Interaction-Centered Intelligence:

Toward Interaction as the Primary Unit of Analysis in Co-Creative AI Systems

-Nicholas Davis

Abstract

Traditional artificial intelligence has largely conceptualized intelligence as isolated internal computation occurring within bounded agents. Across classical AI, machine learning, and many generative systems, the dominant unit of analysis remains the individual model, algorithm, or autonomous system evaluated primarily through outputs, benchmarks, prediction accuracy, or optimization performance. While these approaches have produced major advances in artificial intelligence, they often under-theorize the role of interaction itself in the emergence of intelligence, creativity, meaning, and adaptive behavior.

This paper proposes interaction as the primary unit of analysis in co-creative AI systems and interaction-centered intelligence more broadly. Drawing from distributed cognition, embodied cognition, enaction, participatory sense-making, computational creativity, and human-computer interaction, the paper traces a historical progression in cognitive and AI frameworks toward increasingly relational and interaction-centered models of intelligence. Building upon prior work involving Creative Sense-Making, quantified co-creation, participatory interaction, and co-creative AI systems such as the Drawing Apprentice and AI Drawing Partner, the paper argues that intelligence emerges dynamically through interaction between agents, environments, and evolving socio-technical systems rather than solely through isolated internal computation.

The paper introduces interaction-centered intelligence as a broader framework for understanding human-AI co-creation, adaptive participation, collaborative emergence, and interaction dynamics in co-creative systems. Rather than evaluating intelligence solely through generated outputs, the framework emphasizes interaction trajectories, coordination dynamics, participatory engagement, adaptive regulation, and evolving collaborative structures unfolding through time. Implications for explainable co-creative AI, hybrid intelligence, enactive AI, and future human-centered AI systems are discussed.

1. Introduction

Artificial intelligence research has historically focused on isolated intelligent agents performing tasks through internal computation. Across symbolic artificial intelligence, expert systems, machine learning, and contemporary deep learning architectures, intelligence has often been conceptualized primarily through: representation, prediction, optimization, planning, classification, and autonomous generation (Russell & Norvig, 2021).

Within these paradigms, cognition is generally treated as something occurring inside bounded computational systems, while interaction functions largely as a secondary mechanism for transferring information between otherwise separate entities. Human users become external operators, prompts become inputs, and collaboration becomes reduced to exchanges between independently operating systems. Intelligence is therefore often evaluated primarily through outputs such as: benchmark performance, prediction accuracy, generative capability, or task completion (LeCun et al., 2015).

This output-centered framing has produced major advances in artificial intelligence research. However, many forms of intelligence cannot be adequately explained solely through isolated internal computation. Creativity, improvisation, collaborative adaptation, meaning construction, participatory engagement, and human-AI co-creation emerge dynamically through interaction itself rather than solely inside individual agents. In these contexts, intelligence unfolds relationally across evolving systems involving: humans, AI systems, environments, interfaces, materials, representations, and temporal interaction dynamics.

A growing body of work across cognitive science, human-computer interaction, and computational creativity increasingly challenges purely isolated models of cognition. Distributed cognition expanded the unit of analysis beyond individual minds by proposing that cognition distributes across people, artifacts, tools, and socio-technical systems (Hutchins, 1995). Rather than locating cognition exclusively inside individuals, distributed cognition conceptualizes intelligent activity as emerging across coordinated systems involving humans, representations, environments, and technologies. This represented a major shift away from purely internalist models of cognition and strongly influenced later work in creativity support systems, collaborative interaction, and human-computer interaction.

Embodied cognition further challenged classical information-processing assumptions by emphasizing the role of bodily interaction and sensorimotor coupling in shaping cognition and meaning-making (Clark, 1997; Varela et al., 1991). Rather than treating cognition as abstract symbolic reasoning detached from physical experience, embodied approaches argued that cognition emerges through active engagement between bodies and environments. Enaction extended these ideas further by proposing that cognition arises through ongoing perception-action coupling and adaptive interaction between agents and the world (Varela et al., 1991). From this perspective, meaning is not internally represented and then applied to a pre-given environment; rather, meaningful worlds are enacted dynamically through interaction itself.

Participatory sense-making extended enaction into social interaction by proposing that meaning can emerge relationally through coordination dynamics between participants rather than solely through isolated cognition (De Jaegher & Di Paolo, 2007). Interaction becomes partially autonomous and constitutive of cognition itself rather than merely a communication channel between pre-existing minds. These frameworks increasingly suggest that intelligence cannot be fully understood through isolated computation alone because cognition emerges through ongoing adaptive participation within dynamically coupled systems.

At the same time, recent advances in generative AI and co-creative systems have intensified the need for interaction-centered approaches to intelligence. Modern generative systems are often evaluated primarily through the quality of generated outputs while under-theorizing the collaborative interaction processes through which humans and AI systems jointly construct meaning, creativity, and adaptive behavior. Recent work in human-AI co-creation increasingly argues that collaborative AI systems should not be understood merely as autonomous generators, but as participatory systems engaged in: improvisation, turn-taking, coordination, adaptation, and shared creative interaction with humans (Davis et al., 2015; Yannakakis et al., 2014).

This paper argues that interaction itself should become the primary unit of analysis in co-creative AI systems and interaction-centered intelligence more broadly. Rather than treating intelligence as something contained entirely within humans or machines, interaction-centered approaches conceptualize intelligence as emerging through ongoing participation, coordination, adaptation, and collaborative sense-making between dynamically coupled systems. From this perspective, creativity and intelligence are not reducible to isolated internal processes or final outputs alone, but instead emerge through evolving interaction trajectories unfolding through time.

Drawing from distributed cognition, embodied cognition, enaction, participatory sense-making, computational creativity, and human-computer interaction, this paper traces the historical evolution of increasingly interaction-centered paradigms of cognition and AI. The paper further synthesizes prior work involving: Creative Sense-Making, quantified co-creation, activity traces, creative trajectories, participatory interaction, and co-creative AI systems into a broader framework for interaction-centered intelligence. Building upon earlier work involving co-creative systems such as the Drawing Apprentice (Davis et al., 2015), Creative Sense-Making (Davis et al., 2017), quantified co-creation, and explainable co-creative AI systems, the paper proposes that interaction itself may serve as the generative substrate through which intelligence, creativity, and meaning emerge in collaborative human-AI systems.

2. Limitations of Output-Centered AI

Traditional artificial intelligence systems are often evaluated primarily through output-oriented metrics such as benchmark performance, classification accuracy, optimization efficiency, generative capability, or task completion (Russell & Norvig, 2021). Within these paradigms, intelligence becomes inferred from outputs rather than understood through the evolving interaction processes that produce them. Whether in symbolic AI, machine learning, or modern generative systems, the dominant assumption frequently remains that intelligence resides primarily inside bounded computational agents performing internal computation on external inputs (LeCun et al., 2015).

This output-centered framing has enabled substantial advances across artificial intelligence research, particularly in areas involving prediction, optimization, pattern recognition, and autonomous generation. Contemporary large language models and generative systems further extend this paradigm by producing increasingly sophisticated text, images, music, and multimedia artifacts through large-scale statistical modeling and generative inference. However, several limitations emerge when considering co-creative systems, human-AI collaboration, and interaction-centered forms of intelligence.

First, output-centered paradigms frequently obscure the interaction dynamics underlying collaboration itself. In co-creative systems, creativity does not emerge solely from isolated internal generation processes, but through evolving interaction between participants involving reciprocal influence, adaptive coordination, improvisation, turn-taking, collaborative negotiation, and dynamically unfolding creative trajectories (Davis et al., 2015; Yannakakis et al., 2014). Final artifacts alone cannot fully capture: how collaboration evolved, how meaning emerged, how participants adapted to one another, or how interaction shaped creative outcomes through time.

As a result, many important dimensions of co-creative interaction remain invisible when evaluation focuses exclusively on outputs rather than interaction dynamics. This limitation becomes especially significant in collaborative creativity systems where: timing, responsiveness, participatory engagement, and adaptive interaction play central roles in the creative process itself (Lubart, 2005).

Second, output-centered paradigms often treat human participation as external to intelligence rather than constitutive of it. Human users become reduced to prompt providers, evaluators, selectors, or consumers of machine-generated outputs rather than active participants within evolving collaborative systems. Interaction becomes framed as a mechanism for controlling or querying intelligent systems rather than as the primary site through which intelligence and meaning emerge relationally. This framing risks under-theorizing the participatory and adaptive nature of human-AI collaboration by positioning AI systems as fundamentally autonomous generators operating independently of human interaction dynamics (Deterding et al., 2017).

This limitation becomes increasingly visible in contemporary generative AI systems where human contribution may become obscured within opaque generation pipelines. As generative systems become more capable, questions surrounding: authorship, explainability, participation, collaboration, and hybrid intelligence

become increasingly difficult to address through output-oriented frameworks alone. Several recent approaches in explainable AI similarly struggle because they focus primarily on exposing internal model features rather than modeling the evolving interaction trajectories between humans and AI systems during collaborative activity (Miller, 2019).

Third, output-centered approaches struggle to account for temporal interaction dynamics unfolding through time. Creativity, collaboration, and participatory meaning-making are not static events but temporally evolving processes involving: adaptation, coordination, breakdown, recovery, conceptual divergence, and changing collaborative relationships.

Traditional evaluation paradigms frequently compress these dynamics into static endpoint assessments or isolated artifact evaluations, thereby obscuring the evolving structures of interaction itself (Davis et al., 2017). This becomes particularly problematic in co-creative systems, hybrid intelligence environments, collaborative design systems, and interaction-centered AI contexts where intelligence emerges through sustained participation and adaptive interaction over time.

More broadly, output-centered paradigms often struggle to account for collaborative emergence — situations in which novel meaning, creativity, or adaptive behavior emerge relationally through interaction rather than being attributable solely to individual participants. Participatory sense-making frameworks argue that interaction itself can become partially autonomous and constitutive of cognition, meaning that collaborative dynamics cannot always be reduced to isolated internal processes within separate agents (De Jaegher & Di Paolo, 2007). In co-creative systems, meaning and creativity frequently emerge within the interaction itself rather than solely inside either the human or AI participant independently.

These limitations suggest the need for a broader interaction-centered perspective on intelligence. Rather than evaluating intelligence exclusively through isolated outputs, interaction-centered approaches propose that: interaction trajectories, coordination dynamics, participatory engagement, adaptive regulation, and collaborative emergence should become central phenomena for analysis within co-creative AI systems and human-AI collaboration research more broadly.

3. Historical Evolution of the Unit of Analysis

The history of cognitive science and artificial intelligence can be partially understood as a gradual expansion of the unit of analysis used to explain cognition and intelligence. Across the twentieth and early twenty-first centuries, multiple theoretical movements progressively shifted explanations of intelligence away from isolated internal computation toward increasingly relational, embodied, distributed, and interaction-centered frameworks. This progression reflects a broader conceptual transition from understanding cognition as something occurring solely inside bounded minds or computational systems toward understanding intelligence as emerging dynamically through interaction between agents, environments, bodies, artifacts, and social systems.

Early information-processing approaches in cognitive science and symbolic artificial intelligence primarily conceptualized cognition as internal symbolic computation occurring within isolated agents (Newell & Simon, 1976). Classical cognitive science frequently modeled the mind as an information-processing system analogous to a computer, where cognition involved the manipulation of symbolic representations according to formal computational rules. Similarly, early artificial intelligence systems focused heavily on: logical reasoning, symbolic planning, rule-based inference, search, and internal representation.

Within these paradigms, intelligence was largely treated as an internal property of bounded systems, while interaction with environments functioned primarily as input and output channels for otherwise self-contained computational processes (Russell & Norvig, 2021).

While these approaches enabled major advances in formal reasoning and computational modeling, they increasingly struggled to explain: situated cognition, adaptive interaction, embodied behavior, collaborative intelligence, and real-world human activity.

In response, distributed cognition expanded the unit of analysis beyond isolated minds by proposing that cognition distributes across socio-technical systems involving people, tools, artifacts, representations, and environments (Hutchins, 1995). Rather than viewing cognition solely as internal mental processing, distributed cognition conceptualized intelligent activity as emerging across coordinated systems of humans and material structures. Edwin Hutchins’s influential studies of navigation teams demonstrated how cognition unfolds across interactions between individuals, external representations, instruments, and environmental structures rather than residing entirely inside individual minds.

Distributed cognition significantly influenced: human-computer interaction, creativity support systems, collaborative work research, and socio-technical systems theory.

Importantly, it shifted attention toward interaction between humans and technologies as a central phenomenon for understanding cognition. However, distributed cognition often emphasized the structural distribution of cognition across systems more heavily than the evolving temporal dynamics of interaction itself.

Embodied cognition further expanded the unit of analysis by emphasizing the role of bodily interaction and sensorimotor coupling in shaping cognition and meaning-making (Clark, 1997; Lakoff & Johnson, 1999). Rather than treating cognition as abstract symbolic manipulation detached from physical experience, embodied approaches argued that cognition emerges through active bodily engagement with environments. Perception, action, movement, and physical constraints became central to understanding intelligent behavior.

Embodied cognition challenged purely representational views of intelligence by proposing that cognition depends fundamentally upon: bodily structure, sensorimotor capacities, environmental interaction, and situated activity.

This perspective strongly influenced: robotics, ecological psychology, interaction design, and theories of situated cognition.

However, many embodied approaches remained focused primarily on the relationship between individual bodies and environments rather than collaborative interaction dynamics between multiple participants.

Enaction extended these ideas further by proposing that cognition emerges through dynamic perception-action coupling and adaptive engagement between agents and environments (Varela et al., 1991). Rather than representing a pre-given external world internally, enactive cognition argued that organisms enact meaningful worlds through ongoing interaction with their environments. Cognition became conceptualized not as detached representation, but as: adaptive sense-making, embodied interaction, and relational engagement with the world.

Within enactive frameworks, intelligence emerges through continuous interaction between agents and environments rather than through isolated internal symbolic processing. Enaction further emphasized: autonomy, emergence, embodiment, adaptivity, and participatory engagement as foundational aspects of cognition.

Participatory sense-making later extended enaction into social interaction by proposing that meaning can emerge relationally through coordination dynamics between participants themselves (De Jaegher & Di Paolo, 2007). Interaction became partially autonomous and constitutive of cognition rather than merely serving as a communication channel between pre-existing minds. Meaning was no longer treated solely as something internally generated and externally transmitted; instead, collaborative interaction itself became a site of cognitive emergence.

This shift proved particularly influential for: social cognition, human-computer interaction, co-creative systems, and collaborative AI research.

Participatory sense-making increasingly suggested that intelligence may emerge through evolving coordination dynamics between participants rather than solely within isolated individuals.

More recent work in co-creative AI and Creative Sense-Making further extended these interaction-centered approaches into human-AI collaboration and quantified interaction dynamics. Rather than evaluating intelligence exclusively through generated outputs or isolated system performance, Creative Sense-Making proposed that co-creative interaction itself should become a primary phenomenon for analysis (Davis et al., 2017). This work introduced frameworks for modeling: activity traces, creative trajectories, interaction dynamics, participatory engagement, and sense-making curves unfolding through time during human-AI collaboration.

Quantified co-creation further operationalized these ideas computationally by proposing methods for analyzing evolving collaborative interaction dynamics rather than focusing solely on final artifacts. Interaction itself became observable, measurable, and computationally modelable within co-creative systems.

Across these theoretical movements, the unit of analysis progressively expanded: from isolated symbolic computation, to distributed socio-technical systems, to embodied interaction, to organism-environment coupling, to participatory social interaction, and finally toward interaction-centered intelligence itself.

This historical progression increasingly suggests that intelligence may not reside solely inside isolated systems, but instead emerge dynamically through interaction between agents, environments, bodies, representations, technologies, and evolving collaborative systems. Co-creative AI systems therefore provide an important domain for studying intelligence not merely as internal computation, but as a relational and participatory phenomenon unfolding through interaction itself.

4. Distributed Cognition

Distributed cognition challenged the assumption that cognition occurs solely within isolated minds by proposing that cognitive processes distribute across coordinated socio-technical systems involving people, tools, artifacts, representations, workflows, and environments (Hutchins, 1995). Rather than conceptualizing cognition as an exclusively internal mental phenomenon, distributed cognition argued that intelligent activity unfolds dynamically across systems of interaction between humans and material structures. This represented a major shift away from classical information-processing approaches in cognitive science, which frequently treated cognition as symbolic computation occurring inside bounded agents (Newell & Simon, 1976).

Edwin Hutchins’s influential work on distributed cognition demonstrated how cognition emerges collectively across navigation teams, instruments, representational systems, and collaborative workflows during real-world activity (Hutchins, 1995). In these settings, problem solving and reasoning are not reducible to isolated mental operations performed by individual participants. Instead, cognition becomes distributed across: social coordination, external representations, technological systems, environmental structures, and collaborative interaction.

Within distributed cognition, tools and artifacts are not merely passive supports for cognition; they actively participate in cognitive processes themselves. Maps, diagrams, interfaces, instruments, workflows, and representational systems become integral components of intelligent activity. Cognitive processes therefore extend beyond individual minds into larger socio-technical systems involving humans and technologies operating together (Clark & Chalmers, 1998).

Distributed cognition strongly influenced several research domains including: creativity support systems, collaborative systems research, computer-supported cooperative work, human-computer interaction, and socio-technical systems theory.

In creativity support systems research, distributed cognition helped shift attention toward how creativity emerges through interaction between people, tools, environments, and evolving constraints rather than solely through isolated individual inspiration (Shneiderman, 2007). Similarly, within human-computer interaction, distributed cognition contributed to understanding interfaces not merely as communication channels, but as active participants within larger cognitive systems.

Importantly, distributed cognition represented one of the first major theoretical movements to significantly expand the unit of analysis beyond isolated minds toward coordinated systems of interaction. Intelligence became understood as distributed across systems rather than located exclusively inside individual agents.

However, while distributed cognition expanded the spatial distribution of cognition, it often focused more heavily on structural organization and representational distribution than on evolving temporal interaction dynamics themselves. Many distributed cognition approaches emphasized: coordination structures, representational flow, task decomposition, and socio-technical organization without fully modeling how interaction unfolds dynamically through time during collaborative participation.

This limitation becomes especially important in co-creative systems and human-AI collaboration where: improvisation, reciprocal adaptation, interaction timing, participatory engagement, and evolving coordination dynamics play central roles in the emergence of creativity and meaning. While distributed cognition successfully expanded cognition beyond isolated minds, later frameworks such as enaction, participatory sense-making, and Creative Sense-Making would further extend the unit of analysis toward interaction dynamics themselves rather than solely distributed cognitive structure.

5. Embodied Cognition

Embodied cognition further challenged classical computational theories of mind by emphasizing the role of the body in shaping cognition through perception-action coupling and sensorimotor interaction (Clark, 1997; Varela et al., 1991). Rather than treating cognition as abstract symbolic manipulation detached from physical experience, embodied approaches argued that bodily structure, movement, perception, and environmental interaction fundamentally shape cognition and meaning-making.

Classical cognitive science often conceptualized cognition as computational information processing independent of bodily form. In contrast, embodied cognition proposed that intelligent behavior emerges through active bodily engagement with environments. Cognition therefore cannot be fully separated from: sensorimotor capacities, physical movement, perceptual systems, bodily constraints, and situated environmental interaction (Lakoff & Johnson, 1999).

Embodied cognition strongly influenced: robotics, ecological psychology, interaction design, human-computer interaction, and adaptive systems research.

In robotics, embodied approaches demonstrated that intelligent behavior can emerge through direct sensorimotor interaction with environments rather than relying exclusively on complex internal symbolic representations (Brooks, 1991). Rodney Brooks famously argued against purely representational AI architectures by emphasizing situated embodied interaction as foundational to adaptive intelligence.

Similarly, embodied cognition contributed significantly to ecological psychology and affordance theory, particularly through the work of James Gibson (1979), who proposed that perception is fundamentally relational and action-oriented. From this perspective, organisms perceive environments not merely as objective external spaces, but in terms of affordances — opportunities for action emerging through organism-environment relationships.

Embodied cognition also had substantial influence within human-computer interaction by emphasizing how interfaces, technologies, and environments shape cognition through bodily interaction and situated engagement. Interaction design increasingly shifted toward understanding cognition as emerging through embodied activity rather than abstract information processing alone.

Importantly, embodied cognition expanded the unit of analysis beyond internal symbolic reasoning toward active bodily engagement with environments. Cognition became understood as fundamentally grounded in perception-action coupling rather than detached computation.

However, many embodied cognition approaches remained focused primarily on individual embodied agents interacting with environments rather than collaborative interaction dynamics between multiple participants. While embodiment emphasized the relational coupling between organisms and environments, less attention was often given to: social coordination, collaborative emergence, participatory interaction, and interaction dynamics unfolding between multiple agents.

These limitations became increasingly important in research involving: social cognition, collaborative creativity, participatory interaction, and human-AI co-creation.

Subsequent frameworks such as enaction and participatory sense-making would extend embodiment further by emphasizing how meaning and cognition emerge not only through embodied interaction with environments, but also through evolving coordination dynamics between interacting participants.

6. Enaction and Participatory Sense-Making

Enactive cognition further extended interaction-centered approaches to cognition by proposing that cognition emerges through dynamic interaction between agents and environments rather than through internal representations of a pre-given world (Varela et al., 1991). Within enactive frameworks, cognition became understood as: adaptive engagement, sense-making, embodied interaction, and ongoing perception-action coupling.

Rather than conceptualizing cognition as the manipulation of internal symbolic representations, enaction proposed that organisms actively enact meaningful worlds through adaptive interaction with their environments. Meaning is therefore not internally stored and later applied to the world; rather, meaning emerges dynamically through participation and embodied engagement itself.

Enactive cognition emphasized several key concepts including: autonomy,  emergence, embodiment, adaptivity, and relational sense-making.

Autonomous systems actively regulate their relationships with environments in ways that sustain their viability and continued interaction (Di Paolo, 2005). Cognition therefore becomes inseparable from adaptive interaction between agents and environments.

Importantly, enaction shifted the unit of analysis away from isolated internal representations toward ongoing relational interaction itself. Intelligence became conceptualized as something emerging dynamically through interaction rather than residing statically inside bounded systems.

Participatory sense-making later extended enaction into social interaction by arguing that meaning can emerge relationally through coordination dynamics between participants themselves (De Jaegher & Di Paolo, 2007). Interaction became partially autonomous and constitutive of cognition rather than merely serving as a medium for information exchange between pre-existing minds.

Within participatory sense-making, social interaction is not reducible to isolated cognition occurring independently inside separate individuals. Instead, interaction dynamics themselves contribute to the emergence of: meaning, coordination, collaborative adaptation, and shared understanding.

This represented a major conceptual shift because interaction itself became understood as a site of cognitive emergence rather than merely a transmission mechanism between isolated agents.

These frameworks became highly influential for: co-creative AI, human-computer interaction, computational creativity, collaborative systems research, and hybrid intelligence systems because they positioned interaction as central to: meaning construction, collaboration, creativity, participatory engagement, and adaptive intelligence.

Within co-creative AI research, enaction and participatory sense-making provided theoretical foundations for understanding creativity as something emerging through interaction between humans and AI systems rather than solely through isolated computational generation (Davis et al., 2015). Co-creative systems such as the Drawing Apprentice increasingly explored: improvisational collaboration, reciprocal interaction, turn-taking, adaptive coordination, and participatory creativity between humans and AI agents.

However, while enactive cognition and participatory sense-making provided powerful theoretical frameworks for interaction-centered cognition, many enactive approaches remained difficult to operationalize computationally. Concepts such as: sense-making, participatory emergence, adaptive coordination, and relational intelligence were often theoretically rich but challenging to formally model, quantify, or computationally implement within AI systems.

This limitation helped motivate later work involving: Creative Sense-Making, quantified co-creation, activity traces, creative trajectories, and interaction dynamics modeling, which attempted to operationalize interaction-centered cognition computationally within co-creative AI systems.

7. Creative Sense-Making

Creative Sense-Making (CSM) was introduced as a cognitive framework for modeling and quantifying interaction dynamics during co-creation (Davis et al., 2017). Emerging from earlier work in enactive cognition, participatory sense-making, computational creativity, and co-creative AI, the framework proposed that creativity should not be understood solely through isolated internal cognition or final creative artifacts. Instead, CSM argued that creativity emerges dynamically through evolving interaction between participants during collaborative activity itself (Davis et al., 2017; Davis, 2017).

Prior work in computational creativity had often focused primarily on evaluating generated outputs or autonomous creative capability. However, co-creative systems introduced fundamentally different interaction dynamics because creativity unfolded relationally between human and AI participants through: improvisation, reciprocal influence, adaptation, turn-taking, coordination, and collaborative meaning construction (Yannakakis et al., 2014).

CSM emerged in response to the need for a cognitive framework capable of modeling these evolving interaction processes rather than focusing solely on completed artifacts.

Drawing heavily from enactive cognition and participatory sense-making, Creative Sense-Making proposed that co-creative interaction involves continuous processes of collaborative sense-making unfolding through time (Davis et al., 2017). Rather than conceptualizing creativity as the isolated production of novel outputs, the framework treated co-creation as an evolving interaction trajectory shaped by adaptive participation between collaborators. Creativity therefore became understood as: temporally unfolding, relational, participatory, and interaction-centered.

A central contribution of CSM involved introducing methods for modeling interaction dynamics computationally through: activity traces, interaction histories, creative trajectories, interaction trends, and sense-making curves unfolding dynamically through time (Davis et al., 2017; Davis, 2017). These constructs allowed researchers to visualize and analyze how collaboration evolves moment-by-moment during open-ended co-creative interaction.

Activity traces captured observable interaction events occurring during collaboration, including contributions, responses, interruptions, adaptations, and collaborative exchanges between participants. Interaction histories recorded the evolving sequence of interaction states over time, allowing researchers to examine how co-creative processes developed dynamically during participation. Creative trajectories modeled larger collaborative patterns emerging across interaction sequences, while sense-making curves provided visual representations of evolving collaboration rhythms and interaction dynamics during co-creation.

Importantly, CSM proposed that these interaction dynamics could reveal important cognitive and collaborative structures that remain invisible when evaluating final outputs alone. Creativity became understood not merely as a property of isolated artifacts, but as an emergent process unfolding through interaction itself.

The framework was developed and empirically grounded through studies involving: pretend play, collaborative improvisation, and co-creative drawing systems (Davis, 2017).

One of the most influential applications involved the Drawing Apprentice, an enactive co-creative drawing agent developed at Georgia Institute of Technology (Davis et al., 2015). Within this system, humans and AI agents collaborated on a shared drawing canvas through reciprocal interaction and improvisational sketch exchange. CSM provided methods for modeling how: collaboration rhythms, interaction timing, conceptual divergence, participatory engagement, and adaptive coordination evolved dynamically during co-creation.

CSM further introduced computational approaches for visualizing: collaboration rhythms, participatory interaction, conceptual shifts, adaptive coordination, and evolving co-creative states during interaction itself (Davis et al., 2017). These visualizations allowed researchers to analyze not only what creative artifacts were produced, but how collaboration unfolded through time between participants.

This represented a major shift within computational creativity research because it reframed creativity as an interaction-centered phenomenon rather than an isolated generative capability. The framework increasingly positioned interaction itself as a central object of analysis in co-creative systems.

Creative Sense-Making also contributed toward bridging: cognitive science, computational creativity, human-computer interaction, and co-creative AI. By synthesizing enactive cognition with computational interaction modeling, the framework helped operationalize previously difficult-to-model concepts such as: participatory sense-making, collaborative emergence, adaptive coordination, and interaction-centered creativity.

More recent work has continued extending the framework into broader models of co-creative AI, quantified collaboration, and interaction-centered intelligence. Davis (2024) further generalized CSM into a broader framework for modeling interaction dynamics in co-creative AI systems, while newer systems such as the AI Drawing Partner operationalized quantified interaction modeling directly inside co-creative systems themselves (Davis & Rafner, 2025).

8. Quantified Co-Creation

Quantified co-creation extended Creative Sense-Making into practical computational methods for modeling collaborative dynamics in co-creative AI systems. Rather than evaluating creativity solely through isolated outputs or completed artifacts, quantified co-creation focused on capturing, visualizing, and computationally modeling interaction itself as it unfolded dynamically through time (Davis et al., 2017).

Within traditional computational creativity research, evaluation frequently focused on: novelty, value, artifact quality, or autonomous generative capability. However, these approaches often struggled to capture the evolving collaborative dynamics underlying co-creative interaction. Quantified co-creation emerged in response to this limitation by proposing that: interaction trajectories, participatory engagement, coordination dynamics, adaptation, and collaborative emergence should themselves become observable and measurable computational phenomena.

This framework positioned interaction itself as: observable, measurable, quantifiable, and computationally modelable. Rather than treating co-creative systems as static generators, quantified co-creation analyzed: turn-taking behavior, timing, response patterns, conceptual shifts, interaction trajectories, collaborative rhythms, adaptive coordination, and evolving co-creative structures emerging during participation between humans and AI systems.

One of the major contributions of quantified co-creation involved introducing computational approaches for modeling collaboration dynamically through time. These methods enabled researchers to analyze how co-creative interaction evolves rather than merely evaluating final outcomes after collaboration had concluded.

For example, interaction trajectories captured evolving patterns of collaboration between participants across time, while turn-taking analysis examined how initiative and participation shifted dynamically between humans and AI systems during co-creation. Timing analysis further examined: pauses, interruptions, rhythmic interaction patterns, and coordination dynamics during improvisational collaboration.

Conceptual shifts became another major focus within quantified co-creation research. Building upon work in conceptual blending and creative divergence, researchers explored how AI systems could intentionally provoke: novelty, reinterpretation, ambiguity, and conceptual divergence during collaborative interaction (Karimi et al., 2018a; Karimi et al., 2019). Deep learning methods were integrated into co-creative systems to identify structurally similar but conceptually distinct sketch relationships capable of encouraging more creative interaction trajectories during collaborative design.

This work significantly expanded co-creative AI research beyond static generation toward adaptive collaborative interaction. AI systems increasingly became understood as: improvisational partners, interaction participants, and collaborative contributors rather than isolated autonomous creators.

Quantified co-creation also strongly influenced emerging work in: explainable co-creative AI, hybrid intelligence, interaction-centered evaluation, and participatory AI systems.

Because interaction itself became computationally observable, researchers could increasingly visualize and explain: how collaboration evolved, how meaning emerged, how conceptual divergence occurred, and how adaptive interaction shaped creative outcomes.

This represented an important departure from black-box generative paradigms where collaborative processes remain largely invisible inside opaque generation pipelines. The AI Drawing Partner further extended quantified co-creation into modern co-creative AI systems by integrating Creative Sense-Making directly into the interaction architecture itself (Davis & Rafner, 2025). The system automatically captured and visualized quantified interaction data during collaborative drawing sessions, including: interaction curves, participation dynamics, trend sequences, and evolving co-creative trajectories.

This transformed quantified co-creation from a post-hoc analytical framework into an active computational layer embedded directly within co-creative systems themselves.

More broadly, quantified co-creation helped establish interaction-centered intelligence as a computationally tractable research direction. By operationalizing interaction dynamics through measurable structures and visual models, the framework provided one of the earliest systematic approaches for studying: participatory interaction, collaborative emergence, adaptive co-creation, and interaction-centered cognition within human-AI systems.

9. Interaction-Centered Intelligence

Interaction-centered intelligence proposes that intelligence emerges through interaction itself rather than solely through isolated internal computation. Rather than conceptualizing cognition as a property contained entirely within bounded agents, this framework argues that intelligence unfolds dynamically across evolving systems of participation involving: humans, AI systems, environments, interfaces, artifacts, social systems, and temporal interaction dynamics.

Within this perspective, intelligence is not reduced to internal symbolic manipulation, prediction, or autonomous generation alone. Instead, cognition emerges relationally through ongoing interaction, adaptive coordination, participatory engagement, and collaborative sense-making between dynamically coupled systems (De Jaegher & Di Paolo, 2007; Varela et al., 1991).

Interaction-centered intelligence synthesizes several historical theoretical movements discussed throughout this paper, including: distributed cognition, embodied cognition, enaction, participatory sense-making, co-creative AI, and Creative Sense-Making.

Distributed cognition expanded cognition across socio-technical systems involving humans and artifacts (Hutchins, 1995), while embodied cognition emphasized the role of bodily interaction and perception-action coupling in shaping cognition (Clark, 1997). Enaction further reframed cognition as adaptive sense-making emerging through organism-environment interaction (Varela et al., 1991), and participatory sense-making extended these ideas into collaborative social interaction (De Jaegher & Di Paolo, 2007). Interaction-centered intelligence builds upon these frameworks by proposing that interaction itself becomes the primary phenomenon through which intelligence emerges and should therefore become the primary unit of analysis for co-creative AI and hybrid intelligence systems.

This perspective is particularly important in co-creative AI because collaborative intelligence cannot be fully explained through isolated outputs alone. In co-creative systems, creativity emerges through: reciprocal influence, adaptive coordination, improvisation, turn-taking, collaborative negotiation, conceptual divergence, and evolving interaction trajectories (Davis et al., 2017).

These interaction processes frequently generate forms of meaning and creativity that cannot be attributed solely to either the human or AI participant independently. Instead, intelligence emerges relationally through interaction itself. Interaction-centered intelligence therefore emphasizes: participatory engagement, adaptive regulation, collaborative emergence, coordination dynamics, interaction rhythms, and sustained interaction trajectories as central explanatory phenomena. This perspective reframes co-creative AI systems not merely as creative applications or productivity tools, but as experimental platforms for studying intelligence as an emergent relational phenomenon.

Recent work in co-creative AI increasingly supports this interaction-centered perspective. Frameworks such as Creative Sense-Making (Davis et al., 2017), quantified co-creation (Davis, 2017), the Co-Creative Design Framework (Davis et al., 2025), and COFI (Rezwana & Maher, 2022) all emphasize interaction dynamics as central to co-creative collaboration. These approaches increasingly shift attention away from isolated AI generation toward: interaction structure, participatory adaptation, communication, coordination, and collaborative emergence.

For example, the AI Drawing Partner operationalizes interaction-centered intelligence directly within a co-creative system by modeling interaction dynamics during collaborative drawing sessions through quantified interaction curves, activity traces, and evolving collaboration trajectories (Davis & Rafner, 2025). Rather than functioning solely as an autonomous generator, the system participates dynamically in co-creation while simultaneously modeling how interaction evolves through time.

Interaction-centered intelligence also aligns closely with emerging work in hybrid intelligence and human-centered AI. Hybrid intelligence frameworks increasingly emphasize that future intelligent systems may depend upon effective collaboration between humans and AI systems rather than autonomous machine intelligence alone (Dellermann et al., 2019). Similarly, human-centered AI increasingly focuses on augmenting human capabilities and supporting collaborative interaction rather than replacing human participation entirely (Shneiderman, 2022). Interaction-centered intelligence extends these perspectives further by proposing that interaction itself constitutes the generative substrate through which intelligence unfolds.

Importantly, this perspective also reframes intelligence as fundamentally temporal and process-oriented. Intelligence becomes understood not as static capability, but as an evolving interaction trajectory involving: adaptation, coordination, participation, regulation, and collaborative emergence unfolding through time.

From this perspective, intelligence cannot be fully reduced to isolated outputs or internal model states because meaning and adaptive behavior emerge dynamically through interaction itself.

10. Implications for Human-AI Co-Creation

Interaction-centered intelligence has several important implications for future human-AI systems, particularly within co-creative AI, hybrid intelligence, human-centered AI, and collaborative interaction design.

First, evaluation frameworks for AI systems should increasingly consider: interaction dynamics, participatory engagement, adaptive coordination, collaboration rhythms, interaction trajectories, and sustained participatory interaction rather than evaluating intelligence solely through isolated outputs or benchmark performance. Traditional AI evaluation paradigms frequently assess systems through: accuracy, optimization metrics, output quality, or autonomous capability.

However, these approaches often fail to capture the collaborative interaction dynamics through which intelligence emerges in co-creative systems (Davis et al., 2017; Guzdial & Riedl, 2019). In co-creative AI, important dimensions of intelligence involve: responsiveness, improvisation, coordination, communication, and adaptive participation during interaction itself.

This shift suggests that future evaluation frameworks may increasingly require longitudinal and interaction-centered methods capable of modeling how collaboration evolves dynamically through time.

Second, interaction-centered intelligence has major implications for explainable AI systems. Traditional explainable AI approaches often focus on exposing: model weights, feature importance, internal representations, or decision pathways.

While these approaches may improve interpretability at the level of model computation, they often under-theorize the collaborative interaction processes through which humans and AI systems jointly construct meaning during co-creation (Miller, 2019). Interaction-centered approaches instead suggest that explainability should increasingly involve revealing: evolving interaction dynamics, participatory trajectories, collaborative adaptation, conceptual divergence, and co-regulation processes underlying human-AI collaboration itself.

Recent work in quantified co-creation and explainable co-creative AI increasingly moves in this direction by visualizing interaction curves, activity traces, participation dynamics, and creative trajectories during collaboration (Davis et al., 2025). This reframes explainability not merely as exposing internal model states, but as making collaborative interaction itself visible and interpretable.

Third, interaction-centered intelligence suggests that future hybrid intelligence systems may increasingly depend upon: sustained participatory interaction, adaptive collaboration, co-regulation, and relational coordination between humans and AI systems (Dellermann et al., 2019). Rather than treating AI systems as autonomous replacements for human cognition, hybrid intelligence frameworks increasingly conceptualize intelligence as emerging through collaboration between humans and computational systems. Interaction-centered intelligence extends this idea further by arguing that collaborative interaction itself becomes the central site through which adaptive intelligence emerges.

This perspective may significantly influence: co-creative systems, educational AI, collaborative design systems, adaptive interfaces, creativity support tools, and human-AI teamwork environments.

Rather than optimizing solely for autonomous generation or efficiency, future systems may increasingly prioritize: participatory engagement, collaborative fluency, adaptive responsiveness, and sustained interaction quality.

Fourth, co-creative systems may function as important empirical environments for studying interaction-centered cognition itself. Because co-creative AI systems involve ongoing collaboration between humans and AI agents, they provide unique research environments for investigating: participatory sense-making, adaptive coordination, collaborative emergence, interaction dynamics, and interaction-centered intelligence.

Systems such as the Drawing Apprentice and AI Drawing Partner demonstrate how co-creative AI platforms can simultaneously function as: collaborative creative systems, cognitive research platforms, and interaction modeling environments (Davis et al., 2015; Davis & Rafner, 2025).

This suggests that co-creative systems may play an increasingly important role in operationalizing and studying interaction-centered cognition computationally.

More broadly, interaction-centered intelligence challenges the dominant assumption that intelligence resides primarily inside isolated computational systems. Instead, it proposes that intelligence emerges dynamically through participation between humans, AI systems, environments, and evolving socio-technical structures. As AI systems become increasingly integrated into creative, educational, organizational, and social environments, understanding interaction itself may become essential for developing future human-centered artificial intelligence systems.

11. Future Directions

Several future directions emerge from interaction-centered intelligence research, particularly as AI systems become increasingly integrated into collaborative creative, cognitive, educational, and organizational environments. Collectively, these directions suggest a broader shift away from isolated-output paradigms toward frameworks capable of modeling intelligence as an emergent relational and participatory process unfolding through interaction itself.

One important direction involves the development of interaction-centered evaluation frameworks for co-creative AI systems. Traditional AI evaluation methods frequently emphasize: benchmark performance, output quality, optimization efficiency, or autonomous capability (Russell & Norvig, 2021).

However, these approaches often fail to capture the evolving collaborative dynamics through which intelligence emerges during human-AI interaction. Interaction-centered intelligence instead suggests that evaluation frameworks should increasingly analyze: interaction dynamics, participatory engagement, coordination quality, adaptive responsiveness, collaborative trajectories, and long-term interaction stability (Davis et al., 2017; Guzdial & Riedl, 2019).

This shift may require new computational metrics capable of modeling interaction itself as a temporally evolving phenomenon rather than reducing intelligence to isolated outputs alone.

A second major direction involves explainable co-creative AI systems. Existing explainable AI approaches frequently focus on exposing: feature importance, internal model weights, attention maps, or decision pathways (Miller, 2019).

While these approaches improve interpretability at the level of internal computation, they often under-theorize the collaborative interaction processes through which humans and AI systems jointly construct meaning during co-creation. Interaction-centered approaches instead suggest that explainability should increasingly involve visualizing: interaction trajectories, collaboration rhythms, participatory dynamics, conceptual divergence, and adaptive coordination emerging during interaction itself (Davis & Rafner, 2025). This reframes explainability not merely as exposing internal computation, but as revealing the evolving relational structures through which intelligence emerges collaboratively.

Future work may also significantly expand quantified collaboration modeling. Creative Sense-Making and quantified co-creation introduced early methods for modeling: activity traces, interaction histories, creative trajectories, and sense-making curves (Davis, 2017; Davis et al., 2017).

However, broader computational frameworks are still needed for modeling long-term collaborative interaction across diverse domains beyond creative drawing systems. Future systems may increasingly incorporate: real-time interaction modeling, adaptive interaction analysis, longitudinal collaboration tracking, multi-agent participation modeling, and dynamic co-regulation analysis within collaborative AI environments.

Another important direction involves adaptive regulation systems capable of dynamically regulating interaction quality during collaboration. Current co-creative systems frequently emphasize generation and response production while under-theorizing: interaction pacing, participatory balance, coordination breakdown, over-engagement, or collaborative drift.

Interaction-centered intelligence suggests that future AI systems may increasingly require mechanisms for regulating interaction itself in order to sustain coherent collaboration over time. Such systems may involve: adaptive pacing, participatory balancing, dynamic initiative regulation, collaborative repair, and interactional stabilization during human-AI interaction.

This direction also connects closely with emerging work involving structural drift and interaction degradation in long-term human-AI systems. Sustained interaction trajectories may involve gradual changes in: coordination quality, participation balance, interaction rhythms, collaborative coherence, and meaning construction over time.

Future research may therefore require new models capable of analyzing how collaborative interaction structures evolve, stabilize, or deteriorate during extended human-AI participation.

Hybrid intelligence systems also represent a major future direction for interaction-centered intelligence research. Hybrid intelligence frameworks increasingly argue that future intelligent systems may depend not upon autonomous AI alone, but upon effective collaboration between humans and computational systems (Dellermann et al., 2019). Interaction-centered intelligence extends this perspective by proposing that intelligence itself emerges through participatory interaction between humans and AI systems rather than residing exclusively inside either participant independently.

Rather than optimizing exclusively for automation or autonomous generation, future systems may increasingly prioritize: participatory engagement, adaptive coordination, collaborative fluency, and sustained interaction quality.

Another important future direction involves participatory interaction visualization. Current AI systems often conceal the interaction dynamics underlying collaborative participation inside opaque generation pipelines. Interaction-centered approaches instead suggest that collaborative interaction itself should become visible, interpretable, and computationally explorable. Future systems may therefore increasingly incorporate: interaction visualizations, collaborative trajectory mapping, participation dynamics modeling, interaction rhythm analysis, and co-creative state visualization directly into collaborative AI environments.

Long-term interaction trajectories also represent an important area for future research. Most current AI systems remain optimized for short-term tasks or isolated interactions. However, interaction-centered intelligence suggests that meaningful collaborative intelligence may emerge through sustained interaction unfolding across: extended timescales, repeated participation, evolving coordination structures, and long-term adaptive relationships.

This introduces new research questions involving: longitudinal co-creation, adaptive co-regulation, collaborative learning, trust formation, interaction stability, and evolving participatory dynamics within human-AI systems.

Finally, further work is needed to operationalize interaction-centered cognition computationally across broader domains beyond creative collaboration. Much of the existing work involving Creative Sense-Making, quantified co-creation, and co-creative AI has focused primarily on creative systems such as collaborative drawing and design environments (Davis et al., 2015; Davis et al., 2017). However, interaction-centered intelligence may also apply more broadly to: educational systems, organizational collaboration, scientific discovery, adaptive decision support, human-robot interaction, and hybrid intelligence systems more generally.

As AI systems become increasingly embedded within human social and cognitive environments, understanding interaction itself may become essential for developing future human-centered AI systems capable of sustained participatory collaboration.

12. Conclusion

This paper argued that interaction itself should become the primary unit of analysis in co-creative AI systems and interaction-centered intelligence more broadly. Across much of the history of cognitive science and artificial intelligence, intelligence has frequently been conceptualized as isolated internal computation occurring within bounded agents. From symbolic AI to modern generative systems, dominant paradigms have often emphasized: representation, prediction, optimization, classification, and autonomous generation as central explanatory mechanisms (Russell & Norvig, 2021).

While these paradigms have enabled major advances in artificial intelligence, they often under-theorize the interaction dynamics through which creativity, collaboration, participatory meaning-making, and adaptive intelligence emerge during human-AI interaction.

Drawing from distributed cognition, embodied cognition, enaction, participatory sense-making, Creative Sense-Making, and quantified co-creation, this paper traced a historical progression toward increasingly relational and interaction-centered paradigms of cognition and AI. Distributed cognition expanded cognition across socio-technical systems involving humans and artifacts (Hutchins, 1995), embodied cognition emphasized bodily interaction and perception-action coupling (Clark, 1997), and enaction reframed cognition as adaptive sense-making emerging through organism-environment interaction (Varela et al., 1991). Participatory sense-making further extended these ideas into collaborative interaction by proposing that meaning can emerge relationally through coordination dynamics themselves (De Jaegher & Di Paolo, 2007).

Building upon these foundations, Creative Sense-Making and quantified co-creation introduced computational methods for modeling interaction dynamics during human-AI collaboration through: activity traces, interaction histories, creative trajectories, participation dynamics, and sense-making curves (Davis et al., 2017).

These frameworks increasingly shifted attention away from isolated outputs toward interaction itself as a central phenomenon for understanding co-creative intelligence.

Interaction-centered intelligence extends this progression further by proposing that intelligence emerges dynamically through participation, coordination, adaptation, and collaborative sense-making between dynamically coupled systems. Within this perspective, intelligence is not reducible to isolated internal computation occurring independently inside humans or machines. Instead, intelligence unfolds relationally across: humans, AI systems, environments, interfaces, social systems, and evolving interaction trajectories.

This perspective reframes co-creative AI systems not merely as creative applications, but as important empirical environments for studying interaction-centered cognition itself. Co-creative systems increasingly provide computational platforms for investigating: participatory interaction, collaborative emergence, adaptive regulation, hybrid intelligence, and evolving coordination dynamics within human-AI systems.

As AI systems become increasingly integrated into human creative, cognitive, educational, and social environments, understanding interaction itself may become essential for developing future human-centered artificial intelligence systems. Rather than optimizing solely for isolated autonomous capability, future AI systems may increasingly depend upon: sustained participatory interaction, adaptive coordination, collaborative fluency, and dynamic co-regulation between humans and computational systems. More broadly, interaction-centered intelligence suggests a fundamental conceptual shift within artificial intelligence research. Intelligence may not reside solely inside isolated systems at all. Instead, intelligence may emerge dynamically through interaction itself.

References

Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1–3), 139–159.

Clark, A. (1997). Being there: Putting brain, body, and world together again. MIT Press.

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19.

Davis, N. (2017). Creative sense-making: A cognitive framework for quantifying interaction dynamics in co-creation (Doctoral dissertation, Georgia Institute of Technology).

Davis, N. (2024). Creative sense-making: A cognitive framework for modeling interaction dynamics in co-creative AI. In Artificial Intelligence, Co-Creation and Creativity. Routledge.

Davis, N., & Rafner, J. (2025). AI drawing partner: Co-creative drawing agent and research platform to model co-creation. arXiv preprint arXiv:2501.06607.

Davis, N., Hsiao, C.-P., Singh, K., Liapis, A., & Yannakakis, G. N. (2015). An enactive model of creativity for computational collaboration and co-creation. In Proceedings of the Sixth International Conference on Computational Creativity (pp. 360–367).

Davis, N., Hsiao, C.-P., Singh, K. Y., Li, L., Moningi, S., & Magerko, B. (2015). Drawing apprentice: An enactive co-creative agent for artistic collaboration. In Proceedings of the ACM SIGCHI Conference on Creativity and Cognition (pp. 185–186).

Davis, N., Popova, Y., & Sysoeva, K. (2017). Creative sense-making: Quantifying interaction dynamics in co-creation. In Proceedings of the ACM SIGCHI Conference on Creativity and Cognition (pp. 356–366).

Davis, N., Rezwana, J., Deshpande, M., & Magerko, B. (2025). The co-creative design framework for hybrid intelligence. Proceedings of the ACM on Human-Computer Interaction.

De Jaegher, H., & Di Paolo, E. (2007). Participatory sense-making: An enactive approach to social cognition. Phenomenology and the Cognitive Sciences, 6(4), 485–507.

Dellermann, D., Ebel, P., Söllner, M., & Leimeister, J. M. (2019). Hybrid intelligence. Business & Information Systems Engineering, 61(5), 637–643.

Deterding, S., Hook, J., Fiebrink, R., Gillian, N., Caires, C., Katan, S., & McPherson, A. (2017). Mixed-initiative creative interfaces. In Proceedings of the ACM SIGCHI Conference on Creativity and Cognition (pp. 628–635).

Di Paolo, E. (2005). Autopoiesis, adaptivity, teleology, agency. Phenomenology and the Cognitive Sciences, 4(4), 429–452.

Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin.

Guzdial, M., & Riedl, M. (2019). An interaction framework for studying co-creative AI. arXiv preprint arXiv:1903.09709.

Hutchins, E. (1995). Cognition in the wild. MIT Press.

Karimi, P., Davis, N., Grace, K., & Maher, M. L. (2018a). Deep learning for identifying potential conceptual shifts for co-creative drawing. arXiv preprint arXiv:1801.00723.

Karimi, P., Grace, K., Davis, N., & Maher, M. L. (2018b). Evaluating creativity in computational co-creative systems. arXiv preprint arXiv:1807.09886.

Karimi, P., Maher, M. L., Davis, N., & Grace, K. (2019). Deep learning in a computational model for conceptual shifts in a co-creative design system. arXiv preprint arXiv:1906.10188.

Lakoff, G., & Johnson, M. (1999). Philosophy in the flesh: The embodied mind and its challenge to Western thought. Basic Books.

LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444.

Lubart, T. (2005). How can computers be partners in the creative process: Classification and commentary on the special issue. International Journal of Human-Computer Studies, 63(4–5), 365–369.

Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267, 1–38.

Newell, A., & Simon, H. A. (1976). Computer science as empirical inquiry: Symbols and search. Communications of the ACM, 19(3), 113–126.

Rezwana, J., & Maher, M. L. (2022). Designing creative AI partners with COFI: A framework for modeling interaction in human-AI co-creative systems. arXiv preprint arXiv:2204.07666.

Russell, S., & Norvig, P. (2021). Artificial intelligence: A modern approach (4th ed.). Pearson.

Shneiderman, B. (2007). Creativity support tools: Accelerating discovery and innovation. Communications of the ACM, 50(12), 20–32.

Shneiderman, B. (2022). Human-centered AI. Oxford University Press.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind: Cognitive science and human experience. MIT Press.

Yannakakis, G. N., Liapis, A., & Alexopoulos, C. (2014). Mixed-initiative co-creativity. In Proceedings of the 9th International Conference on the Foundations of Digital Games.