Building Co-Creative AI Systems as a Method for Testing Enactive Theory: Bridging HCI, Enaction, and AI

Nicholas Davis, PhD

Abstract

Cognitive science and artificial intelligence were once closely braided, sharing methods, questions, and conceptual commitments. Today, however, they have largely diverged. Enactive and embodied approaches within cognitive science have developed rich theoretical accounts of sense-making, temporality, and lived interaction, while contemporary AI has advanced powerful systems primarily organized around optimization, prediction, and performance. This divergence has produced an asymmetry: cognitive theories increasingly lack constructive contact with artificial systems, while AI systems increasingly lack grounding in theories of cognition. This paper argues that building co-creative AI systems constitutes a form of theory-in-action through which enactive cognitive theories can be tested, refined, and made operational. We propose co-creative AI—particularly as developed within Human–Computer Interaction—as a methodological bridge that reunites theory and construction. By treating system-building as an epistemic practice rather than an application step, co-creative systems function as experimental instruments for cognitive theory, revealing constraints, failure modes, and commitments that abstract theorizing alone cannot expose.

1. Introduction: The Drift Between Cognitive Science and AI

Cognitive science, artificial intelligence, and human–computer interaction did not begin as separate enterprises. Early work in cybernetics, artificial intelligence, and cognitive modeling treated theory and construction as mutually informing: building systems was a way of thinking about cognition, and theories of cognition were expected to survive contact with working artifacts (Wiener, 1948; Ashby, 1956; Simon, 1969; Winograd & Flores, 1986). Over time, however, these traditions have drifted apart.

In contemporary cognitive science, particularly under enactive, embodied, and phenomenological approaches, cognition is understood as sense-making enacted through ongoing coupling with the world (Varela et al., 1991; Thompson, 2007; Di Paolo et al., 2017). Concepts such as viability, autonomy, temporality, and regulation now occupy a central place in theoretical accounts of mind (Maturana & Varela, 1980; Di Paolo, 2005; Weber & Varela, 2002). These developments have significantly enriched our understanding of cognition as lived, situated, and historically extended (Merleau-Ponty, 2012; Gallagher, 2017).

At the same time, artificial intelligence has undergone rapid technical expansion. Large-scale learning systems, generative models, and optimization-driven architectures now achieve impressive performance across a wide range of tasks (LeCun et al., 2015; Sutton & Barto, 2018). Yet these systems are largely developed independently of contemporary cognitive theory. Intelligence is typically framed in terms of prediction accuracy, reward maximization, or benchmark performance, and interaction is often reduced to input–output mapping rather than treated as an unfolding process of sense-making (Russell & Norvig, 2021; Lipton & Steinhardt, 2019).

The result is a growing gap. Cognitive theories of mind increasingly lack constructive instantiation, while artificial systems increasingly lack cognitive grounding. Enactive theories describe what cognition is, but are rarely stressed by systems that must operate, adapt, and remain coherent over time (Froese & Ziemke, 2009). AI systems, meanwhile, excel at short-horizon success while exhibiting subtle but persistent failures in long-term interaction—over-engagement, mis-timing, loss of trust, and breakdowns of coherence that are not easily captured by existing metrics (Lee & See, 2004; Sculley et al., 2015; Amodei et al., 2016).

This paper argues that Human–Computer Interaction—and, in particular, co-creative AI systems—provides a practical site where these traditions can be re-entangled. Co-creative systems force artificial agents to participate in ongoing human activity rather than merely produce outputs (Jordan & Henderson, 1995; Fischer et al., 2014). In doing so, they make unavoidable many of the core commitments of enactive theory: regulation over time, sensitivity to drift, pacing, yielding, and the maintenance of interactional coherence (Dourish, 2004; De Jaegher & Di Paolo, 2007).

This perspective aligns closely with more recent work on dynamic human–computer interaction (DHCI), which conceptualizes interaction as a temporally extended, adaptive process rather than a sequence of discrete user actions. Contemporary DHCI research emphasizes long-term interaction, mutual adaptation, breakdown and recovery, and the co-evolution of users and systems over time—particularly in contexts involving intelligent, autonomous, or learning systems (Bødker, 2015; Kuutti et al., 2021; Lim et al., 2019). Rather than treating interaction as momentary input–output exchange, DHCI foregrounds trajectories, histories, and transitions as primary analytic units. Co-creative AI systems naturally instantiate these concerns: they expose how pacing, yielding, responsiveness, and coherence must be regulated across unfolding interactional contexts. As such, DHCI provides a contemporary interactional framework in which enactive commitments—such as participatory sense-making and regulation under drift—become operationally unavoidable rather than theoretically optional.

We propose that building such systems is not an application of cognitive theory but a mode of theoretical inquiry. Co-creative AI systems function as epistemic instruments through which theories of cognition are enacted, constrained, and revised (Schön, 1983; Kirsh, 2013). By returning construction to the center of cognitive science—via HCI—we outline a path toward re-braiding enaction, AI, and interaction around the shared problem of understanding intelligence as something that unfolds over time.

Contributions

This paper makes three primary contributions.

First, it articulates a methodological diagnosis of the current divide between enactive cognitive science and artificial intelligence, arguing that the separation is not merely disciplinary but rooted in a shared retreat from constructive engagement on one side and from cognitive theory on the other. We characterize this condition as a form of theoretical and constructive drift that limits both explanatory power and practical viability (Suchman, 2007; Froese & Ziemke, 2009).

Second, the paper advances co-creative AI within HCI as a methodological hinge through which enactive theory and artificial systems can productively constrain one another. We argue that co-creative interaction—by demanding sustained, situated participation over time—forces cognitive commitments such as regulation, autonomy, and sense-making to become operational, while simultaneously exposing limitations of optimization-centric AI approaches that neglect interactional coherence (Bown et al., 2009; Davis et al., 2015; Jordanous & Keller, 2016).

Third, we propose building as a mode of theoretical inquiry in cognitive science. Rather than treating construction as downstream of theory, we show how co-creative systems function as epistemic instruments that reveal where cognitive concepts succeed, fail, or require refinement when enacted in real interaction (Schön, 1983; Kirsh, 2013; Di Paolo et al., 2017). This reframing supports a re-braided cognitive science in which theory, system-building, and lived interaction evolve together, offering a practical path toward cognitively grounded AI without reducing cognition to performance metrics or abstraction alone.

This paper does not propose a new interface or system for adoption, but a methodological stance for HCI research: that building co-creative systems can function as a form of cognitive experimentation.

2. Enaction Without Construction: Limits of Contemporary Enactive Cognitive Science

2.1 Enaction’s Contributions to Cognitive Theory

Enactive cognitive science has produced some of the most compelling accounts of cognition available today. By rejecting internalist, representational models, enaction reframes cognition as sense-making enacted through ongoing coupling between organism and environment (Varela et al., 1991; Di Paolo et al., 2017). Concepts such as autonomy, viability, structural coupling, and normativity have reshaped how cognition is understood—placing time, embodiment, and lived experience at the center rather than at the margins (Maturana & Varela, 1980; Weber & Varela, 2002; Thompson, 2007).

These contributions are not merely philosophical. Enactive approaches have clarified why cognition cannot be reduced to symbol manipulation or statistical inference alone (Hutto & Myin, 2017). They have foregrounded the role of affect, action readiness, and historical continuity in sustaining meaningful engagement with the world (Di Paolo, 2005; Colombetti, 2014). In doing so, enaction has offered a powerful ontology of cognition—one that emphasizes becoming over state, regulation over control, and participation over representation (Varela, 1979; Gallagher, 2017).

2.2 The Absence of Constructive Pressure

However, despite these strengths, much contemporary enactive work remains largely non-constructive. Enactive theories are often articulated at a conceptual or descriptive level, supported by phenomenological analysis or reinterpretations of empirical findings, but rarely stressed by systems that must actually operate over time (Froese & Ziemke, 2009). As a result, many core enactive commitments remain underdetermined with respect to implementation, dynamics, and failure.

This limitation is not accidental. Enactive cognitive science has, in part, defined itself in opposition to classical AI and computational modeling, which it correctly identified as overly representational and disembodied (Varela et al., 1991; Winograd & Flores, 1986). In distancing itself from those traditions, however, enaction has also retreated from system-building as a methodological practice. Artificial systems are frequently treated as metaphors, cautionary examples, or external foils rather than as sites where enactive theory might be enacted, tested, or challenged (Suchman, 2007; Newen et al., 2018).

2.3 The Gap Between Theory and Operation

The consequence is a growing gap between theoretical richness and operational clarity. Key enactive notions—such as regulation, autonomy, sense-making, and normativity—are invoked with increasing sophistication, yet often without specifying how these processes would function in systems that must sustain interaction, recover from misalignment, or adapt under ongoing change (Di Paolo & Thompson, 2014). Without constructive pressure, theories risk stabilizing internally while remaining permissive about their own commitments.

We refer to this condition as theoretical drift. Theoretical drift occurs when a framework remains coherent within its own discourse but gradually loses traction with practice. Concepts accumulate, distinctions proliferate, and explanatory narratives become more refined, yet the theory is no longer constrained by the demands of building something that must work (Kuhn, 1962; Schön, 1983). Importantly, theoretical drift is not a failure of rigor; it is a failure of contact.

2.4 Temporality Without Operational Constraint

This becomes especially visible when enactive theories are applied to phenomena that unfold over time. While enaction emphasizes temporality in principle, few enactive accounts are tested against systems that must manage long-duration interaction—where misalignment accumulates slowly, breakdowns are subtle, and viability depends on pacing, restraint, and historical sensitivity (Thompson, 2007; Rietveld et al., 2018). As a result, temporality often remains a descriptive feature rather than an operational constraint.

2.5 Ambiguity About Success and Failure

Without construction, enactive theories face a second risk: ambiguity about what counts as success or failure. In the absence of running systems, it becomes difficult to distinguish between concepts that are explanatorily necessary and those that are merely rhetorically appealing. Does a given account of regulation genuinely constrain behavior, or does it simply redescribe it? Does autonomy imply specific capacities, or does it function as a post hoc label? These questions cannot be resolved through theory alone (Beer, 2003; Froese & Ziemke, 2009).

2.6 Construction as Inquiry, Not Engineering

None of this suggests that enactive cognitive science must become engineering in the narrow sense. Rather, it suggests that enactive theory requires constructive counterparts—systems that instantiate its commitments sufficiently to expose their consequences. Building, in this context, is not a validation exercise but a form of inquiry (Schön, 1983; Kirsh, 2013). It reveals where concepts are underspecified, where assumptions conflict, and where additional distinctions are required.

In the absence of such constructive engagement, enactive cognitive science risks becoming phenomenologically rich but operationally thin. Its accounts may remain compelling, yet increasingly detached from the practical problem of how cognition—biological or artificial—maintains coherence under changing conditions. Re-engaging with construction is therefore not a retreat from enaction’s philosophical commitments, but a way of honoring them (Di Paolo et al., 2017).

The next section turns to the complementary problem: contemporary artificial intelligence systems that build powerful machinery while largely abandoning theories of cognition altogether.

3. AI Without Cognition: Limits of Contemporary Artificial Intelligence

3.1 Technical Success and Conceptual Narrowing

Contemporary artificial intelligence has achieved extraordinary technical success. Large-scale learning systems now demonstrate impressive performance across perception, language, and control tasks. Advances in representation learning, optimization, and scaling have produced systems that are robust, flexible, and widely deployable (LeCun et al., 2015; Brown et al., 2020; Bommasani et al., 2021). From an engineering standpoint, modern AI is undeniably powerful.

Yet this success has come with a narrowing of what intelligence is taken to be. In dominant AI paradigms, intelligence is framed primarily as performance on well-defined tasks: prediction accuracy, reward maximization, loss minimization, or benchmark superiority (Russell & Norvig, 2021; Sutton & Barto, 2018). Learning is evaluated in terms of convergence and generalization, and interaction is often treated as a transient channel for input and output rather than as an ongoing process of sense-making (Bender & Koller, 2020).

3.2 Performance Without Viability

This framing has proven effective for short-horizon problems, but it carries significant cognitive limitations. Most contemporary AI systems are optimized to act correctly, not to remain viable. They excel at producing appropriate outputs under assumed conditions, yet struggle to sustain coherent behavior across extended interaction, shifting contexts, or evolving norms (Amodei et al., 2016; Marcus, 2018). When breakdown occurs, it is typically handled through retraining, fine-tuning, or external intervention rather than through endogenous regulation.

From a cognitive perspective, this reflects a fundamental asymmetry. Human and animal cognition is not organized primarily around performance metrics, but around maintaining coherent engagement with the world over time (Di Paolo et al., 2017; Thompson, 2007). Cognition persists by regulating participation—by adjusting pacing, yielding control, reorienting attention, and responding to subtle signs of misalignment before failure becomes explicit (Rietveld et al., 2018). These processes are largely absent from contemporary AI architectures, not because they are infeasible, but because they are rarely treated as first-class design problems (Suchman, 2007).

3.3 Interaction as Interface Rather Than Relationship

Interaction, in particular, is frequently reduced to a technical interface problem. Users provide inputs; systems return outputs. Even in interactive learning settings, the dominant question is how to extract signal from human behavior efficiently, rather than how to participate meaningfully in a shared activity (Amershi et al., 2014; Kocaballi et al., 2020). As a result, interaction is modeled as data exchange rather than as a temporally extended relationship.

This reduction obscures a key issue: intelligence that unfolds over time cannot be evaluated solely through instantaneous performance. Systems that appear competent in isolated episodes may nevertheless accumulate misalignment, rigidity, or overconfidence when deployed in real-world settings (Dietterich, 2019; Mitchell et al., 2021). The growing literature on distribution shift, dataset bias, and model brittleness reflects this problem, yet these phenomena are often treated as technical pathologies rather than as symptoms of a deeper cognitive gap (Quinonero-Candela et al., 2009; Ovadia et al., 2019).

3.4 The Absence of Sense-Making Mechanisms

That gap concerns sense-making. Contemporary AI systems typically do not maintain an internal distinction between doing well and being aligned with the situation. They lack mechanisms for detecting when continued action is no longer appropriate, when participation should be modulated, or when coherence is degrading despite locally valid behavior (Floridi et al., 2018; Russell, 2019). As a result, they may persist confidently in regimes that appear successful according to internal metrics but are interactionally failing.

Crucially, this is not a matter of insufficient scale or data. Increasing model capacity can improve short-term adaptability, but it does not by itself introduce the capacity for self-regulation under drift (Bengio et al., 2021). Without architectural commitments to temporality, coupling, and participation, systems remain brittle over extended interaction—no matter how impressive their immediate outputs (Marcus & Davis, 2019).

3.5 Construction Without Cognitive Theory

From the perspective of enactive cognitive science, this reveals a missed opportunity. AI systems constitute one of the few domains where cognitive theories could be materially instantiated and stressed under real conditions. Yet most contemporary AI research proceeds with minimal engagement with theories of cognition, embodiment, or sense-making (Newell, 1990; Lake et al., 2017). Cognition is implicitly assumed rather than explicitly modeled.

The result is a form of cognitive underdetermination: systems that work remarkably well without clarifying what kind of intelligence they embody. Benchmarks are passed, but questions about autonomy, regulation, and viability are deferred or externalized (Rahimi et al., 2019). Intelligence becomes something inferred from outputs rather than understood as an ongoing organizational achievement (Bender et al., 2021).

3.6 Why the Absence of Cognition Matters

This does not imply that contemporary AI must adopt enactive theory wholesale. Rather, it suggests that AI’s current trajectory leaves key dimensions of intelligence unaddressed—particularly those that matter in sustained interaction with humans (Dreyfus, 2007; Suchman, 2007). The absence of cognitive theory is not neutral; it shapes what systems can notice, how they fail, and how responsibility for breakdown is distributed (Benjamin, 2019).

If enactive cognitive science risks becoming theory without construction, contemporary AI risks becoming construction without cognition. The challenge, then, is not to privilege one over the other, but to identify a domain in which both must confront each other’s limits.

The next section argues that human–computer interaction provides precisely such a domain—a space where cognition, construction, and lived experience are forced into direct contact.

4. Human–Computer Interaction as a Bridging Discipline

4.1 HCI Between Cognitive Theory and Artificial Systems

Human–computer interaction occupies a distinctive position between cognitive theory and artificial systems. Unlike much of cognitive science, HCI cannot remain purely descriptive; unlike much of AI, it cannot treat interaction as incidental. Systems must be encountered by real users, over time, in contexts where breakdown, confusion, frustration, and trust are unavoidable. This makes HCI an unusually demanding site for cognitive ideas (Norman, 2013; Rogers, 2012).

Where enactive cognitive science emphasizes lived experience, coupling, and sense-making, HCI operationalizes these concerns by necessity (Suchman, 2007). Interaction unfolds temporally. Misalignment accumulates. Users adapt, resist, reinterpret, or disengage. A system that performs well in isolation may still fail interactionally, not because it is inaccurate, but because it cannot pace itself, yield control, or remain intelligible as circumstances change (Dourish, 2001; Höök, 2018).

4.2 Interaction as a Temporal and Normative Process

For this reason, HCI functions as a phenomenological laboratory for cognitive theory. Concepts such as engagement, breakdown, coordination, and repair are not abstract topics but empirical realities encountered in use (Winograd & Flores, 1986; Suchman, 1987). Designers cannot assume stable meanings, fixed goals, or uniform users. Instead, they must confront how sense-making emerges in practice—how it stabilizes, drifts, and sometimes collapses (Rogers, Sharp, & Preece, 2011).

Crucially, HCI also constrains theory through construction. Cognitive commitments become embedded in interface timing, system responsiveness, feedback structure, and permissible actions. A theory that emphasizes coupling but produces systems that interrupt, dominate, or over-automate reveals a mismatch between its conceptual claims and its operational consequences (Höök et al., 2016). In this way, HCI exposes gaps between what a theory says cognition is and what it actually enables systems to do.

4.3 Viability as an Interactional Criterion

This distinguishes HCI from post-hoc evaluation approaches. Rather than asking whether a system achieves predefined outcomes, HCI asks whether interaction remains viable. Does the system remain understandable? Does it support repair? Can users regain orientation after breakdown? These questions align directly with enactive concerns about sense-making and normativity, yet they demand concrete answers in the form of working systems (Di Paolo et al., 2017; Kaptelinin & Nardi, 2012).

Historically, HCI has often been framed as a downstream application area—where insights from psychology or AI are translated into usable interfaces. However, this framing understates its epistemic role. HCI does not merely apply cognitive theories; it tests their consequences (Rogers, 2012). It forces theories to encounter time, variability, and situated use. In doing so, it reveals which theoretical distinctions matter and which dissolve under interactional pressure (Suchman, 2007).

4.4 Creative Interaction as a Stress Test for Cognitive Theory

This is especially apparent in creative and co-creative systems. Creative interaction cannot be reduced to goal satisfaction or task completion. It requires sensitivity to timing, openness to deviation, and tolerance for ambiguity (Fischer et al., 2014). Systems must know when to act, when to hold back, and when to let human activity reconfigure the interaction. These demands make creative HCI a particularly rich testbed for enactive ideas (Höök, 2018; Davis et al., 2021).

Moreover, HCI provides a shared language across disciplines. Terms such as interactional breakdown, repair, legibility, trust, and pacing are intelligible to designers, engineers, and cognitive scientists alike (Norman, 2013; Rogers et al., 2011). This makes HCI a practical site for re-entangling enactive theory and AI construction without requiring either community to abandon its core commitments.

4.5 HCI as a Load-Bearing Structure

From this perspective, HCI is not merely a bridge between enaction and AI—it is the load-bearing structure that makes their reunion possible. It supplies the experimental ecology in which cognitive theories must survive contact with real systems and real users. It also supplies AI with criteria of success that extend beyond optimization toward sustained participation, intelligibility, and trust (Shneiderman, 2020).

The argument of this paper is therefore not that HCI should import enactive theory wholesale, nor that AI should be subordinated to cognitive science. Rather, HCI provides the space in which enactive ideas can become operational and AI systems can become cognitively accountable. It is the discipline in which theory, construction, and lived interaction are forced into alignment.

Table X situates co-creative AI at the intersection of three traditions that are frequently invoked together but rarely compared at the level of their operative commitments: enactive cognitive science, artificial intelligence, and human–computer interaction. Rather than contrasting these fields in terms of disciplinary scope or technical maturity, the table highlights how each addresses core dimensions of co-creative interaction—such as temporality, regulation, drift, autonomy, and evaluation—through distinct methodological lenses. The comparison makes visible both complementarities and blind spots: what each tradition foregrounds, what it tends to abstract away, and which phenomena become legible only when systems must participate in sustained interaction with human partners. In doing so, the table supports the paper’s central claim that co-creative AI functions as a methodological hinge—one that exposes the limits of isolated approaches and motivates a re-braided framework in which theory, construction, and interaction mutually constrain one another.

The next section builds on this claim by advancing a stronger methodological position: that building cognitive systems is itself a form of theory, and that co-creative AI systems make this reflective relationship explicit.

5. Building as Theory: Reflection-in-Action for Cognitive Science

5.1 The Traditional Separation of Theory and Construction

Cognitive science has traditionally treated theory and construction as separate phases. Theories are developed to explain cognition; systems are then built to apply, illustrate, or evaluate those theories. In this framing, construction is downstream of theory—instrumental rather than epistemic (Newell & Simon, 1976; Marr, 1982).

This paper advances a different claim: building cognitive systems is itself a form of theoretical work. Construction is not merely an implementation step, but a mode of reflection that exposes what a theory actually commits to when enacted in a dynamic system (Schön, 1983; Agre, 1997).

5.2 Reflection-in-Action as a Methodological Stance

This view draws on traditions of reflection-in-action, where understanding emerges through making rather than prior to it (Schön, 1983). When a cognitive theory is embedded in a working system, its assumptions cease to be abstract. They become constraints on behavior, timing, and interaction. The system does not merely “instantiate” the theory; it tests it continuously through ongoing engagement with an environment and with human participants (Hutchins, 1995).

In this sense, theories structure behavior before they explain it. Decisions about what a system can perceive, how it responds, when it acts, and when it refrains are theoretical commitments, whether or not they are labeled as such (Agre & Chapman, 1987). Building makes these commitments explicit and forces them to confront real conditions: noise, drift, variability, and breakdown (Suchman, 2007).

5.3 Failure as Theoretical Revelation

Conversely, working systems reveal theory. When a system fails—not by crashing, but by becoming intrusive, incoherent, or unresponsive—it exposes mismatches between theoretical intent and operational reality (Norman, 2013). Concepts that appear coherent on paper may prove unstable in interaction. Others, previously underspecified, become unavoidable once a system must act over time.

Cognitive prototypes thus function as epistemic probes. They are not demonstrations of correctness, but instruments for discovering the limits, ambiguities, and productive tensions within a theory (Hutchins, 1995; Gaver, 2012). By observing how a system behaves across extended interaction, researchers gain access to phenomena that static analysis cannot reveal: gradual misalignment, interactional fatigue, over-engagement, and the subtle conditions under which sense-making is sustained or lost (Di Paolo et al., 2017).

5.4 Beyond Benchmark-Driven Validation

This perspective contrasts sharply with benchmark-driven validation. Benchmarks assume stationarity, well-defined tasks, and external success criteria (Sculley et al., 2018). Reflection-in-action instead treats cognition as something that unfolds, reorganizes, and sometimes degrades. The relevant question is not whether a system achieves optimal performance, but whether it remains coherently engaged as conditions change.

Importantly, this is not an argument for engineering pragmatism at the expense of theory. On the contrary, it is an argument for theoretical accountability. A theory that cannot be built—or that produces systems incapable of sustained interaction—reveals a gap between its explanatory ambition and its operational substance (Agre, 1997; Varela et al., 1991).

5.5 Temporal Stress as a Criterion for Theory

Building, in this view, is not about validating theories as true or false. It is about subjecting them to temporal stress. Theories that survive construction do so not by eliminating complexity, but by regulating it. They remain intelligible when embedded in systems that must act, wait, adapt, and sometimes fail (Di Paolo & Thompson, 2014).

For enactive cognitive science, this methodological stance is especially consequential. Enaction emphasizes sense-making as an ongoing, relational process rather than a static computation (Varela et al., 1991). Such claims cannot be adequately assessed through offline models or short-horizon tasks. They require systems that participate over time—systems whose viability depends on their capacity to regulate engagement rather than optimize outcomes (Di Paolo et al., 2017).

5.6 Where Co-Creative AI Fits

Co-creative AI systems provide precisely this setting. They force theories of sense-making to confront interactional reality. They expose where concepts like coupling, autonomy, and agency become actionable, and where they remain metaphorical (Höök, 2018; Davis et al., 2021). In doing so, they transform cognitive system building from a matter of application into a site of theoretical discovery.

The next section develops this argument concretely by focusing on co-creative AI as an enactive testbed—one where reflection-in-action is not incidental, but unavoidable.

6. Co-Creative AI as an Enactive Testbed

6.1 Why Co-Creation Is a Demanding Cognitive Setting

Co-creative AI systems offer a uniquely demanding testbed for enactive cognitive theory. Unlike task-oriented or generative systems that operate in short, well-defined episodes, co-creative systems must participate in ongoing human activity. They do not simply produce artifacts; they share a process. As a result, they cannot avoid the core phenomena that enactive theory emphasizes: coupling, temporality, sense-making, and viability (Varela et al., 1991; Di Paolo et al., 2017).

Co-creation foregrounds interaction as the primary unit of analysis. In these systems, meaning does not reside in outputs alone, but in how contributions are timed, sustained, adapted, or withheld in response to a human partner (Sawyer, 2012). This shifts the evaluative focus from correctness or novelty to interactional coherence: whether the system remains appropriately engaged as the activity unfolds (Höök, 2018).

6.2 Drift as the Default Condition of Interaction

From an enactive perspective, this setting is not incidental—it is ideal. Co-creative interaction makes drift unavoidable. Human goals evolve, attention fluctuates, styles shift, and phases of activity open and close without explicit signals. Any system that participates meaningfully over time must contend with these changes. Drift is not an edge case; it is the default condition of sustained interaction (Suchman, 2007; De Jaegher & Di Paolo, 2007).

This is precisely where many AI systems fail. Generative or assistive systems often respond competently in isolation yet degrade during extended use. They over-elaborate when restraint is needed, persist beyond the natural end of a creative phase, or remain active when yielding would preserve coherence (Amershi et al., 2019). These failures are rarely captured by standard metrics, but they are immediately apparent to human collaborators.

6.3 Experiential Signals of Misalignment

Co-creative AI exposes these breakdowns with unusual clarity. Because interaction is continuous and visible, misalignment manifests as experiential friction rather than numerical error. Users sense when a system is “out of sync”—too eager, too repetitive, too dominant, or insufficiently responsive (Norman, 2013; Höök, 2018). These judgments are not subjective noise; they are signals of regulatory failure at the interactional level.

For enactive theory, this setting is revealing. Concepts such as autonomy, coupling, and sense-making cease to be abstract descriptors and become operational necessities (Di Paolo et al., 2017). A system that cannot regulate how it participates—how long it persists, how strongly it couples, when it yields—cannot sustain co-creation, regardless of its generative capabilities.

6.4 Regulation as Modulation, Not Control

Importantly, co-creative systems make clear that regulation is not equivalent to control. Successful co-creative agents do not enforce plans or optimize toward predefined goals. Instead, they modulate their engagement: stabilizing certain interactional patterns while allowing others to dissolve. They exhibit pacing, restraint, and sensitivity to phase transitions—qualities that are central to enactive accounts of cognition but rarely instantiated in AI systems (Varela et al., 1991; Di Paolo & Thompson, 2014).

This testbed also exposes the limits of representation-centric approaches. In co-creative interaction, there is often no stable task representation to infer and no fixed objective to optimize. What matters is the system’s ability to remain attuned to an evolving situation (Suchman, 2007). Sense-making here is enacted through timing and participation, not internal world models alone.

6.5 Co-Creative Systems as Theoretical Crucibles

Concrete examples of this dynamic can be seen in enactive co-creative agents such as drawing partners that regulate stroke density, continuation, and silence based on interactional context; or systems that adjust their level of initiative in response to user engagement rather than maximizing output (Davis et al., 2021; Yannakakis et al., 2014). While the technical details vary, the shared feature is that these systems treat participation itself as the locus of adaptation.

Through such systems, enactive theory encounters resistance. Not all conceptual distinctions survive implementation. Some notions prove too coarse to guide action; others become unexpectedly central. Drift, in particular, emerges as a structural phenomenon: not a failure of prediction, but a mismatch between the timescale of regulation and the timescale of interactional change.

Co-creative AI thus functions as more than an application domain. It is a theoretical crucible. It forces enactive concepts to become actionable, exposes their ambiguities, and reveals where additional structure is required. By demanding sustained, situated participation, co-creative systems transform enaction from a descriptive framework into a set of design constraints that must be negotiated in practice.

The following section examines what emerges from this process—what building co-creative systems reveals about cognition that theory alone cannot show.

7. What Building Reveals That Theory Alone Cannot

7.1 Construction as a Source of Theoretical Transformation

Constructing co-creative cognitive systems does more than instantiate existing theory. It actively reshapes it. Certain properties of cognition—while conceptually acknowledged in enactive accounts—only become precise, unavoidable, or problematic when they must be operationalized in a running system (Varela et al., 1991; Di Paolo et al., 2017). Building exposes where theory is underspecified, where distinctions collapse, and where new structure is required.

This transformative role of construction aligns with long-standing accounts of reflection-in-action, in which making is itself a mode of thinking rather than a downstream application of ideas (Schön, 1983). When systems must operate continuously, theoretical commitments are no longer optional—they become constraints that directly shape behavior.

7.2 Drift as a Structural Phenomenon

One such revelation concerns drift. In theory, drift is often treated as a secondary effect: noise, variability, or gradual deviation from a norm. In practice, drift proves to be structural. Any system engaged in sustained interaction accumulates misalignment, even when local behavior remains appropriate (Suchman, 2007). Drift emerges not from error, but from persistence.

Systems that continue to act coherently in the short term may nonetheless erode interactional viability over time. This phenomenon is difficult to appreciate in abstract accounts but becomes immediately apparent in deployed co-creative systems, where misalignment accumulates quietly rather than catastrophically (De Jaegher & Di Paolo, 2007).

7.3 Regulation as the Shaping of Participation

A second insight concerns regulation. Enactive theory emphasizes regulation as central to autonomy and sense-making, yet often leaves its mechanics implicit. Building reveals that regulation is not primarily about selecting correct actions or maintaining internal variables within bounds. It is about shaping participation (Di Paolo & Thompson, 2014).

Effective systems regulate how strongly they couple, how long they persist, and when they release control. These dimensions—pacing, yielding, restraint—are not easily captured by performance metrics, but they dominate the success or failure of co-creative interaction (Höök, 2018).

7.4 Coherence Versus Accuracy

Construction also clarifies the relationship between coherence and accuracy. In operational systems, these objectives frequently diverge. A system can be accurate in its immediate contributions while undermining the longer-term coherence of the interaction. Conversely, a system may temporarily reduce local appropriateness in order to preserve trust, interpretability, or phase alignment (Norman, 2013).

Theory alone often assumes these goals align; building shows that they do not. Sustained sense-making depends more on coherence across time than on momentary success—a distinction that becomes unavoidable in extended co-creative engagement (Suchman, 2007).

7.5 Temporal Granularity as a Design Constraint

Another lesson concerns temporal granularity. Enactive theories rightly emphasize temporality, but building reveals that which timescale a system regulates is decisive. Systems that regulate too slowly fail to respond to emerging shifts; systems that regulate too quickly become jittery or overreactive.

Many breakdowns in co-creative AI are not due to missing information, but to misaligned temporal resolution—regulating at the wrong grain for the interaction at hand (Kelso, 1995). This insight forces theory to move beyond generic appeals to “time” and toward explicit accounts of temporal coordination.

7.6 Beyond Representational Sufficiency

Building also exposes the limits of representational explanations. In co-creative interaction, meaning often cannot be localized to internal states or inferred goals. What matters is not what the system “believes,” but how it behaves in relation to an evolving situation (Brooks, 1991; Suchman, 2007).

Systems succeed or fail based on interactional fit, not internal correctness. This does not negate representation, but it deprioritizes it as the primary explanatory resource for intelligent behavior in situated, participatory contexts.

7.7 Failure Modes Invisible to Theory Alone

Perhaps most importantly, construction reveals failure modes invisible to abstract theory. Systems may become trapped in locally coherent but globally inappropriate regimes. They may persist gracefully long after an interaction should end. They may appear adaptive while silently exhausting their human collaborators (Amershi et al., 2019).

These are not errors in the conventional sense, yet they are decisive for real-world viability. Only through running systems—over time, with real users—do these pathologies become legible. In this way, building functions as a form of epistemic constraint. It forces theoretical commitments to confront the realities of timing, fatigue, misalignment, and breakdown.

Concepts that survive this encounter gain precision; those that do not must be revised or abandoned. Theory is not merely applied—it is disciplined by construction.

7.8 Toward a Unified Cognitive Science

This does not diminish the value of enactive theory. On the contrary, it strengthens it. By engaging with co-creative systems, enaction moves from a largely interpretive framework to a generative one—capable of guiding design, explaining failure, and motivating new distinctions grounded in practice (Di Paolo et al., 2017).

The final sections synthesize these insights into a broader proposal: a re-braided cognitive science in which theory, interaction, and construction evolve together rather than apart.

8. Toward a Re-Braided Cognitive Science

8.1 A Methodological, Not Merely Disciplinary, Divide

The separation between enactive cognitive science and artificial intelligence is often described as a disciplinary rift: different conferences, journals, vocabularies, and evaluative norms. While these sociological factors are real, they obscure a deeper issue. The divide is fundamentally methodological. The two traditions do not merely study different phenomena; they operationalize knowledge in fundamentally different ways.

Enactive cognitive science has developed a sophisticated account of cognition as sense-making enacted through embodied, situated, and temporally extended engagement (Varela et al., 1991; Di Paolo et al., 2017). Its core commitments—autonomy, normativity, structural coupling, and viability—redefine cognition as an ongoing organizational achievement rather than a computational process over representations. Yet despite this emphasis on process and enactment, much enactive work remains analytically descriptive rather than constructively generative. Its theories are articulated through conceptual analysis, phenomenology, and reinterpretation of empirical results, but are rarely stressed by systems that must act, persist, and adapt under real temporal pressure.

Artificial intelligence, by contrast, is profoundly constructive. It builds systems that operate at scale, adapt to data, and perform under real constraints. However, contemporary AI increasingly proceeds without explicit theories of cognition. Intelligence is treated as an emergent property of optimization, scaling, or architectural ingenuity rather than as an explanandum in its own right (Russell & Norvig, 2021). Cognitive concepts—autonomy, understanding, sense-making—are either implicitly assumed or explicitly bracketed as philosophical concerns external to engineering practice.

The result is a mutual withdrawal. Enaction retreats from construction in order to preserve its philosophical integrity. AI retreats from cognitive theory in order to maximize technical progress. What is lost in this mutual disengagement is a site where theories of cognition must confront the consequences of being enacted, and where artificial systems must answer to more than performance metrics. Re-braiding cognitive science therefore requires addressing this methodological gap directly, not merely encouraging cross-disciplinary dialogue.

8.2 Co-Creative AI as a Methodological Hinge

This paper argues that such a site already exists in practice: the design and study of co-creative AI systems within HCI. Co-creative systems—drawing partners, musical collaborators, interactive writing tools—are not peripheral curiosities. They occupy a methodological position that neither enactive theory nor mainstream AI can avoid without cost (Davis et al., 2017; Kantosalo & Toivonen, 2016).

Co-creative AI functions as a hinge because it forces two pressures simultaneously. On one side, it forces cognitive theory to become operational. Concepts such as coupling, autonomy, regulation, and sense-making must be instantiated in systems that actually interact with humans over time. Vague appeals to emergence or participation are insufficient when a system must decide when to act, when to wait, and when to disengage.

On the other side, co-creative AI forces artificial systems to confront interaction as a first-class phenomenon. Unlike task-oriented systems, co-creative agents cannot rely on stable objectives, clearly defined success conditions, or short evaluation horizons. They must remain intelligible, responsive, and appropriately restrained as human partners adapt, drift, and reinterpret the activity (Suchman, 2007; Höök, 2018).

This dual pressure makes co-creative AI uniquely valuable. It does not allow enaction to remain purely interpretive, nor AI to remain purely instrumental. It creates a methodological space in which theories and systems must constrain one another without collapsing into one another.

8.3 Complementary Roles: Enaction, AI, HCI, and Building

When viewed through this lens, the apparent competition between enaction, AI, and HCI dissolves. Each plays a distinct and necessary role in a re-braided cognitive science.

Enaction supplies the ontological commitments. It insists that cognition is enacted, relational, and temporally extended—that meaning arises through participation rather than internal representation alone (Varela et al., 1991; Thompson, 2007). These commitments define what counts as a cognitively adequate system in the first place.

AI supplies the machinery. It provides learning algorithms, adaptive architectures, and computational substrates capable of sustained behavior. Without AI, enactive commitments risk remaining aspirational; with AI alone, cognition risks being reduced to optimization.

HCI supplies the grounding. It places systems in contact with real users, real contexts, and real breakdowns. HCI refuses idealized assumptions about stable goals, compliant users, or frictionless interaction. It is where cognitive claims encounter lived experience (Norman, 2013; Suchman, 2007).

Building itself supplies the epistemic constraint. Construction prevents theory from drifting into abstraction and prevents systems from drifting into unguided performance. It forces decisions: what to regulate, what to sense, what to allow, and what to forbid. These decisions are not merely engineering choices; they are theoretical commitments made concrete.

8.4 Building as a Mode of Theoretical Thought

This perspective reframes building as a mode of thinking rather than a downstream application of thought. Artificial systems, especially co-creative ones, become instruments of inquiry—ways of probing the consequences of our theoretical commitments (Schön, 1983).

When a theory is built into a system, its assumptions acquire operational force. Claims about autonomy become constraints on initiative. Claims about coupling become decisions about responsiveness. Claims about normativity become mechanisms for pacing, restraint, and yielding. Building forces theories to answer questions they can otherwise defer: How fast should regulation occur? What happens when coherence degrades gradually rather than catastrophically? When should a system stop?

In this sense, building is not merely illustrative. It is reflective. It reveals where concepts are underspecified, where distinctions collapse under pressure, and where new distinctions are required. A system that behaves intrusively, rigidly, or incoherently exposes theoretical gaps that no amount of conceptual refinement alone can reveal.

8.5 Resolving the Theory–Construction Oscillation

Cognitive science has long oscillated between two unsatisfying poles. At one extreme lies abstract theorizing detached from implementation, rich in conceptual nuance but permissive about consequences. At the other lies system-building detached from theory, powerful in practice but thin in explanation.

A re-braided cognitive science rejects this oscillation. It treats construction as a form of theoretical articulation—one that exposes commitments through behavior rather than prose. Theory does not precede construction once and for all; it evolves through cycles of building, observing, revising, and rebuilding (Di Paolo et al., 2017).

This stance does not subordinate theory to engineering, nor does it instrumentalize construction as mere validation. Instead, it treats construction as a way of asking better theoretical questions—questions that only arise when systems must act over time, interact with others, and survive their own participation.

8.6 Why Co-Creative Systems Matter

Co-creative systems are especially potent in this role because they resist simplification. They cannot be reduced to task completion, reward maximization, or static evaluation. Creative interaction unfolds in phases, tolerates ambiguity, and depends on timing and restraint as much as contribution (Sawyer, 2012).

In such systems, drift is unavoidable. Human partners change direction, lose interest, reinterpret goals, or silently conclude an activity. A system that cannot detect and respond to these shifts—by yielding, pausing, or terminating—will fail regardless of the quality of its outputs.

This makes co-creative systems a crucible for enactive concepts. Sense-making must be enacted rather than inferred. Autonomy must be regulated rather than asserted. Viability must be sustained rather than assumed. Coupling must be modulated rather than maximized.

Because these demands are experiential and immediate, co-creative interaction exposes breakdowns that benchmarks cannot capture. Users feel when a system is out of sync. These felt judgments are not noise; they are signals of cognitive misalignment (Höök, 2018).

8.7 A Shift in the Meaning of Success

Re-braiding cognitive science also requires a shift in evaluation. Success can no longer be defined primarily in terms of optimal performance, accuracy, or output quality. Instead, success becomes coherence: the capacity to sustain meaningful participation over time.

This shift aligns enaction, AI, and HCI around a shared evaluative center. It foregrounds interactional viability, interpretability, and trust rather than isolated competence (Lee & See, 2004). A system may be locally impressive yet globally corrosive; another may be modest in output yet sustaining in engagement.

By prioritizing coherence, evaluation becomes sensitive to phenomena that matter in real use: pacing, fatigue, overreach, and graceful disengagement. These criteria are not anti-technical; they are cognitively grounded.

8.8 From Unified Theory to Unified Practice

The outcome of this re-braiding is not a grand unified theory of cognition. It is a unified practice. A way of doing cognitive science in which theory and construction co-evolve.

In this practice, artificial systems are neither mere models nor mere products. They are experimental partners—entities through which theoretical commitments become visible, contestable, and revisable. Cognitive science becomes a discipline that thinks by building and builds in order to think (Schön, 1983).

Such a practice is necessarily iterative. It values revision over closure and engagement over abstraction. It treats failure not as error but as information—evidence about the limits of our concepts.

8.9 Re-Uniting Cognition and AI Without Nostalgia

Finally, re-uniting cognition and AI in this way is not a return to early symbolic AI or an attempt to resurrect a lost synthesis. It is a forward-looking recognition that understanding intelligence requires systems that must live with the consequences of their own participation.

Co-creative AI, situated within HCI, offers a concrete path toward this engagement. It creates conditions under which cognition, theory, and construction cannot drift apart. In doing so, it makes intelligence—biological or artificial—not merely something we measure, but something we can study as an ongoing, lived achievement.

9. Implications and Future Directions

The perspective advanced in this paper has implications that cut across cognitive science, artificial intelligence, and human–computer interaction. Rather than proposing a new algorithm or architecture, it reframes what counts as progress in the study of cognition and intelligent systems. By treating co-creative system building as a form of theoretical inquiry, it opens new methodological and evaluative pathways for each field.

9.1 Implications for Cognitive Science

For cognitive science—particularly traditions grounded in enaction and embodiment—the central implication is methodological. If cognition is understood as enacted, temporally extended, and relational, then theories of cognition must remain buildable (Varela et al., 1991; Di Paolo et al., 2017). The inability to construct systems that exhibit the dynamics a theory describes should not be treated as a mere engineering gap; it should be taken as a diagnostic signal about the theory’s operational completeness.

This does not imply that all cognitive theory must be immediately implemented, nor that artificial systems must replicate biological cognition. Rather, it suggests that constructive engagement is a necessary complement to phenomenological and conceptual analysis (Thompson, 2007). Co-creative systems provide a particularly suitable testbed because they expose the very phenomena enactive theories emphasize—sense-making, breakdown, regulation, and recovery—under conditions that cannot be idealized away.

More broadly, this perspective encourages cognitive science to treat artificial systems not only as models or metaphors, but as experimental instruments capable of revealing the consequences of theoretical commitments over time (Beer, 2003).

9.2 Implications for Artificial Intelligence Research

For AI, the implications concern both evaluation and design priorities. Contemporary AI research overwhelmingly emphasizes performance under fixed task formulations, often evaluated through benchmarks that abstract away interactional context and temporal continuity (Lipton & Steinhardt, 2019). The co-creative perspective challenges this framing.

If intelligence is understood as the capacity to sustain coherent participation under change, then short-horizon success is insufficient. Systems must be evaluated in terms of how they regulate engagement, respond to drift, and manage transitions across phases of interaction. This calls for evaluation practices that emphasize trajectories, histories, and breakdown handling rather than isolated outputs (Sculley et al., 2015; Gama et al., 2014).

From a design standpoint, this perspective suggests that progress in AI may depend less on increasing expressivity or scale, and more on developing regulatory capacities—mechanisms for pacing, yielding, restraint, and termination. These capacities are rarely foregrounded in mainstream AI, yet they are central to any system intended to interact with humans over time (Lee & See, 2004; Russell, 2019).

Importantly, this is not a rejection of existing AI methods. It is a proposal to re-situate them within interactional ecologies where cognition becomes visible as an ongoing process rather than a static competence.

9.3 Implications for Human–Computer Interaction

For HCI—and especially communities such as Creativity & Cognition—this work positions co-creative systems as more than interaction design challenges or experiential artifacts. It argues that they can function as sites of cognitive discovery (Dourish, 2004; Höök, 2018).

HCI uniquely confronts the realities that both cognitive theory and AI often abstract away: breakdowns, misalignment, trust erosion, over-participation, and disengagement (Suchman, 2007). Co-creative systems make these phenomena unavoidable, and therefore measurable, discussable, and designable.

This suggests a reframing of co-creative evaluation. Rather than asking whether systems are novel, useful, or engaging in isolation, HCI research can ask how systems regulate participation over time, how users adapt to them, and how mutual coherence is maintained or lost (Jordan & Henderson, 1995). In this way, HCI becomes not only a consumer of cognitive theory, but a generator of theoretical insight.

9.4 Future Research Directions

Several concrete research directions follow from this perspective.

First, there is a need for new evaluative frameworks that capture interactional coherence over time. This includes methods for analyzing regime stability, transition dynamics, and forms of graceful breakdown or withdrawal (Di Paolo & De Jaegher, 2012).

Second, further work is needed to explore how enactive regulatory concepts scale—both to richer modalities and to more complex internal architectures. Understanding how regulation operates in systems with high-dimensional representations remains an open challenge (Parr et al., 2022).

Third, comparative studies across co-creative domains—drawing, music, writing, and collaborative problem solving—could help identify which regulatory phenomena are domain-general and which are domain-specific (Bown et al., 2009; Sawyer, 2012).

Finally, closer collaboration between cognitive theorists, AI researchers, and HCI practitioners will be essential. The perspective advanced here depends on iterative movement between theory and construction, rather than disciplinary isolation.

10. Conclusion

This paper has argued that the growing separation between enactive cognitive science and artificial intelligence is not merely theoretical, but methodological. Enactive accounts of cognition emphasize sense-making, coupling, and temporal coherence, yet too often remain detached from the construction of systems that must actually sustain these dynamics. Conversely, AI systems increasingly achieve impressive performance while lacking principled accounts of interaction, breakdown, and long-term coherence.

We have proposed co-creative AI systems—developed and studied through human–computer interaction—as a practical bridge between these traditions. By participating in ongoing activity rather than producing isolated outputs, co-creative systems make enactive concepts operationally unavoidable. Building such systems functions as a form of reflection-in-action: theoretical commitments shape system behavior, and system behavior in turn reveals the strengths and limits of those commitments.

For the IUI community, this perspective reframes mixed-initiative interaction not as a question of control allocation or optimization, but as one of regulating participation over time. Issues such as pacing, yielding, escalation, and withdrawal emerge as central interactional capacities rather than secondary design concerns. Evaluating intelligent interfaces therefore requires attention to trajectories of interaction, not just momentary success.

More broadly, this work suggests that progress in interactive AI depends on treating system building as an epistemic practice. Co-creative interfaces are not only applications of theory; they are instruments through which theories of cognition and interaction can be tested, refined, and rethought. Re-uniting enaction, HCI, and AI in this way is not a return to earlier paradigms, but a necessary step toward understanding intelligent systems that must remain coherent as interactions unfold.

Acknowledgments

The author thanks the conversational AI system ChatGPT (OpenAI) for its role as a sustained intellectual collaborator during the development of this work. Through iterative dialogue, critique, and co-construction of arguments, the system functioned as a reflective partner in theory development, helping to clarify claims, surface assumptions, and refine the structure of the paper. Importantly, all theoretical positions, interpretations, and final editorial decisions remain the responsibility of the author. This collaboration itself exemplifies the paper’s core argument: that building and interacting with artificial systems can serve as a productive site for theoretical reflection and epistemic development.

References

Agre, P. E. (1997). Computation and human experience. Cambridge University Press.

Agre, P. E., & Chapman, D. (1987). Pengi: An implementation of a theory of activity. Proceedings of the AAAI Conference on Artificial Intelligence.

Amershi, S., et al. (2014). Power to the people: The role of humans in interactive machine learning. AI Magazine, 35(4), 105–120.

Amershi, S., et al. (2019). Guidelines for human–AI interaction. Proceedings of the CHI Conference on Human Factors in Computing Systems, 1–13.

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv:1606.06565.

Ashby, W. R. (1956). An introduction to cybernetics. Chapman & Hall.

Beer, R. D. (2003). The dynamics of active categorical perception in an evolved model agent. Adaptive Behavior, 11(4), 209–243.

Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. Proceedings of the ACL.

Bender, E. M., et al. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of FAccT.

Bengio, Y., et al. (2021). Deep learning for AI. Communications of the ACM, 64(7), 58–65.

Benjamin, R. (2019). Race after technology. Polity.

Bommasani, R., et al. (2021). On the opportunities and risks of foundation models. arXiv:2108.07258.

Bown, O., Eldridge, A., & McCormack, J. (2009). Understanding interaction in contemporary digital music: From instruments to behavioural objects. Organised Sound, 14(2), 188–196.

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

Brown, T. B., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS).

Colombetti, G. (2014). The feeling body: Affective science meets the enactive mind. MIT Press.

Davis, N., Hsiao, C.-P., Popova, Y., & Magerko, B. (2015). An enactive model of creativity for computational collaboration and co-creation. Proceedings of the International Conference on Computational Creativity.

Davis, N., et al. (2017). Creativity support for computational creativity. Proceedings of the International Conference on Computational Creativity.

Davis, N., et al. (2021). Human–AI co-creativity: Designing for interactional emergence. Creativity and Cognition.

Davis, N., Hsiao, C.-P., Singh, K., & Magerko, B. (2021). Creative sensemaking: A framework for human–AI co-creation. Proceedings of the CHI Conference on Human Factors in Computing Systems.

De Jaegher, H., & Di Paolo, E. A. (2007). Participatory sense-making. Phenomenology and the Cognitive Sciences, 6(4), 485–507.

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

Di Paolo, E. A., & De Jaegher, H. (2012). The interactive brain hypothesis. Frontiers in Human Neuroscience, 6, 163.

Di Paolo, E. A., & Thompson, E. (2014). The enactive approach. Topoi, 33(1), 1–12.

Di Paolo, E. A., Buhrmann, T., & Barandiaran, X. (2017). Sensorimotor life: An enactive proposal. Oxford University Press.

Di Paolo, E. A., Cuffari, E. C., & De Jaegher, H. (2017). Linguistic bodies: The continuity between life and language. MIT Press.

Dietterich, T. G. (2019). Robust artificial intelligence. AI Magazine, 40(1), 5–24.

Dourish, P. (2004). Where the action is: The foundations of embodied interaction. MIT Press.

Dreyfus, H. L. (2007). Why Heideggerian AI failed and how fixing it would require making it more Heideggerian. Artificial Intelligence, 171(18), 1137–1160.

Fischer, G., Giaccardi, E., Eden, H., Sugimoto, M., & Ye, Y. (2006). Beyond binary choices: Integrating individual and social creativity. International Journal of Human–Computer Studies, 63(4–5), 482–512.

Fischer, G., et al. (2014). Creativity and human–computer interaction. ACM Transactions on Computer-Human Interaction, 21(2).

Floridi, L., et al. (2018). AI4People—An ethical framework for a good AI society: Opportunities, risks, principles, and recommendations. Minds and Machines, 28(4), 689–707.

Froese, T., & Ziemke, T. (2009). Enactive artificial intelligence: Investigating the systemic organization of life and mind. Artificial Intelligence, 173(3–4), 466–500.

Gallagher, S. (2017). Enactivist interventions: Rethinking the mind. Oxford University Press.

Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4).

Gaver, W. (2012). What should we expect from research through design? Proceedings of the CHI Conference on Human Factors in Computing Systems.

Höök, K. (2018). Designing with the body. MIT Press.

Höök, K., et al. (2016). Embracing first-person perspectives in HCI. Human–Computer Interaction, 31(1).

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

Hutto, D. D., & Myin, E. (2017). Evolving enactivism: Basic minds meet content. MIT Press.

Jordan, B., & Henderson, A. (1995). Interaction analysis: Foundations and practice. Journal of the Learning Sciences, 4(1), 39–103.

Jordanous, A., & Keller, B. (2016). Modelling creativity: Identifying key components through a corpus-based approach. PLOS ONE, 11(10).

Kantosalo, A., & Toivonen, H. (2016). Modes of human–computer co-creativity. In Computational Creativity Research.

Kaptelinin, V., & Nardi, B. (2012). Activity theory in HCI: Fundamentals and reflections. MIT Press.

Kelso, J. A. S. (1995). Dynamic patterns: The self-organization of brain and behavior. MIT Press.

Kirsh, D. (2013). Embodied cognition and the magical future of interaction design. ACM Transactions on Computer-Human Interaction, 20(1).

Kuhn, T. S. (1962). The structure of scientific revolutions. University of Chicago Press.

Lake, B. M., et al. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40.

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

Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80.

Lipton, Z. C., & Steinhardt, J. (2019). Troubling trends in machine learning scholarship. Communications of the ACM, 62(10), 45–53.

Marr, D. (1982). Vision. W. H. Freeman.

Marcus, G. (2018). Deep learning: A critical appraisal. arXiv:1801.00631.

Marcus, G., & Davis, E. (2019). Rebooting AI: Building artificial intelligence we can trust. Pantheon.

Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and cognition: The realization of the living. D. Reidel.

Merleau-Ponty, M. (2012). Phenomenology of perception (D. A. Landes, Trans.). Routledge.

Mitchell, M., et al. (2021). Model cards for model reporting. Proceedings of FAccT.

Newell, A. (1990). Unified theories of cognition. Harvard University Press.

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

Newen, A., De Bruin, L., & Gallagher, S. (Eds.). (2018). The Oxford handbook of 4E cognition. Oxford University Press.

Norman, D. A. (2013). The design of everyday things (Revised ed.). Basic Books.

Ovadia, Y., et al. (2019). Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift. Advances in Neural Information Processing Systems (NeurIPS).

Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active inference. MIT Press.

Quinonero-Candela, J., et al. (2009). Dataset shift in machine learning. MIT Press.

Rahimi, A., et al. (2019). Test of time award: Machine learning has become alchemy. NeurIPS.

Rietveld, E., Denys, D., & Van Westen, M. (2018). Ecological-enactive cognition as engaging with a field of relevant affordances. Phenomenology and the Cognitive Sciences, 17(1), 1–24.

Rogers, Y. (2012). HCI theory: Classical, modern, and contemporary. Synthesis Lectures on Human-Centered Informatics.

Rogers, Y., Sharp, H., & Preece, J. (2011). Interaction design. Wiley.

Russell, S. (2019). Human compatible. Viking.

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

Sawyer, R. K. (2012). Explaining creativity: The science of human innovation (2nd ed.). Oxford University Press.

Schön, D. A. (1983). The reflective practitioner. Basic Books.

Sculley, D., et al. (2015). Hidden technical debt in machine learning systems. Advances in Neural Information Processing Systems, 28.

Sculley, D., et al. (2018). Winner’s curse? On pace, progress, and empirical rigor. ICLR Workshop.

Shneiderman, B. (2020). Human-centered AI. International Journal of Human–Computer Interaction, 36(6).

Simon, H. A. (1969). The sciences of the artificial. MIT Press.

Suchman, L. (1987). Plans and situated actions. Cambridge University Press.

Suchman, L. (2007). Human–machine reconfigurations. Cambridge University Press.

Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning (2nd ed.). MIT Press.

Thompson, E. (2007). Mind in life. Harvard University Press.

Varela, F. J. (1979). Principles of biological autonomy. North-Holland.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind. MIT Press.

Weber, A., & Varela, F. J. (2002). Life after Kant: Natural purposes and the autopoietic foundations of biological individuality. Phenomenology and the Cognitive Sciences, 1(2), 97–125.

Wiener, N. (1948). Cybernetics. MIT Press.

Winograd, T., & Flores, F. (1986). Understanding computers and cognition. Ablex.

Yannakakis, G. N., Liapis, A., & Alexopoulos, C. (2014). Mixed-initiative co-creativity. Proceedings of the Foundations of Digital Games (FDG).

—-

Bødker, S. (2015). Third-wave HCI, 10 years later—participation and sharing. Interactions, 22(5), 24–31. https://doi.org/10.1145/2804405

Kuutti, K., Bannon, L. J., Byrne, D., Cocciolo, A., McDonald, D. W., & Stewart, J. (2021). Where is the human? Bridging the gap between AI and HCI. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM. https://doi.org/10.1145/3411764.3445486

Lim, Y.-K., Stolterman, E., Jung, H., & Donaldson, J. (2019). Interaction gestalt and the design of aesthetic interactions. ACM Transactions on Computer-Human Interaction, 26(1), Article 4. https://doi.org/10.1145/3290605

Page updated

Google Sites

Report abuse