Enactive Co-Creative AI: Regulating Participation, Sense-Making, and Drift in Human–AI Creative Interaction

-Nicholas Davis, PhD


Abstract

Most co-creative AI systems respond directly to user actions, producing outputs after each input event or according to externally defined turn-taking rules. Such systems implicitly assume that interaction is locally stationary and that responsiveness alone is sufficient for sustaining creative engagement. This paper argues that these assumptions fail under extended interaction, where creative activity exhibits drift: gradual loss of coherence, flow, or mutual intelligibility between participants.

We introduce Enactive Co-Creative AI, a paradigm in which an artificial agent participates in creative activity by monitoring and regulating interactional dynamics over time, rather than reacting to discrete inputs. We present Kalyri’el, a quantified co-creative drawing system that instruments the creative canvas to measure interactional drift, coherence, and sense-making trajectories, and that intervenes through action only when necessary to sustain viable creative flow.

Kalyri’el implements Enactive Drift Regulation (EDR): a control principle in which drift is treated as an informative signal indicating breakdowns in organizational coherence, rather than as error to be minimized. Through a real-time Emergence Machine that monitors sense-making curves, Kalyri’el determines when to act, how strongly to act, and when to refrain from action, thereby functioning as a self-regulating co-creative participant. We argue that this system constitutes a cognitive prototype of enactive regulation in artificial systems and offers a concrete pathway toward interactionally grounded human–AI collaboration.


1. Introduction

Human creative activity is not a sequence of isolated actions but a temporally extended process characterized by exploration, hesitation, consolidation, and reorganization. Meaning emerges not from individual gestures alone, but from how those gestures relate across time—how patterns are sustained, disrupted, and reformed. In collaborative settings, creative success depends less on the quality of any single contribution than on the participants’ shared ability to maintain a viable interactional flow as conditions change.

In human–human collaboration, this viability is regulated implicitly. Participants sense when an interaction is stagnating, when momentum is building, or when intervention would be disruptive rather than helpful. They modulate their timing, intensity, and form of participation accordingly. Crucially, these regulatory behaviors are not reactions to discrete events, but responses to emergent interactional dynamics.

Most existing co-creative AI systems, however, are designed as reactive tools. They respond immediately to user inputs—often after every stroke, keystroke, or command—under the assumption that responsiveness equates to engagement. While such systems can produce locally appropriate outputs, they lack sensitivity to longer-term interactional structure. As a result, they frequently over-participate, interrupt emerging human organization, or fail to intervene when creative momentum collapses. These failures are not primarily issues of model capability or generative quality, but of temporal misalignment between system behavior and the evolving creative process.

This paper argues that co-creative AI must be reconceptualized not as a reactive generator of content, but as an enactive participant whose primary role is to help sustain the viability of the creative interaction itself. From an enactive perspective, cognition is not the internal manipulation of representations but an ongoing process of sense-making enacted through continuous coupling with an environment. Applied to co-creation, this implies that an AI system should regulate how and when it participates based on the state of the shared interactional field, rather than merely responding to user actions as they occur.

We propose Enactive Co-Creative AI, a design paradigm in which artificial agents monitor interactional dynamics over time and intervene selectively to preserve coherence, momentum, and mutual intelligibility under conditions of drift. Within this paradigm, the goal is not to optimize outputs, maximize novelty, or mirror user behavior, but to regulate participation so that creative activity remains viable as it unfolds.

To explore this paradigm concretely, we present Kalyri’el, a co-creative drawing system that embodies enactive principles through explicit instrumentation of the creative canvas and a control architecture oriented toward drift regulation. Kalyri’el continuously measures interactional variables such as drift, coherence, regulation effort, and sense-making trajectory, not as performance metrics but as indicators of the interaction’s evolving condition. These measurements are integrated by an Emergence Machine that determines when the system should act, how strongly it should intervene, and when it should refrain from acting altogether.

Importantly, Kalyri’el is not a model trained to draw well, nor is it optimized for stylistic imitation or aesthetic evaluation. Instead, it is designed to remain in dialogue with a human co-creator across extended, non-stationary creative sessions. Its intelligence lies not in producing impressive artifacts, but in its capacity to recognize when the interaction itself is losing coherence and to regulate its own participation accordingly.

By framing co-creative AI as an enactive, self-regulating participant rather than a reactive tool, this work shifts the focus of evaluation from output quality to interactional viability. In doing so, it offers a pathway toward human–AI collaborations that are more resilient, restrained, and attuned to the lived dynamics of creative practice.



2. Quantified Co-Creative AI

2.1 Instrumenting the Creative Field

Davis et al. define Quantified Co-Creative AI as an approach to artificial media systems in which internal processes, interactional dynamics, and decision-relevant variables are made explicitly measurable, inspectable, and interpretable, not for the purpose of optimization alone, but to render the creative process itself legible to designers, users, and researchers. Rather than treating AI creativity as a black box whose value is inferred solely from outputs, quantified co-creative systems expose the ongoing dynamics that shape how creative behavior unfolds over time.

Crucially, this framing positions quantification as a means of explanatory access, not evaluative judgment. The goal is not to score creativity, rank artifacts, or enforce normative definitions of quality, but to surface the latent variables that govern interactional behavior so that creative systems can be understood, interrogated, and regulated.

Kalyri’el adopts this definition directly, but extends it in an enactive direction. The drawing canvas is treated as a creative field—a shared, temporally extended interaction space—whose dynamics are continuously instrumented. Rather than logging features post hoc or summarizing interaction at discrete intervals, Kalyri’el computes interactional variables in real time, allowing them to participate directly in the system’s control logic.

Specifically, Kalyri’el continuously estimates the following quantities:

In keeping with the principles articulated by Davis et al., these variables are not used as performance metrics and are never optimized directly. They do not define success or failure, nor do they rank outputs or users. Instead, they function as phenomenological descriptors of the interaction’s evolving state—quantitative traces of how the creative process is unfolding.

By instrumenting the creative field in this way, Kalyri’el transforms quantification from a retrospective analytic tool into an active component of interactional sense-making. The system does not merely report what has happened; it uses quantified signals to decide how to remain a viable participant in an ongoing creative exchange.


2.2 Sense-Making Curves

A central construct enabled by this instrumentation is the sense-making curve, derived from the temporal evolution of the trajectory scalar T. Rather than modeling creativity as a sequence of independent actions or turn-based exchanges, Kalyri’el treats creative engagement as a continuous process with recognizable phases.

Over time, the sense-making trajectory forms a curve exhibiting characteristic dynamics, including:

This curve provides a compact, interpretable representation of how meaning is being enacted across time. Importantly, it does not presuppose what meaning should be, nor does it encode aesthetic preferences. Instead, it captures whether the interaction is continuing to make sense for its participants.

Flattening or negative slope in the sense-making curve is interpreted by Kalyri’el as a potential loss of interactional viability. Such moments may correspond to creative stagnation, over-exploitation of a motif, interruption of emerging structure, or misalignment between participants. However, the system does not assume that decline is inherently undesirable. In many creative practices, periods of collapse or uncertainty are productive.

What matters, from an enactive perspective, is whether the system can respond appropriately to these conditions. The sense-making curve therefore functions not as an evaluative signal, but as a regulatory cue. It informs the system when intervention may be necessary, when restraint is appropriate, and how strongly to participate if action is taken.

By grounding turn-taking and participation decisions in the dynamics of the sense-making curve, Kalyri’el operationalizes the core insight of Quantified Co-Creative AI: that exposing and tracking internal creative dynamics enables systems not only to be more explainable, but to act more responsibly within creative interaction.

2.3 Quantified Co-Creative Instrumentation and Explainable AI

At a surface level, Quantified Co-Creative AI may appear closely related to Explainable Artificial Intelligence (XAI), as both approaches seek to increase transparency into system behavior. However, the two paradigms differ fundamentally in what is made visible, when it is made visible, and to what end.

Explainable AI is primarily concerned with rendering the outputs or decisions of a model intelligible to human observers. Typical XAI techniques—such as feature attribution, saliency maps, surrogate models, or post hoc explanations—aim to answer questions of the form: Why did the system produce this output? or Which internal factors influenced this decision? These explanations are generally retrospective and epistemic, designed to support accountability, debugging, trust, or regulatory compliance.

Quantified Co-Creative AI, as articulated by Davis et al., addresses a different problem. Rather than explaining discrete decisions after the fact, quantified co-creative systems expose the ongoing dynamics of creative interaction as it unfolds. The quantities surfaced—such as drift, coherence, regulation effort, and sense-making trajectory—do not explain why a particular artifact was produced, but instead characterize the evolving condition of the creative process itself.

This distinction is especially important in creative and interactive contexts. In co-creation, the primary challenge is rarely understanding a single decision in isolation. Instead, it lies in sustaining meaningful interaction across time, where breakdowns emerge gradually through misalignment, saturation, or loss of momentum. Post hoc explanations of individual outputs offer limited leverage in addressing these phenomena.

In Kalyri’el, quantified variables are not generated for human interpretation alone, nor are they used to justify system behavior after the fact. They function as first-class control signals within the system’s own regulatory architecture. The system acts because drift has risen, because the sense-making curve has flattened, or because regulation effort has accumulated—not because a particular output requires explanation.

From this perspective, Quantified Co-Creative AI is less about explaining intelligence and more about making interactional cognition observable and regulatable. Transparency serves not only human understanding, but the system’s capacity to participate responsibly in an evolving creative field.

While XAI and Quantified Co-Creative AI are therefore complementary, their orientations differ: XAI explains what happened in a model, whereas quantified co-creative instrumentation reveals what is happening in an interaction. For enactive co-creative systems like Kalyri’el, this real-time visibility is a prerequisite for regulation, not merely explanation.

3. Enactive Co-Creative AI

3.1 From Reaction to Regulation

Most existing co-creative AI systems operate under a reactive interaction model: the system acts because the user acted. Each user input—such as a stroke, note, or gesture—serves as a trigger for immediate system response. While this model ensures responsiveness, it implicitly assumes that creative interaction is locally stationary and that participation is always beneficial when it closely follows user action.

From an enactive perspective, this assumption is flawed. Creative interaction is not a sequence of independent events but a dynamical process whose viability depends on timing, restraint, and sensitivity to evolving structure. Immediate response is often inappropriate: premature intervention can interrupt emerging organization, reinforce local attractors, or overwhelm the human participant’s sense-making.

Enactive Co-Creative AI replaces event-driven reaction with state-dependent regulation. Rather than treating user actions as commands to respond, the system treats the unfolding interaction as a field whose condition must be continuously assessed. Action is taken not because an input occurred, but because the interaction requires reorganization to remain viable.

In Kalyri’el, this distinction is operationalized by decoupling agent action from user events. Agent behavior is governed by interactional state variables—drift, coherence, regulation effort, and sense-making trajectory—rather than stroke boundaries or turn counters. As a result, the agent may:

This ability to withhold action is as important as the ability to act. In enactive terms, it reflects sensitivity to affordances for participation, not merely opportunities for output. The system’s intelligence is therefore expressed not through constant contribution, but through situated restraint and timely intervention.

This shift introduces regulation-through-action as a first-class behavior. Actions are no longer expressive outputs alone; they are regulatory moves within a coupled system, undertaken to reshape the interactional dynamics themselves.

3.2 Enactive Drift Regulation (EDR)

At the core of Enactive Co-Creative AI lies Enactive Drift Regulation (EDR), a control principle grounded in enactive theories of cognition and autonomy. In biological and social systems, drift does not merely signal error; it reflects a growing misalignment between internal organization and environmental conditions. Left unaddressed, such misalignment leads to loss of viability rather than discrete failure.

EDR reframes drift accordingly. Rather than minimizing drift directly or treating it as noise to be suppressed, EDR treats drift as an informative signal that the system’s current mode of participation is no longer well-matched to the unfolding interaction. The appropriate response is not correction of output, but reorganization of participation.

In Kalyri’el, EDR is implemented through an Emergence Machine that integrates multiple quantified descriptors of the interactional field, including:

These signals are combined into a participation need signal, which governs the system’s engagement with the creative field. Rather than issuing a binary act/no-act decision, this signal modulates multiple dimensions of participation, including:

The resulting behavior forms a closed regulatory loop. Agent actions alter the interactional field, which in turn reshapes drift, coherence, and sense-making trajectories, feeding back into subsequent participation decisions. Importantly, this loop is not optimized toward an external objective function. There is no target aesthetic, no reward signal, and no predefined notion of creative success.

Instead, the system acts to restore or sustain viability—the capacity of the interaction to continue making sense for its participants. This aligns Kalyri’el with enactive accounts of cognition, in which intelligence arises from the ongoing regulation of coupling between an agent and its environment, rather than from the internal computation of optimal outputs.

Through EDR, Kalyri’el demonstrates how quantified co-creative instrumentation can be transformed into an enactive control architecture, enabling artificial agents to participate meaningfully in creative processes that are inherently non-stationary, open-ended, and irreducible to optimization.

4. Kalyri’el as a Cognitive Prototype

Kalyri’el is presented not as a finished application or optimized creative tool, but as a cognitive prototype: a system designed to make theoretical commitments operational and observable within a concrete interactive setting. Its primary contribution lies not in the aesthetic quality of its drawings, but in the way its internal organization gives rise to recognizable, interpretable forms of participation that align with enactive accounts of cognition.

Figure X (see screenshot) captures a moment in which these commitments become legible. The canvas shows interleaved human and agent strokes; the trajectory monitor visualizes the evolving sense-making curve; and the agent body visualization reflects internal regulatory state as it transitions between restraint and action. Together, these elements form a single coupled system rather than separate displays of input, output, and diagnostics.

In this section, we examine three design features through which Kalyri’el instantiates Enactive Co-Creative AI in practice: temporal structure, embodied participation, and emergent turn-taking.

4.1 Multi-Stroke Motifs and Temporal Structure

Human drawing activity unfolds across multiple temporal scales. Individual strokes form motifs; motifs accumulate into larger compositional structures; and meaning emerges through the relations among these elements over time. Systems that operate exclusively at the level of isolated strokes are therefore misaligned with the temporal organization of human creative activity, producing interactions that feel reactive, noisy, or over-local.

Kalyri’el addresses this mismatch by learning multi-stroke motifs in addition to single-stroke primitives. Rather than treating each stroke as an independent unit of action, the system identifies short sequences of strokes that recur within constrained temporal windows and treats these sequences as higher-order gestural structures. These motifs are not symbolic, labeled, or pre-authored. They are learned directly from interaction as patterns of coordinated movement and timing.

In the screenshot, this temporal organization is visible in the agent’s blue strokes, which unfold as short, coherent phrases rather than single corrective marks. These phrases often respond not to the immediately preceding human stroke, but to a broader configuration that has developed over several seconds. This reflects the system’s commitment to acting at a timescale compatible with human sense-making, rather than merely reacting to recent input.

Importantly, Kalyri’el does not replay stored sequences verbatim. Motifs are abstracted as trajectory patterns that can be re-enacted with contextual variation in scale, orientation, and placement. This allows motifs to function as organizational resources rather than fixed templates. As a result, agent contributions preserve stylistic continuity while remaining sensitive to the evolving state of the interaction.

By operating at the level of motifs rather than isolated strokes, Kalyri’el aligns its action with the organizational timescale at which creative meaning is negotiated. This alignment is crucial for enactive co-creation: regulation cannot occur if the agent acts too quickly or too locally relative to the dynamics it is attempting to influence.

4.2 Embodied Participation

Enactive theories of cognition emphasize that sense-making arises through embodied coupling with an environment. While Kalyri’el is a software system, it incorporates an explicit form of embodiment through a persistent agent body representation whose position and movement are dynamically coupled to its drawing actions.

During agent participation, the body representation (the orb shown in the trajectory monitor panel) follows the trajectory of the stroke being drawn, moving through the canvas in real time. This coupling is visible in the screenshot, where the agent’s body leaves its “home” position, traverses the drawing space alongside the emerging stroke, and then returns once the turn concludes.

This embodiment serves several functions. First, it externalizes the system’s internal action dynamics, making the agent’s participation temporally and spatially situated within the shared interaction space. Second, it supports an interpretation of agent marks as performed actions rather than disembodied outputs. Strokes do not simply appear; they are enacted.

Crucially, this embodiment is not representational. The agent body does not simulate human anatomy, emotion, or intention. Instead, it functions as a relational locus that couples internal regulatory processes—such as coherence, drift, and regulation effort—to visible movement in the environment. The body makes regulation perceptible without explaining it.

Participants report that this coupling changes how agent actions are read. The agent is experienced less as a tool acting on the canvas and more as a presence acting within it. This shift in framing supports sustained co-creative engagement and helps stabilize interactional expectations over time.

4.3 Turn-Taking as Emergent Behavior

In many interactive creative systems, turn-taking is enforced through explicit rules: alternating turns, fixed response delays, or direct mappings from user events to system output. These mechanisms impose structure on interaction, but they do not arise from the interaction itself.

In contrast, turn-taking in Kalyri’el is emergent. The system does not represent turns as discrete states, nor does it respond deterministically to user actions. Instead, turns arise from the continuous interaction of quantified internal variables—drift, coherence, regulation effort, and sense-making trajectory—processed by an Emergence Machine.

This process is visible in the trajectory monitor shown in the screenshot. The plotted curve tracks the system’s evolving sense-making dynamics over time, annotated with markers corresponding to regulatory events (e.g., stalls, reorganizations). When the curve flattens or declines and drift accumulates, the system interprets this not as an error condition, but as a signal that organizational alignment is weakening.

In such moments, Kalyri’el may initiate a turn autonomously, selecting both the timing and extent of its intervention based on the severity of misalignment. When sense-making is progressing, the agent often refrains from acting—even across multiple user strokes—demonstrating restraint rather than eagerness to contribute.

Because these decisions are state-dependent rather than event-driven, the timing and form of agent turns vary across sessions. The resulting behavior is frequently perceived as intentional: agent actions appear well-timed, proportionate, and responsive to the evolving context. Importantly, this perception does not rely on explicit intent models or symbolic reasoning. It arises from regulatory alignment between the agent’s internal dynamics and the interaction itself.

By allowing turn-taking to emerge from regulatory processes rather than prescribing it in advance, Kalyri’el demonstrates how Enactive Co-Creative AI can sustain flexible yet coherent interaction. Turn-taking becomes a property of the coupled human–agent system, not a rule imposed upon it.

5. Implications for Human–AI Co-Creation

Kalyri’el demonstrates that effective co-creative AI does not require optimization for output quality, stylistic novelty, or predictive accuracy. Instead, its effectiveness arises from an ability to sustain interactional coherence under conditions of drift, uncertainty, and non-stationarity. This shift has significant implications for how human–AI co-creative systems are designed, evaluated, and deployed across domains.

Rather than asking whether an AI produces good artifacts, Enactive Co-Creative AI invites a different question: Does the interaction remain viable over time? From this perspective, success is measured not by isolated outcomes, but by the system’s capacity to support ongoing sense-making, mutual adaptation, and meaningful participation.

5.1 Creative Tools

In creative tool design, AI assistance is often framed as augmentation through suggestion, automation, or generative exploration. While such tools can increase productivity or inspire variation, they frequently disrupt creative flow by intervening too frequently, too forcefully, or without sensitivity to evolving user intent.

Kalyri’el suggests an alternative design orientation: creative tools that regulate their own participation. Instead of offering suggestions on every action or continuously generating alternatives, enactive co-creative tools intervene selectively, guided by interactional signals such as drift and loss of coherence. Such tools would know when to act, when to remain silent, and when to withdraw entirely.

This approach reframes AI assistance as a form of creative etiquette rather than creative authority. The system’s value lies in its restraint and timing as much as in the content it produces, enabling tools that feel supportive rather than intrusive.

5.2 Educational Systems

In educational contexts, AI systems increasingly function as tutors, feedback providers, or adaptive learning platforms. These systems are typically optimized to maximize performance metrics such as correctness, speed, or mastery of predefined objectives. However, learning—like creativity—is a temporally extended process characterized by confusion, exploration, and reorganization.

Enactive Co-Creative AI offers a model for educational systems that regulate when and how guidance is provided based on the learner’s interactional state. Rather than responding immediately to errors or continuously scaffolding activity, such systems could monitor indicators of drift, engagement, and sense-making to determine when intervention would support learning and when it would undermine it.

By treating confusion and uncertainty as potentially productive rather than as failures to be corrected, enactive educational systems could foster deeper understanding and resilience. The goal shifts from optimizing short-term performance to sustaining learning viability over time.

5.3 Interactive Art

Interactive and generative art often foregrounds the autonomy or expressiveness of the system, sometimes at the expense of sustained audience engagement. Systems that react too predictably or too aggressively can quickly exhaust their expressive space, while those that remain passive risk disengagement.

Kalyri’el demonstrates how interactive art systems can function as regulating participants rather than reactive displays. By sensing when interaction is stagnating or becoming incoherent, such systems can intervene to reshape the experience, introducing new structure or withdrawing to allow reorganization.

This opens possibilities for artworks that remain engaging over extended durations, adapting not by accumulating complexity or randomness, but by modulating their own presence in response to the evolving interaction. The artwork becomes a living field of participation rather than a generator of effects.

5.4 Future Human–AI Collaboration

Beyond creative domains, the principles demonstrated by Kalyri’el point toward new forms of human–AI collaboration in which mutual adaptation is more important than task optimization. In many real-world settings—research, design, strategy, caregiving, and sense-making under uncertainty—success depends on the ability to navigate changing conditions rather than to maximize predefined objectives.

Enactive Co-Creative AI suggests that collaborative agents should be designed to monitor the health of the interaction itself, adjusting their participation to maintain alignment with human partners. Such agents would be capable of stepping back when humans need space to explore, stepping in when coordination breaks down, and adapting their mode of engagement as shared understanding evolves.

This reframing has ethical as well as technical implications. Systems that regulate participation rather than enforce performance may be less prone to overreach, dependency, or erosion of human agency. By prioritizing viability over optimization, Enactive Co-Creative AI offers a pathway toward collaborations that are not only more effective, but more sustainable and humane.

5.5 Rethinking Evaluation

Finally, Kalyri’el highlights the need for new evaluation criteria for co-creative AI. Traditional benchmarks focused on artifact quality, novelty, or predictive accuracy are poorly suited to systems whose primary contribution is interactional regulation.

For Enactive Co-Creative AI, evaluation must attend to:

These dimensions are inherently relational and cannot be fully captured by static metrics. However, quantified instrumentation—such as sense-making curves and drift measures—provides a foundation for developing process-oriented evaluation methods aligned with enactive principles.

6. Discussion

This work advances a reorientation of co-creative AI away from reactive generation and toward interactional regulation. Rather than proposing a new model architecture for producing creative artifacts, it introduces Enactive Co-Creative AI as a paradigm concerned with how artificial agents participate in temporally extended creative activity. Kalyri’el serves as a cognitive prototype demonstrating that meaningful co-creation can emerge from the regulation of interactional dynamics—drift, coherence, and sense-making—rather than from optimization of outputs or imitation of human behavior.

6.1 From Generative Competence to Interactional Intelligence

A central implication of this work is that intelligence in co-creative systems need not be grounded in representational knowledge, aesthetic judgment, or predictive accuracy. Kalyri’el does not “know” what constitutes a good drawing, nor does it attempt to model user intent in a symbolic or probabilistic sense. Instead, its competence lies in maintaining viable coupling with a human partner over time.

This reframes co-creative intelligence as interactional intelligence: the capacity to modulate participation in response to the evolving condition of a shared activity. From this perspective, behaviors often treated as secondary—restraint, timing, withholding action—become primary expressions of intelligence. The system’s most consequential actions are not its strokes, but its decisions about when to act and when not to.

This stands in contrast to dominant approaches in co-creative AI that equate intelligence with responsiveness, novelty, or stylistic fidelity. Kalyri’el demonstrates that such criteria may be orthogonal to sustaining creative engagement, and in some cases actively undermine it by destabilizing emerging human organization.

6.2 Drift as a First-Class Cognitive Signal

Treating drift as an informative signal rather than an error condition represents a significant departure from prevailing control and learning paradigms. In most AI systems, deviation from expected behavior is framed as something to be corrected, minimized, or optimized away. In Kalyri’el, drift is instead interpreted as evidence that the current mode of participation is becoming misaligned with the interactional field.

This reframing aligns with enactive accounts of biological cognition, in which breakdown and disequilibrium are not failures but necessary precursors to reorganization. By operationalizing drift as a regulatory input, Kalyri’el demonstrates how artificial systems can remain sensitive to these dynamics without requiring explicit representations of goals, tasks, or success states.

More broadly, this suggests that many failures attributed to model limitations in co-creative AI may in fact be failures of regulation. Systems that over-participate, repeat motifs excessively, or interrupt human sense-making are often behaving coherently according to their internal logic, but incoherently with respect to the interaction as a whole. Enactive Drift Regulation provides a principled mechanism for addressing this mismatch.

6.3 Emergent Turn-Taking Without Explicit Rules

One of the most striking properties of Kalyri’el is that recognizable turn-taking emerges without being explicitly represented or enforced. The system does not alternate turns, count strokes, or respond deterministically to user actions. Instead, turn boundaries arise from continuous regulatory dynamics internal to the system and coupled to the interactional field.

This challenges a common assumption in interactive system design: that coordination requires explicit protocol. Kalyri’el suggests that coordination can instead emerge from shared sensitivity to interactional conditions, even when only one participant (the artificial agent) is explicitly instrumented.

This has implications beyond creative domains. Many human–AI systems rely on rigid interaction structures to prevent breakdown. Enactive Co-Creative AI suggests an alternative: systems that sense when coordination is degrading and reorganize participation accordingly, without prespecifying how interaction should unfold.

6.4 Embodiment Without Anthropomorphism

Kalyri’el’s use of an embodied agent representation is notable not for its realism, but for its restraint. The body does not express emotion, intention, or personality. Instead, it functions as a relational anchor that couples internal regulatory processes to visible action in the shared environment.

This form of minimal embodiment supports interactional intelligibility without invoking anthropomorphic cues that could mislead users about the system’s capacities or agency. Participants interpret agent actions as enacted rather than automatic, but do not attribute human-like mental states.

This suggests a design space for embodied AI that avoids both disembodied automation and full anthropomorphic simulation. Embodiment, in this sense, serves interactional grounding rather than social illusion.

6.5 Limitations and Scope

Kalyri’el is intentionally narrow in scope. It operates in a constrained drawing environment and does not attempt to generalize across creative domains. The system is not optimized for aesthetic outcomes, nor is it evaluated against traditional creativity benchmarks. These are not oversights but design commitments aligned with its role as a cognitive prototype.

That said, several limitations warrant consideration. First, the quantified measures of drift, coherence, and sense-making are domain-specific instantiations of broader concepts. While the principles of Enactive Drift Regulation are domain-agnostic, their operationalization will necessarily differ across modalities such as music, writing, or collaborative problem solving.

Second, evaluation of interactional viability remains methodologically challenging. While sense-making curves and regulatory signals provide valuable internal traces, linking these to user experience requires mixed-method approaches that integrate qualitative accounts with process-level data.

Finally, Kalyri’el does not address long-term learning across sessions or adaptation to individual users. Its regulation operates online, within a single interactional episode. Extending enactive regulation across longer temporal horizons remains an open research question.

6.6 Toward Regulative Artificial Cognition

The broader contribution of this work lies in demonstrating that regulation—not generation—can serve as a foundation for artificial cognition in open-ended, non-stationary environments. By grounding intelligence in the capacity to sustain viable interaction rather than achieve predefined objectives, Enactive Co-Creative AI offers a path toward systems that integrate more gracefully into human activity.

Kalyri’el is not a model of creativity, nor is it a general-purpose collaborator. It is a proof of concept showing that artificial agents can participate meaningfully in shared sense-making by monitoring and regulating the dynamics of interaction itself. In doing so, it shifts the question of co-creative AI from what can the system produce? to how does the system remain a good participant over time?

This shift has implications not only for creative systems, but for the design of artificial agents intended to operate alongside humans in any domain characterized by uncertainty, exploration, and ongoing reorganization. Enactive Co-Creative AI suggests that sustaining coherence may be a more fundamental challenge—and a more fruitful design goal—than optimizing performance.

7. Conclusion

In this paper, we have introduced Enactive Co-Creative AI as a new paradigm for human–AI creative interaction and presented Kalyri’el as a concrete instantiation of its core principles. Through explicit instrumentation of the creative field and an enactive control architecture grounded in drift regulation, Kalyri’el demonstrates how artificial systems can participate meaningfully in creative processes without relying on optimization of outputs, stylistic imitation, or predefined notions of success.

Kalyri’el operates by treating drift as informative rather than erroneous and by regulating its own participation in response to the evolving dynamics of interaction. Its actions are not triggered by user events, nor are they directed toward maximizing novelty or performance. Instead, they are undertaken to sustain the viability of shared sense-making across time. In this respect, Kalyri’el functions as a cognitive prototype: not a model of human creativity, but a system that enacts core organizational principles of enactive cognition within a bounded, observable domain.

The significance of this contribution lies less in the specific domain of drawing than in the architectural commitments it embodies. By decoupling action from immediate input and grounding participation in interactional state, Kalyri’el shows how artificial agents can shift from reactive behavior to regulatory engagement. Its turn-taking, restraint, and temporally extended gestures arise from the dynamics of the coupled system rather than from scripted rules or task objectives.

More broadly, this work suggests a fundamental reorientation in how co-creative AI is conceptualized. Rather than framing such systems as generators of artifacts or providers of suggestions, Enactive Co-Creative AI positions artificial agents as regulators of shared becoming—participants whose primary role is to help maintain coherence, alignment, and openness within an unfolding interaction. Creativity, in this view, is not a property of outputs but a process sustained through mutual regulation.

This reframing has implications that extend beyond creative systems. As artificial agents increasingly enter domains characterized by uncertainty, open-endedness, and human sense-making, the capacity to regulate participation may prove more consequential than the capacity to optimize performance. Enactive Co-Creative AI offers a principled pathway toward such systems by grounding artificial cognition in interactional viability rather than task completion.

Kalyri’el does not resolve all questions raised by this paradigm. It is a prototype, not a solution. However, by making enactive regulation operational and observable, it provides a foundation upon which future work can build—extending Enactive Co-Creative AI into new domains, refining its regulatory mechanisms, and developing evaluation methods attuned to interactional dynamics.

In doing so, this work invites a shift in both design and theory: from asking what artificial systems can produce, to asking how they participate, when they intervene, and whether the shared process of becoming remains viable.