everyday devices, AI needs social understanding and cooperative intelligence to
integrate into society and our daily lives. AI is, however, a novice when it comes to collaborating with humans [
37
]. The
term ’human-AI collaboration’ has emerged in recent work studying user interaction with AI systems [
7
,
24
,
118
,
151
].
This marks both a shift to a collaborative from an automated perspective of AI, and the advancement of AI capabilities to
be a collaborative partner in some domains. Ashktorab et al. asserted that human-AI co-creation could be a starting point
of designing and developing AI that can cooperate with humans [
8
]. Human-AI interaction has many challenges and is
dicult to design [
160
]. HCI deals with complex technologies, including research to mitigate unexpected consequences.
A critical rst step in designing valuable human-AI interactions is to identify technical challenges, articulate the unique
qualities of AI that make it dicult to design, and then develop insights for future research [
160
]. Building a fair and
eective AI application is considered dicult due to the complexity both in dening the goals and algorithmically
achieving the dened goals. Prior research has addressed these challenges by promoting interaction design guidelines
[
6
,
111
]. In this paper, we provide COFI as a framework to describe the possible interaction spaces in human-AI creative
collaboration and identify existing trends and gaps in existing interaction designs. COFI can also be useful in AI research
and HCI research to design cooperative AI in dierent domains. COFI will expand as we learn and identify more aspects
of human-AI collaboration.
6 LIMITATIONS
The identication of clusters of interaction models in human-AI co-creative systems is limited to the specic dataset
that we used for the analysis. Although we believe this sample contains a large population, the systems in the dataset
are limited by the expectations and technologies at the time of publication. We expect the clusters and descriptions of
interaction models for co-creative systems will change over time.
7 CONCLUSIONS
This paper develops and describes the COFI as a framework for modeling interaction in co-creative systems. COFI
was used to analyze the interaction design of 92 co-creative systems from the literature. Three interaction models for
co-creative systems were identied: generative pleasing agents, improvisational agents, and advisory agents. When
developing a co-creative system, these interaction models can be useful to choose suitable interaction components for
corresponding co-creative systems. COFI is broader than the interaction designs utilized in any specic co-creative
system in the data set. The ndings show that the space of possibilities is underutilized. While the analysis is limited
to the data set, it demonstrates that COFI can be a tool for identifying research directions and research gaps in the
current space of co-creativity. COFI revealed a general lack of communication in co-creative systems within the dataset.
In particular, very few systems incorporate AI to human communication, communication channels other than direct
manipulation for collecting intentional information from humans and gathering consequential communication data,
such as eye gaze, biometric data, gesture, and emotion. This gap demonstrates an area of future research for the eld of
co-creativity. We argue that COFI will provide useful guidelines for interaction modeling while developing co-creative
systems. As a framework, COFI is expandable as other interaction components can be added to it in the future. User
22 Rezwana and Maher
experiments with dierent interaction models can help identify eective interaction design for dierent types of
co-creative systems and lead to insights into factors that aect user engagement.
REFERENCES
[1] [n.d.]. Dictionary, Encyclopedia and Thesaurus. https://www.thefreedictionary.com/
[2] [n.d.]. Library of Mixed-Initiative Creative Interfaces. http://mici.codingconduct.cc/. (Accessed on 05/31/2020).
[3] 2015. Style Machine Lite. https://metacreativetech.com/products/stylemachine-lite/
[4]
Margareta Ackerman, James Morgan, and Christopher Cassion. 2018. Co-Creative Conceptual Art. In Proceedings of the Ninth International
Conference on Computational Creativity. 1–8.
[5]
Aftab Alam, Sehat Ullah, Shah Khalid, Fakhrud Din, and Ihsan Rabbi. 2013. Computer Supported Collaborative Work (CSCW) and Network Issues:
A Survey. International Information Institute (Tokyo). Information 16, 11 (2013), 7995.
[6]
Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul N Bennett, Kori
Inkpen, et al. 2019. Guidelines for human-AI interaction. In Proceedings of the 2019 chi conference on human factors in computing systems. 1–13.
[7]
Ines Arous, Jie Yang, Mourad Khayati, and Philippe Cudré-Mauroux. 2020. Opencrowd: A human-ai collaborative approach for nding social
inuencers via open-ended answers aggregation. In Proceedings of The Web Conference 2020. 1851–1862.
[8]
Zahra Ashktorab, Q Vera Liao, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. 2020. Human-ai
collaboration in a cooperative game setting: Measuring social perception and outcomes. Proceedings of the ACM on Human-Computer Interaction 4,
CSCW2 (2020), 1–20.
[9] Ronald M Baecker. 1993. Readings in groupware and computer-supported cooperative work: Assisting human-human collaboration. Elsevier.
[10]
Chris L Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B Tenenbaum. 2017. Rational quantitative attribution of beliefs, desires and percepts
in human mentalizing. Nature Human Behaviour 1, 4 (2017), 1–10.
[11]
Kevin Baker, Saul Greenberg, and Carl Gutwin. 2001. Heuristic evaluation of groupware based on the mechanics of collaboration. In IFIP International
Conference on Engineering for Human-Computer Interaction. Springer, 123–139.
[12]
Kim A Bard. 1992. Intentional behavior and intentional communication in young free-ranging orangutans. Child development 63, 5 (1992),
1186–1197.
[13]
Min Basadur and Peter A Hausdorf. 1996. Measuring divergent thinking attitudes related to creative problem solving and innovation management.
Creativity Research Journal 9, 1 (1996), 21–32.
[14]
Steve R Bergen. 2009. Evolving stylized images using a user-interactive genetic algorithm. In Proceedings of the 11th Annual Conference Companion
on Genetic and Evolutionary Computation Conference: Late Breaking Papers. 2745–2752.
[15] Margaret A Boden. 1998. Creativity and articial intelligence. Articial Intelligence 103, 1-2 (1998), 347–356.
[16] Oliver Bown. 2014. Empirically Grounding the Evaluation of Creative Systems: Incorporating Interaction Design.. In ICCC. 112–119.
[17]
Oliver Bown. 2015. Player Responses to a Live Algorithm: Conceptualising computational creativity without recourse to human comparisons?. In
ICCC. 126–133.
[18]
Oliver Bown and Andrew R Brown. 2018. Interaction design for metacreative systems. In New Directions in Third Wave Human-Computer Interaction:
Volume 1-Technologies. Springer, 67–87.
[19]
Oliver Bown, Kazjon Grace, Liam Bray, and Dan Ventura. 2020. A Speculative Exploration of the Role of Dialogue in Human-ComputerCo-creation..
In ICCC. 25–32.
[20]
Ingar Brinck. 2008. The role of intersubjectivity in the development of intentional communication. The shared mind: Perspectives on intersubjectivity
(2008), 115–140.
[21]
Andrew R Brown, Toby Giord, and Rene Wooller. 2010. Generative Music Systems for Live Performance. In First International Conference on
Computational Intelligence. 290.
[22]
Diletta Cacciagrano and Flavio Corradini. 2001. On synchronous and asynchronous communication paradigms. In Italian Conference on Theoretical
Computer Science. Springer, 256–268.
[23]
Sara Helms Cahan and Jennifer H Fewell. 2004. Division of labor and the evolution of task sharing in queen associations of the harvester ant
Pogonomyrmex californicus. Behavioral ecology and sociobiology 56, 1 (2004), 9–17.
[24]
Carrie J Cai, Samantha Winter, David Steiner, Lauren Wilcox, and Michael Terry. 2019. " Hello AI": Uncovering the Onboarding Needs of Medical
Practitioners for Human-AI Collaborative Decision-Making. Proceedings of the ACM on Human-computer Interaction 3, CSCW (2019), 1–24.
[25]
Fuyuan Cao, Jiye Liang, Deyu Li, Liang Bai, and Chuangyin Dang. 2012. A dissimilarity measure for the k-Modes clustering algorithm. Knowledge-
Based Systems 26 (2012), 120–127.
[26]
Luigi Cardamone, Daniele Loiacono, and Pier Luca Lanzi. 2011. Interactive evolution for the procedural generation of tracks in a high-end racing
game. In Proceedings of the 13th annual conference on Genetic and evolutionary computation. 395–402.
[27]
Kristin Carlson, Philippe Pasquier, Herbert H Tsang, Jordon Phillips, Thecla Schiphorst, and Tom Calvert. 2016. Cochoreo: A generative feature in
idanceforms for creating novel keyframe animation for choreography. In Proceedings of the 7th International Conference on Computational Creativity.
380–387.
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
23
[28] O Castaño Pérez, BA Kybartas, and Rafael Bidarra. 2016. TaleBox: A mobile game for mixed-initiative story creation. (2016).
[29]
Lee Cheatley, Margareta Ackerman, Alison Pease, and Wendy Moncur. 2020. Co-creative songwriting for bereavement support. In Eleventh
International Conference on Computational Creativity: ICCC’20. Association for Computational Creativity, 33–41.
[30]
Angelo EM Ciarlini, Cesar T Pozzer, Antonio L Furtado, and Bruno Feijó. 2005. A logic-based tool for interactive generation and dramatization of
stories. In Proceedings of the 2005 ACM SIGCHI International Conference on Advances in computer entertainment technology. 133–140.
[31]
Simon Colton, Michael Cook, and Azalea Raad. 2011. Ludic considerations of tablet-based evo-art. In European Conference on the Applications of
Evolutionary Computation. Springer, 223–233.
[32] Simon Colton, Jeremy Gow, Pedro Torres, and Paul A Cairns. 2010. Experiments in Objet Trouvé Browsing.. In ICCC. 238–247.
[33] Simon Colton, Geraint A Wiggins, et al. 2012. Computational creativity: The nal frontier?. In Ecai, Vol. 12. Montpelier, 21–26.
[34] Kate Compton and Michael Mateas. 2015. Casual Creators.. In ICCC. 228–235.
[35] Michael Cook, Jeremy Gow, and Simon Colton. 2016. Danesh: Helping bridge the gap between procedural generators and their output. (2016).
[36] Luka Crnkovic-Friis and Louise Crnkovic-Friis. 2016. Generative choreography using deep learning. arXiv preprint arXiv:1605.06921 (2016).
[37]
Allan Dafoe, Yoram Bachrach, Gillian Hadeld, Eric Horvitz, Kate Larson, and Thore Graepel. 2021. Cooperative AI: machines must learn to nd
common ground.
[38] Palle Dahlstedt. 2001. A MutaSynth in parameter space: interactive composition through evolution. Organised Sound 6, 2 (2001), 121–124.
[39]
Nicholas Davis, Chih-Pin Hsiao, Yanna Popova, and Brian Magerko. 2015. An enactive model of creativity for computational collaboration and
co-creation. In Creativity in the Digital Age. Springer, 109–133.
[40]
Nicholas Davis, Chih-PIn Hsiao, Kunwar Yashraj Singh, Lisa Li, Sanat Moningi, and Brian Magerko. 2015. Drawing apprentice: An enactive
co-creative agent for artistic collaboration. In Proceedings of the 2015 ACM SIGCHI Conference on Creativity and Cognition. 185–186.
[41]
Nicholas Davis, Chih-PIn Hsiao, Kunwar Yashraj Singh, Lisa Li, and Brian Magerko. 2016. Empirically studying participatory sense-making in
abstract drawing with a co-creative cognitive agent. In Proceedings of the 21st International Conference on Intelligent User Interfaces. 196–207.
[42]
Nicholas Mark Davis. 2013. Human-computer co-creativity: Blending human and computational creativity. In Ninth Articial Intelligence and
Interactive Digital Entertainment Conference.
[43] Hanne De Jaegher. 2013. Embodiment and sense-making in autism. Frontiers in integrative neuroscience 7 (2013), 15.
[44] Hanne De Jaegher and Ezequiel Di Paolo. 2007. Participatory sense-making. Phenomenology and the cognitive sciences 6, 4 (2007), 485–507.
[45]
Manoj Deshpande. 2020. Towards Co-build: An Architecture Machine for Co-creative Form-making. Ph.D. Dissertation. The University of North
Carolina at Charlotte.
[46]
Sebastian Deterding, Jonathan Hook, Rebecca Fiebrink, Marco Gillies, Jeremy Gow, Memo Akten, Gillian Smith, Antonios Liapis, and Kate Compton.
2017. Mixed-initiative creative interfaces. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems.
628–635.
[47] Arne Dietrich. 2004. The cognitive neuroscience of creativity. Psychonomic bulletin & review 11, 6 (2004), 1011–1026.
[48]
Steve DiPaola, Graeme McCaig, Kristin Carlson, Sara Salevati, and Nathan Sorenson. 2013. Adaptation of an Autonomous Creative Evolutionary
System for Real-World Design Application Based on Creative Cognition.. In ICCC. 40–47.
[49]
Malin Eiband, Daniel Buschek, and Heinrich Hussmann. 2021. How to support users in understanding intelligent systems? Structuring the
discussion. In 26th International Conference on Intelligent User Interfaces. 120–132.
[50]
Arne Eigenfeldt and Philippe Pasquier. 2010. Realtime generation of harmonic progressions using controlled Markov selection. In Proceedings of
ICCC-X-Computational Creativity Conference. 16–25.
[51]
Jonathan Eisenmann, Benjamin Schroeder, Matthew Lewis, and Rick Parent. 2011. Creating choreography with interactive evolutionary algorithms.
In European Conference on the Applications of Evolutionary Computation. Springer, 293–302.
[52] Paul Ekman and Wallace V Friesen. 1969. Nonverbal leakage and clues to deception. Psychiatry 32, 1 (1969), 88–106.
[53]
Daniel Fallman. 2008. The interaction design research triangle of design practice, design studies, and design exploration. Design Issues 24, 3 (2008),
4–18.
[54]
Judith E Fan, Monica Dinculescu, and David Ha. 2019. Collabdraw: an environment for collaborative sketching with an articial agent. In
Proceedings of the 2019 on Creativity and Cognition. 556–561.
[55]
Valentina Fantasia, Hanne De Jaegher, and Alessandra Fasulo. 2014. We can work it out: an enactive look at cooperation. Frontiers in psychology 5
(2014), 874.
[56] Frank Fischer and Heinz Mandl. 2003. Being there or being where? Videoconferencing and cooperative learning. na.
[57]
James Gain, Patrick Marais, and Wolfgang Straßer. 2009. Terrain sketching. In Proceedings of the 2009 symposium on Interactive 3D graphics and
games. 31–38.
[58]
Katy Ilonka Gero and Lydia B Chilton. 2019. Metaphoria: An algorithmic companion for metaphor creation. In Proceedings of the 2019 CHI
Conference on Human Factors in Computing Systems. 1–12.
[59]
Stephen Gilroy, Julie Porteous, Fred Charles, and Marc Cavazza. 2012. Exploring passive user interaction for adaptive narratives. In Proceedings of
the 2012 ACM international conference on Intelligent User Interfaces. 119–128.
[60]
Ashok K Goel and Spencer Rugaber. 2017. GAIA: A CAD-like environment for designing game-playing agents. IEEE Intelligent Systems 32, 3 (2017),
60–67.
24 Rezwana and Maher
[61]
Carl Gutwin, Saul Greenberg, and Mark Roseman. 1996. Workspace awareness in real-time distributed groupware: Framework, widgets, and
evaluation. In People and Computers XI. Springer, 281–298.
[62] Matthew Guzdial and Mark Riedl. 2019. An interaction framework for studying co-creative ai. arXiv preprint arXiv:1903.09709 (2019).
[63]
Rania Hodhod and Brian Magerko. 2016. Closing the cognitive gap between humans and interactive narrative agents using shared mental models.
In Proceedings of the 21st International Conference on Intelligent User Interfaces. 135–146.
[64] Guy Homan and Gil Weinberg. 2011. Interactive improvisation with a robotic marimba player. Autonomous Robots 31, 2-3 (2011), 133–153.
[65]
Zhexue Huang. 1997. Clustering large data sets with mixed numeric and categorical values. In Proceedings of the 1st pacic-asia conference on
knowledge discovery and data mining,(PAKDD). Singapore, 21–34.
[66]
Stean Ianigro and Oliver Bown. 2016. Plecto: a low-level interactive genetic algorithm for the evolution of audio. In International Conference on
Computational Intelligence in Music, Sound, Art and Design. Springer, 63–78.
[67]
I Infantino, A Augello, A Manfré, G Pilato, and F Vella. 2016. Robodanza: Live performances of a creative dancing humanoid. In Proceedings of the
Seventh International Conference on Computational Creativity. 388–395.
[68]
Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and AlexeiA Efros. 2017. Image-to-image translation with conditional adversarial networks. In Proceedings
of the IEEE conference on computer vision and pattern recognition. 1125–1134.
[69]
Mikhail Jacob, Alexander Zook, and Brian Magerko. 2013. Viewpoints AI: Procedurally Representing and Reasoning about Gestures.. In DiGRA
conference.
[70]
Kyle E Jennings, Dean Keith Simonton, and Stephen E Palmer. 2011. Understanding exploratory creativity in a visual domain. In Proceedings of the
8th ACM conference on Creativity and cognition. 223–232.
[71]
Youngseung Jeon, Seungwan Jin, Patrick C Shih, and Kyungsik Han. 2021. FashionQ: An AI-Driven Creativity Support Tool for Facilitating Ideation
in Fashion Design. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. 1–18.
[72]
Rex Eugene Jung, Brittany S Mead, Jessica Carrasco, and Ranee A Flores. 2013. The structure of creative cognition in the human brain. Frontiers in
human neuroscience 7 (2013), 330.
[73]
Peter H Kahn, Takayuki Kanda, Hiroshi Ishiguro, Brian T Gill, Solace Shen, Jolina H Ruckert, and Heather E Gary. 2016. Human creativity can
be facilitated through interacting with a social robot. In 2016 11th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE,
173–180.
[74]
Maximos Kaliakatsos-Papakostas, Roberto Confalonieri, Joseph Corneli, Asterios Zacharakis, and Emilios Cambouropoulos. 2016. An argument-
based creative assistant for harmonic blending. arXiv preprint arXiv:1603.01770 (2016).
[75]
Nabil N Kamel and Robert M Davison. 1998. Applying CSCW technology to overcome traditional barriers in group interactions. Information &
Management 34, 4 (1998), 209–219.
[76] Anna Kantosalo and Anna Jordanous. 2020. Role-based perceptions of computer participants in human-computer co-creativity. AISB.
[77]
Anna Kantosalo, Prashanth Thattai Ravikumar, Kazjon Grace, and Tapio Takala. 2020. Modalities, Styles and Strategies: An Interaction Framework
for Human-Computer Co-Creativity.. In ICCC. 57–64.
[78]
Anna Kantosalo, Jukka M Toivanen, Ping Xiao, and Hannu Toivonen. 2014. From Isolation to Involvement: Adapting Machine Creativity Software
to Support Human-Computer Co-Creation.. In ICCC. 1–7.
[79]
Anna Kantosalo and Hannu Toivonen. 2016. Modes for creative human-computer collaboration: Alternating and task-divided co-creativity. In
Proceedings of the seventh international conference on computational creativity. 77–84.
[80]
Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. 2020. Creative sketching partner: an analysis of human-AI
co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces. 221–230.
[81]
Jody Koenig Kellas and April R Trees. 2005. Rating interactional sense-making in the process of joint storytelling. The sourcebook of nonverbal
measures: Going beyond words (2005), 281.
[82] Robert M Keller. 2012. Continuous improvisation and trading with Impro-Visor. (2012).
[83]
Andruid Kerne, Eunyee Koh, Steven M Smith, Andrew Webb, and Blake Dworaczyk. 2008. combinFormation: Mixed-initiative composition of
image and text surrogates promotes information discovery. ACM Transactions on Information Systems (TOIS) 27, 1 (2008), 1–45.
[84] Ahmed Khalifa, Gabriella AB Barros, and Julian Togelius. 2017. Deeptingle. arXiv preprint arXiv:1705.03557 (2017).
[85]
Jeeeun Kim, Haruki Takahashi, Homei Miyashita, Michelle Annett, and Tom Yeh. 2017. Machines as co-designers: A ction on the future of
human-fabrication machine interaction. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems.
790–805.
[86]
Gary Klein, Brian Moon, and Robert R Homan. 2006. Making sense of sensemaking 1: Alternative perspectives. IEEE intelligent systems 21, 4
(2006), 70–73.
[87] Jon Kolko. 2010. Thoughts on interaction design. Morgan Kaufmann.
[88]
Robert Kovacs, Anna Seufert, Ludwig Wall, Hsiang-Ting Chen, Florian Meinel, Willi Müller, Sijing You, Maximilian Brehm, Jonathan Striebel,
Yannis Kommana, et al
.
2017. Trussfab: Fabricating sturdy large-scale structures on desktop 3d printers. In Proceedings of the 2017 CHI Conference
on Human Factors in Computing Systems. 2606–2616.
[89]
Iván M Laclaustra, José Ledesma, Gonzalo Méndez, and Pablo Gervás. 2014. Kill the Dragon and Rescue the Princess: Designing a Plan-based
Multi-agent Story Generator.. In ICCC. 347–350.
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
25
[90] Carlos León. 2011. Stella-a story generation system for generic scenarios. In Proceedings of the Second International Conference on Computational
Creativity.
[91]
Aaron Levisohn and Philippe Pasquier. 2008. BeatBender: subsumption architecture for autonomous rhythm generation. In Proceedings of the 2008
International Conference on Advances in Computer Entertainment Technology. 51–58.
[92]
Zhuying Li, Yan Wang, Wei Wang, Stefan Greuter, and Florian’Floyd’ Mueller. 2020. Empowering a Creative City: Engage Citizens in Creating
Street Art through Human-AI Collaboration. In Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems. 1–8.
[93]
Antonios Liapis, Georgios N Yannakakis, and Julian Togelius. 2012. Co-creating game content using an adaptive model of user taste. In 3rd
International Conference on Computational Creativity.
[94]
Antonios Liapis, Georgios N Yannakakis, and Julian Togelius. 2013. Sentient world: Human-based procedural cartography. In International
Conference on Evolutionary and Biologically Inspired Music and Art. Springer, 180–191.
[95] Antonios Liapis, Georgios N Yannakakis, and Julian Togelius. 2014. Computational game creativity. ICCC.
[96]
Yuyu Lin, Jiahao Guo, Yang Chen, Cheng Yao, and Fangtian Ying. 2020. It is your turn: collaborative ideation with a co-creative robot through
sketch. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 1–14.
[97]
Tao Liu, Hirofumi Saito, and Misato Oi. 2015. Role of the right inferior frontal gyrus in turn-based cooperation and competition: a near-infrared
spectroscopy study. Brain and cognition 99 (2015), 17–23.
[98]
Duri Long, Mikhail Jacob, Nicholas Davis, and Brian Magerko. 2017. Designing for socially interactive systems. In Proceedings of the 2017 ACM
SIGCHI Conference on Creativity and Cognition. 39–50.
[99] Alex Rodriguez Lopez, Antonio Pedro Oliveira, and Amílcar Cardoso. 2010. Real-Time Emotion-Driven Music Engine.. In ICCC. 150–154.
[100]
Ryan Louie, Andy Coenen, Cheng Zhi Huang, Michael Terry, and Carrie J Cai. 2020. Novice-AI music co-creation via AI-steering tools for deep
generative models. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 1–13.
[101] Todd I Lubart. 2001. Models of the creative process: Past, present and future. Creativity research journal 13, 3-4 (2001), 295–308.
[102] Pedro Lucas and Carlos Martinho. 2017. Stay Awhile and Listen to 3Buddy, a Co-creative Level Design Support Tool.. In ICCC. 205–212.
[103]
Penousal Machado and Amílcar Cardoso. 2000. NEvAr–the assessment of an evolutionary art tool. In Proceedings of the AISB00 Symposium on
Creative & Cultural Aspects and Applications of AI & Cognitive Science, Birmingham, UK, Vol. 456.
[104] Tiago Machado, Ivan Bravi, Zhu Wang, Andy Nealen, and Julian Togelius. 2016. Shopping for Game Mechanics. (2016).
[105]
Brian Magerko, Christopher DeLeon, and Peter Dohogne. 2011. Digital improvisational theatre: party quirks. In International Workshop on Intelligent
Virtual Agents. Springer, 42–47.
[106] Mary Lou Maher. 2012. Computational and collective creativity: Who’s being creative?. In ICCC. Citeseer, 67–71.
[107] Lena Mamykina, Linda Candy, and Ernest Edmonds. 2002. Collaborative creativity. Commun. ACM 45, 10 (2002), 96–99.
[108]
Charles Martin, Henry Gardner, Ben Swift, and Michael Martin. 2016. Intelligent agents and networked buttons improve free-improvised ensemble
music-making on touch-screens. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. 2295–2306.
[109]
Michael Mateas and Andrew Stern. 2003. Façade: An experiment in building a fully-realized interactive drama. In Game developers conference,
Vol. 2. 4–8.
[110]
Jon McCormack, Toby Giord, Patrick Hutchings, Maria Teresa Llano Rodriguez, Matthew Yee-King, and Mark d’Inverno. 2019. In a silent way:
Communication between ai and improvising musicians beyond sound. In Proceedings of the 2019 CHI Conference on Human Factors in Computing
Systems. 1–11.
[111]
Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and
Timnit Gebru. 2019. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency. 220–229.
[112]
Asako Miura and Misao Hida. 2004. Synergy between diversity and similarity in group-idea generation. Small Group Research 35, 5 (2004), 540–564.
[113] Fabio Morreale, Raul Masu, et al. 2017. Renegotiating responsibilities in human-computer ensembles. (2017).
[114]
Bilge Mutlu, Fumitaka Yamaoka, Takayuki Kanda, Hiroshi Ishiguro, and Norihiro Hagita. 2009. Nonverbal leakage in robots: communication of
intentions through seemingly unintentional behavior. In Proceedings of the 4th ACM/IEEE international conference on Human robot interaction.
69–76.
[115]
Santiago Negrete-Yankelevich and Nora Morales Zaragoza. 2014. The apprentice framework: planning and assessing creativity.. In ICCC. 280–283.
[116]
Mark Nelson, Swen Gaudl, Simon Colton, Edward Powley, Blanca Perez Ferrer, Rob Saunders, Peter Ivey, and Michael Cook. 2017. Fluidic games in
cultural contexts. (2017).
[117] Laurence Nigay. 2004. Design space for multimodal interaction. In Building the Information Society. Springer, 403–408.
[118]
Changhoon Oh, Jungwoo Song, Jinhan Choi, Seonghyeon Kim, Sungwoo Lee, and Bongwon Suh. 2018. I lead, you help but only with enough
details: Understanding user experience of co-creation with articial intelligence. In Proceedings of the 2018 CHI Conference on Human Factors in
Computing Systems. 1–13.
[119]
Hugo Gonçalo Oliveira, Raquel Hervás, Alberto Díaz, and Pablo Gervás. 2014. Adapting a Generic Platform for Poetry Generation to Produce
Spanish Poems.. In ICCC. 63–71.
[120] Brian O’Neill and Mark Riedl. 2011. Simulating the Everyday Creativity of Readers.. In ICCC. 153–158.
[121]
Hiroyuki Osone, Jun-Li Lu, and Yoichi Ochiai. 2021. BunCho: AI Supported Story Co-Creation via Unsupervised Multitask Learning to Increase
Writers’ Creativity in Japanese. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems. 1–10.
[122] Francois Pachet. 2003. The continuator: Musical interaction with style. Journal of New Music Research 32, 3 (2003), 333–341.
26 Rezwana and Maher
[123] Jéssica Parente, Tiago Martins, Joao Bicker, and Penousal Machado. 2020. Which type is your type?. In ICCC. 476–483.
[124]
Victor M Ruiz Penichet, Ismael Marin, Jose A Gallud, María Dolores Lozano, and Ricardo Tesoriero. 2007. A classication method for CSCW
systems. Electronic Notes in Theoretical Computer Science 168 (2007), 237–247.
[125]
Ken Perlin and Athomas Goldberg. 1996. Improv: A system for scripting interactive actors in virtual worlds. In Proceedings of the 23rd annual
conference on Computer graphics and interactive techniques. 205–216.
[126]
Florian Pinel and Lav R Varshney. 2014. Computational creativity for culinary recipes. In CHI’14 Extended Abstracts on Human Factors in Computing
Systems. 439–442.
[127] David Premack and Guy Woodru. 1978. Does the chimpanzee have a theory of mind? Behavioral and brain sciences 1, 4 (1978), 515–526.
[128]
Walter Reinhard, Jean Schweitzer, Gerd Volksen, and Michael Weber. 1994. CSCW tools: concepts and architectures. Computer 27, 5 (1994), 28–36.
[129]
Jeba Rezwanaa, Mary Lou Mahera, and Nicholas Davisa. 2021. Creative PenPal: A Virtual Embodied Conversational AI Agent to Improve User
Engagement and Collaborative Experience in Human-AI Co-Creative Design Ideation. (2021).
[130] Mel Rhodes. 1961. An analysis of creativity. The Phi delta kappan 42, 7 (1961), 305–310.
[131]
Mark Riedl, Jonathan Rowe, and David K Elson. 2008. Toward intelligent support of authoring machinima media content: story and visualization.
(2008).
[132]
Tom Rodden and Gordon Blair. 1991. CSCW and distributed systems: The problem of control. In Proceedings of the Second European Conference on
Computer-Supported Cooperative Work ECSCW’91. Springer, 49–64.
[133]
Daniel M Russell, Mark J Stek, Peter Pirolli, and Stuart K Card. 1993. The cost structure of sensemaking. In Proceedings of the IN TERACT’93 and
CHI’93 conference on Human factors in computing systems. 269–276.
[134] Tony Salvador, Jean Scholtz, and James Larson. 1996. The Denver model for groupware design. ACM SIGCHI Bulletin 28, 1 (1996), 52–58.
[135]
Ben Samuel, Michael Mateas, and Noah Wardrip-Fruin. 2016. The design of Writing Buddy: a mixed-initiative approach towards computational
story collaboration. In International Conference on Interactive Digital Storytelling. Springer, 388–396.
[136]
Richard Savery, Lisa Zahray, and Gil Weinberg. 2020. Shimon the Rapper: A Real-Time System for Human-Robot Interactive Rap Battles. arXiv
preprint arXiv:2009.09234 (2020).
[137]
R Keith Sawyer and Stacy DeZutter. 2009. Distributed creativity: How collective creations emerge from collaboration. Psychology of aesthetics,
creativity, and the arts 3, 2 (2009), 81.
[138]
Andreas Scheibenpug, Johannes Karder, Susanne Schaller, Stefan Wagner, and Michael Aenzeller. 2016. Evolutionary Procedural 2D Map
Generation using Novelty Search. In Proceedings of the 2016 on Genetic and Evolutionary Computation Conference Companion. 39–40.
[139] Kjeld Schmidt. 2008. Cooperative work and coordinative practices. In Cooperative Work and Coordinative Practices. Springer, 3–27.
[140]
Oliver Schmitt and Daniel Buschek. 2021. CharacterChat: Supporting the Creation of Fictional Characters through Conversation and Progressive
Manifestation with a Chatbot. arXiv preprint arXiv:2106.12314 (2021).
[141]
Noor Shaker, Mohammad Shaker, and Julian Togelius. 2013. Ropossum: An authoring tool for designing, optimizing and solving cut the rope levels.
In Ninth Articial Intelligence and Interactive Digital Entertainment Conference.
[142]
Ruben Michaël Smelik, Tim Tutenel, Klaas Jan de Kraker, and Rafael Bidarra. 2010. Interactive creation of virtual worlds using procedural sketching..
In Eurographics (Short papers). 29–32.
[143]
Gillian Smith, Jim Whitehead, and Michael Mateas. 2010. Tanagra: A mixed-initiative level design tool. In Proceedings of the Fifth International
Conference on the Foundations of Digital Games. 209–216.
[144] Michael R Smith, Ryan S Hintze, and Dan Ventura. 2014. Nehovah: A Neologism Creator Nomen Ipsum.. In ICCC. 173–181.
[145] Frank K Sonnenberg. 1991. Strategies for creativity. Journal of Business Strategy (1991).
[146]
Jannick Kirk Sørensen. 2016. Silent game as Model for Examining Student Online Creativity-Preliminary Results from an Experiment. Think
CROSS. Magdeburg: Change MEDIA 10 (2016).
[147]
Jannick Kirk Sørensen. 2017. Exploring Constrained Creative Communication: The Silent Game as Model for Studying Online Collaboration.
International Journal of E-Services and Mobile Applications (IJESMA) 9, 4 (2017), 1–23.
[148] Paul A Szerlip, Amy K Hoover, and Kenneth O Stanley. 2012. Maestrogenesis: Computer-assisted musical accompaniment generation. (2012).
[149]
Kurt Thywissen. 1999. GeNotator: an environment for exploring the application of evolutionary techniques in computer-assisted composition.
Organised Sound 4, 2 (1999), 127–133.
[150]
Ha Trinh, Darren Edge, Lazlo Ring, and Timothy Bickmore. 2016. Thinking Outside the Box: Co-planning Scientic Presentations with Virtual
Agents. In International Conference on Intelligent Virtual Agents. Springer, 306–316.
[151]
Dakuo Wang, Justin D Weisz, Michael Muller, Parikshit Ram, Werner Geyer, Casey Dugan, Yla Tausczik, Horst Samulowitz, and Alexander Gray.
2019. Human-ai collaboration in data science: Exploring data scientists’ perceptions of automated ai. Proceedings of the ACM on Human-Computer
Interaction 3, CSCW (2019), 1–24.
[152] Peter Wegner. 1997. Why interaction is more powerful than algorithms. Commun. ACM 40, 5 (1997), 80–91.
[153]
Miaomiao Wen, Nancy Baym, Omer Tamuz, Jaime Teevan, Susan T Dumais, and Adam Kalai. 2015. OMG UR Funny! Computer-Aided Humor with
an Application to Chat.. In ICCC. 86–93.
[154] Tom White and Ian Loh. 2017. Generating Animations by Sketching in Conceptual Space.. In ICCC. 261–268.
[155] Alena Widows and Harriet Sandilands. 2009. The Poetry Machine. www.thepoetrymachine.net
Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems
27
[156]
Geraint A Wiggins. 2006. A preliminary framework for description, analysis and comparison of creative systems. Knowledge-Based Systems 19, 7
(2006), 449–458.
[157]
Blake Williford, Abhay Doke, Michel Pahud, Ken Hinckley, and Tracy Hammond. 2019. DrawMyPhoto: assisting novices in drawing from
photographs. In Proceedings of the 2019 on Creativity and Cognition. 198–209.
[158]
Lauren Winston and Brian Magerko. 2017. Turn-taking with improvisational co-creative agents. In Thirteenth Articial Intelligence and Interactive
Digital Entertainment Conference.
[159]
Jun Xiao, Xuemei Zhang, Phil Cheatle, Yuli Gao, and C Brian Atkins. 2008. Mixed-initiative photo collage authoring. In Proceedings of the 16th
ACM international conference on Multimedia. 509–518.
[160] Qian Yang, Aaron Steinfeld, Carolyn Rosé, and John Zimmerman. 2020. Re-examining whether, why, and how human-AI interaction is uniquely
dicult to design. In Proceedings of the 2020 chi conference on human factors in computing systems. 1–13.
[161] Georgios N Yannakakis, Antonios Liapis, and Constantine Alexopoulos. 2014. Mixed-initiative co-creativity. (2014).
[162] Matthew Yee-King and Mark d’Inverno. 2016. Experience driven design of creative systems. (2016).
[163] Wako Yoshida, Ray J Dolan, and Karl J Friston. 2008. Game theory of mind. PLoS computational biology 4, 12 (2008), e1000254.
[164]
Michael W Young and Oliver Bown. 2010. Clap-along: A negotiation strategy for creative musical interaction with computational systems. In
Proceedings of the International Conference on Computational Creativity 2010. Departament of Informatics Engineering University of Coimbra,
215–222.
[165]
Chao Zhang, Cheng Yao, Jianhui Liu, Zili Zhou, Weilin Zhang, Lijuan Liu, Fangtian Ying, Yijun Zhao, and Guanyun Wang. 2021. StoryDrawer: A
Co-Creative Agent Supporting Children’s Storytelling through Collaborative Drawing. In Extended Abstracts of the 2021 CHI Conference on Human
Factors in Computing Systems. 1–6.
[166] Viktor Zoric and Björn Gambäck. 2018. The Image Artist: Computer Generated Art Based on Musical Input.. In ICCC. 296–303.
