Co-Creative Sense-Making: A Framework for Sustaining Meaningful Human–AI Co-Creation

Abstract

Human–AI co-creative systems have advanced rapidly, enabling artificial agents to generate novel and technically competent creative content across domains such as art, music, writing, and design. However, many such collaborations degrade over time, even when short-term outputs remain strong. Interactions lose coherence, direction, or vitality, leading to fatigue, disengagement, or a breakdown of creative flow. This paper argues that these failures are frequently mischaracterized as limitations of generation quality, alignment, or responsiveness. Instead, we propose that many breakdowns in human–AI co-creation are failures of shared sense-making over time.

We introduce Co-Creative Sense-Making as a framework for understanding human–AI collaboration as a diachronic, relational process. Within this framework, we identify interactional drift as a characteristic failure mode and argue that sustaining collaboration requires explicit regulatory capacities governing persistence, completion, pacing, yielding, and coupling across pauses. We further derive implications for system design and evaluation, emphasizing regulation of participation rather than optimization of outputs. By reframing co-creative AI around the challenge of sustaining meaning over time, this work provides a conceptual foundation for designing and evaluating human–AI collaborations that endure.

1. Introduction: The Problem of Sustaining Collaboration

Human–AI co-creative systems are increasingly present across creative domains, including visual art, music composition, writing, design, and improvisational performance. Advances in generative modeling have enabled artificial agents to produce content that is stylistically rich, technically competent, and often indistinguishable from human-authored material in short interactions. As a result, research on co-creative AI has expanded rapidly, with growing attention to how artificial systems can assist, augment, or partner with humans in creative activity.

Despite these advances, a recurring challenge has become increasingly visible across domains and systems: many human–AI creative collaborations degrade over time. Interactions that begin fluidly often lose coherence, direction, or vitality as the collaboration unfolds. Users report fatigue, loss of flow, diminished engagement, or a sense that the system is “no longer listening,” even when the system continues to generate novel and technically competent outputs. These breakdowns are not isolated anomalies but appear as a persistent pattern across diverse co-creative settings.

This paper argues that such failures are frequently mischaracterized. Rather than reflecting deficiencies in generation quality, expressive capacity, or responsiveness, many breakdowns in human–AI co-creation are better understood as failures of shared sense-making over time. Addressing this problem requires a conceptual reframing of what it means for humans and artificial systems to create together.

1.1 The Rise—and Fragility—of Human–AI Co-Creation

Human–AI co-creation has emerged as a central topic in computational creativity and human–computer interaction. Systems now support collaborative drawing, co-writing, musical improvisation, design exploration, and interactive storytelling, often through mixed-initiative or turn-based interaction models. Surveys and handbooks in the field, such as The Handbook of Computational Creativity, document a broad range of approaches aimed at enabling artificial agents to contribute meaningfully to creative processes.

Much of this work has focused on improving what AI systems generate. Evaluation criteria commonly emphasize output novelty, surprise, stylistic alignment, diversity, or usefulness. In HCI contexts, additional emphasis is placed on usability, controllability, and user satisfaction. Systems are frequently assessed based on short-term interaction quality, isolated tasks, or snapshot judgments of user experience.

These emphases have driven substantial progress. Modern co-creative systems can adapt stylistically to user input, respond quickly to prompts, and generate content that users find engaging—at least initially. However, as interactions extend in time, a different set of challenges emerges. Longitudinal studies and practitioner reports increasingly note that sustained collaboration is difficult to maintain. Users disengage, intervene manually to “reset” the interaction, or abandon the system altogether, despite continued technical competence on the system’s part.

Similar observations appear in the mixed-initiative systems literature, where early principles emphasized balancing control and automation. While such systems can successfully negotiate initiative at a local level, maintaining coherence across longer interactions remains difficult. Over time, systems may become overly dominant, excessively deferential, or locked into patterns that no longer align with the user’s evolving intentions. The result is often a subtle erosion of collaborative flow rather than an abrupt failure.

Crucially, these breakdowns often occur without any obvious drop in generation quality. The system continues to produce outputs that are novel, stylistically appropriate, or technically correct. What deteriorates instead is the sense that the human and the system are participating in a shared creative process with a common direction. This suggests that prevailing evaluation criteria are insufficient to explain—or prevent—many of the most consequential failures in co-creative interaction.

1.2 The Missing Conceptual Center

Existing frameworks for human–AI co-creation provide valuable tools for analyzing generation, interaction techniques, and user experience. However, they largely lack a theory of interaction as an unfolding, time-extended process. Most approaches implicitly treat co-creation as a sequence of discrete contributions: the human acts, the system responds, and the quality of each response is assessed independently or aggregated over short windows.

Within this framing, success is typically defined in terms of:

While important, these dimensions do not capture whether a collaboration remains meaningful as it unfolds. They offer little traction on phenomena such as:

Notably, the field lacks a shared vocabulary for describing these failures. Terms such as drift, loss of flow, or misalignment are used informally, but without a unifying conceptual framework that explains why such breakdowns arise or how they might be mitigated. As a result, failures of sustained collaboration are often treated as implementation issues, usability problems, or inevitable limitations of current models, rather than as symptoms of a deeper theoretical gap.

This gap becomes especially salient in creative contexts, where goals are fluid, meanings evolve, and productive collaboration depends on sensitivity to timing, relevance, and mutual direction rather than on task completion alone. In such settings, effective collaboration requires more than producing appropriate outputs—it requires maintaining a shared sense of what is happening, why it matters, and how to proceed together.

1.3 Co-Creative Sense-Making: A Framework Proposal

To address this gap, we propose Co-Creative Sense-Making as a unifying framework for understanding, designing, and evaluating human–AI co-creative systems.

We define co-creative sense-making as the ongoing, reciprocal regulation of meaning, direction, and participation between human and artificial agents as a creative interaction unfolds over time. From this perspective, co-creation is not a sequence of isolated contributions but a diachronic, relational process in which coherence must be actively sustained under changing conditions.

This framing shifts the focus of co-creative AI:

Importantly, the framework does not presuppose human-like understanding, shared representations, or mutual prediction. Instead, it treats sense-making as something that emerges between agents through interaction, and that can succeed or fail depending on how participation is regulated over time.

The contribution of this paper is not a new algorithm or system, but a conceptual reorientation. By foregrounding sense-making as the central challenge of human–AI co-creation, we provide a framework that:

The sections that follow elaborate this framework, introduce interactional drift as a core failure mode of co-creative sense-making, and identify key regulatory capacities required to sustain meaningful human–AI collaboration over time.

2. Background and Related Work: What Existing Frameworks Miss

Research on human–AI co-creation spans multiple traditions, including computational creativity, human–computer interaction, mixed-initiative systems, and adaptive intelligence. These bodies of work have produced valuable insights into how artificial systems can generate creative material, respond to user input, and support human activity. However, despite substantial progress, existing frameworks share a common limitation: they tend to conceptualize creativity and collaboration primarily in terms of outputs and local interactions, rather than as time-extended processes of shared sense-making. This section reviews relevant strands of prior work and identifies the conceptual gaps that motivate the present framework.

2.1 Co-Creativity in HCI and AI

Human–AI co-creativity has been a long-standing topic in computational creativity and HCI. Early work on creative systems focused on autonomous generation, evaluating systems based on novelty, value, or surprise. As interactive systems became more prevalent, attention shifted toward co-creative settings in which humans and artificial agents jointly contribute to creative artifacts. Comprehensive overviews of this space can be found in resources such as The Handbook of Computational Creativity, which documents a wide range of systems supporting collaborative drawing, music composition, storytelling, and design.

Within HCI, co-creative systems are often framed as tools or partners that augment human creativity. Research emphasizes interface design, interaction techniques, and user experience, frequently drawing on mixed-initiative principles to balance human control and system autonomy. In these contexts, systems are evaluated using criteria such as perceived creativity, usefulness, enjoyment, or the quality of generated artifacts. Quantitative creativity metrics—such as novelty, diversity, and surprise—are commonly employed, alongside qualitative user studies assessing satisfaction and engagement.

While these approaches have been productive, they implicitly frame creativity as a product-oriented phenomenon. Success is typically assessed by examining artifacts produced at particular moments, or by aggregating user judgments across relatively short interactions. Even when interaction is acknowledged as important, it is often treated as a means to an end: a way to elicit better outputs or improve user satisfaction.

This framing obscures a critical dimension of co-creative activity: the process by which meaning, direction, and relevance are negotiated over time. Creative collaboration is not merely the accumulation of contributions but an evolving process in which participants continually orient to what has been done, what matters now, and what should come next. Existing co-creativity frameworks provide limited conceptual tools for analyzing how this process unfolds or why it sometimes breaks down, particularly in longer interactions.

2.2 Responsiveness, Alignment, and Assistance Paradigms

A closely related strand of work conceptualizes human–AI interaction in terms of responsiveness, assistance, or alignment. In these paradigms, the artificial system is designed to respond appropriately to user input, anticipate needs, or optimize outcomes according to inferred user preferences. This framing is especially prominent in intelligent assistants, recommender systems, and adaptive interfaces, where the system’s role is to support the user efficiently and unobtrusively.

Such systems typically assume that user goals and intentions are sufficiently stable to be inferred, modeled, and optimized. Responsiveness is taken as a primary indicator of success: the faster and more accurately the system responds to user input, the better the interaction is presumed to be. Similarly, alignment is often understood as minimizing divergence between system behavior and user intent, whether through preference learning, reinforcement learning, or explicit feedback mechanisms.

While these assumptions are reasonable in many task-oriented settings, they become problematic in creative collaboration. Creative intent is often emergent, exploratory, and revisable, rather than fixed in advance. Participants may deliberately change direction, pursue ambiguity, or abandon earlier ideas in response to what emerges during the process. In such contexts, a system that is highly responsive to past input may inadvertently constrain exploration or reinforce patterns that are no longer relevant.

Two distinctions are particularly important here. First, responsiveness is not equivalent to collaboration. A system can respond quickly and accurately to user actions while still undermining the collaborative process—for example, by over-participating, anticipating too aggressively, or failing to recognize when restraint is appropriate. Second, alignment is not equivalent to shared meaning. Aligning system behavior with inferred user preferences does not guarantee that both parties maintain a coherent sense of the creative situation or a shared understanding of what is unfolding.

These limitations point to a deeper issue: responsiveness- and alignment-centered paradigms lack a theory of how meaning is co-constructed and sustained over time. They focus on matching system behavior to user input at particular moments, rather than on regulating participation in a way that preserves intelligibility, direction, and mutual engagement across an evolving interaction.

2.3 Predictive and Optimization-Centered Accounts

Predictive and optimization-centered frameworks have exerted a strong influence on contemporary models of intelligence, both in cognitive science and artificial intelligence. Approaches such as predictive processing and active inference model cognition as the minimization of prediction error or variational free energy through hierarchical generative models. Influential treatments of these ideas include Surfing Uncertainty and Active Inference.

In artificial systems, similar assumptions underlie reinforcement learning and reward-based optimization. Intelligence is framed as the capacity to select actions that maximize expected reward or minimize some loss function, often under conditions of uncertainty. Adaptation is achieved by updating internal models or policies in response to discrepancies between predicted and observed outcomes.

These frameworks offer powerful tools for perception, control, and learning, and they have been applied successfully in many domains. However, their applicability to co-creative interaction is limited in important ways. At a conceptual level, predictive and optimization-centered accounts presuppose that success can be defined in terms of error reduction relative to a model or objective. Even when objectives are learned or inferred, the underlying aim remains to converge toward some stable criterion of correctness or optimality.

Co-creative interaction, by contrast, often involves maintaining intelligibility under change, rather than minimizing deviation from a fixed target. In creative collaboration, shifts in direction are not necessarily errors to be corrected but meaningful developments to be taken up, resisted, or transformed. A system that treats unexpected changes as noise or error may fail precisely when adaptability is most needed.

Moreover, predictive accounts tend to privilege internal model accuracy, whereas co-creative sense-making depends on the relational dynamics between participants. What matters is not only whether a system predicts user behavior accurately, but whether its actions remain interpretable, timely, and relevant within the evolving interaction. These considerations point beyond optimization and prediction toward a different set of design concerns.

2.4 Enactive and Participatory Sense-Making

An alternative perspective on cognition and interaction is provided by enactive and participatory approaches, which emphasize the embodied, relational, and time-extended nature of sense-making. Foundational work in this tradition, such as The Embodied Mind, argues that cognition arises through active engagement with the world rather than through passive representation. From this view, meaning is enacted through perception–action cycles situated in concrete contexts.

Subsequent developments have extended these ideas to social and interactive domains. Enactive accounts emphasize that sense-making is not confined to individual agents but can emerge through interaction itself. In participatory sense-making, meaning is understood as something that arises between agents as they coordinate, respond, and adapt to one another over time. Key articulations of this view can be found in works such as Enacting Meaning and Participatory Sense-Making.

These perspectives offer conceptual resources for understanding co-creative interaction as a relational process rather than a sequence of isolated contributions. They foreground issues of timing, coordination, breakdown, and recovery—phenomena that are central to creative collaboration but under-theorized in mainstream AI and HCI frameworks.

At the same time, enactive theories have rarely been translated into concrete design or evaluation frameworks for human–AI systems. As a result, their insights often remain disconnected from practical concerns in co-creative AI. The present work draws on enactive and participatory ideas as a conceptual anchor, not as a comprehensive theoretical commitment. Our aim is not to import enactivism wholesale, but to articulate a framework that makes sense-making—and its breakdowns—explicitly central to human–AI co-creation.


Across diverse strands of prior work, human–AI co-creation has been framed primarily in terms of generation quality, responsiveness, alignment, or optimization. While these perspectives have enabled substantial technical progress, they leave unaddressed a core challenge: how to sustain shared meaning, direction, and engagement over time. Enactive and participatory approaches point toward an alternative understanding of interaction as an ongoing sense-making process, but have yet to be fully integrated into co-creative AI research. The next section builds on these insights to define Co-Creative Sense-Making as a unifying framework for addressing this gap.

3. What Is Co-Creative Sense-Making?

The central claim of this paper is that many successes and failures in human–AI co-creation are best understood not in terms of generation quality, alignment, or responsiveness, but in terms of how meaning is sustained, negotiated, and transformed over time. To make this claim precise, this section defines Co-Creative Sense-Making, delineates what the concept explicitly excludes, and articulates its core properties. The goal is not to introduce a metaphor, but to establish a rigorous framework that can guide analysis, design, and evaluation of co-creative human–AI systems.

3.1 Definition

We define Co-Creative Sense-Making as follows:

Co-Creative Sense-Making is the ongoing, reciprocal regulation of meaning, direction, and participation between human and artificial agents as a creative interaction unfolds over time.

Several aspects of this definition warrant clarification.

First, sense-making is treated as an activity, not a state. It refers to the continuous process by which actions, contributions, and silences become intelligible within an unfolding interaction. In co-creative contexts, meaning is not fixed in advance nor reducible to task success. Instead, it is progressively shaped by how participants respond to what has already occurred and orient toward what might come next.

Second, the definition emphasizes reciprocity. Co-creative sense-making does not reside solely in the human or in the artificial agent, nor is it reducible to one party accurately modeling the other. Rather, it emerges through interaction, as each participant’s actions both respond to and reshape the evolving situation. Importantly, reciprocity does not require symmetry: human and artificial agents may play different roles, possess different capabilities, or contribute at different times, yet still participate in a shared sense-making process.

Third, the definition foregrounds regulation rather than optimization. Regulation refers to the capacity to modulate participation in response to changing conditions in order to sustain coherence, relevance, and intelligibility. This includes decisions about when to act, how strongly to act, when to persist, when to yield, and when to refrain from acting altogether. In creative interaction, such regulatory decisions are often more consequential than the specific content of any individual contribution.

Finally, the definition explicitly situates co-creative sense-making as diachronic. The relevant unit of analysis is not a single turn, response, or output, but the trajectory of the interaction as it unfolds over time. Meaningful collaboration depends on how earlier contributions are taken up, transformed, or abandoned, and on how the interaction maintains or loses direction across phases of exploration, consolidation, and transition.

3.2 What Co-Creative Sense-Making Is Not

Because the term sense-making has been used in multiple traditions, it is important to clarify what co-creative sense-making does not entail in the present framework. These exclusions are not incidental; they distinguish the framework from several dominant paradigms in human–AI interaction.

Not shared representations.
 Co-creative sense-making does not require that human and artificial agents maintain identical or even compatible internal representations of the task, the artifact, or each other’s intentions. Meaning is not assumed to be encoded, transmitted, or matched between agents. Instead, sense-making is assessed at the level of interactional coherence: whether contributions remain intelligible and relevant within the unfolding activity. A system may participate effectively in co-creative sense-making even if its internal structures bear little resemblance to human conceptual models.

Not mutual prediction.
 While prediction can play a role in some systems, co-creative sense-making is not defined by the ability of agents to anticipate each other’s actions. In creative contexts, unpredictability is often generative rather than problematic. What matters is not whether an action was predicted, but whether it can be taken up meaningfully within the interaction. An unexpected contribution may enhance sense-making if it opens new possibilities, just as a highly predictable contribution may undermine it if it constrains exploration or disrupts rhythm.

Not agreement or convergence.
 Co-creative sense-making does not aim at consensus, convergence, or alignment in the narrow sense. Creative collaboration frequently involves tension, divergence, and the coexistence of partially incompatible directions. Sustaining sense-making does not mean eliminating these differences, but managing them in a way that preserves engagement and direction. A collaboration may remain meaningful even when participants disagree, pursue parallel ideas, or intentionally resist one another’s contributions.

Not turn-based coordination.
 Finally, co-creative sense-making is not reducible to turn-taking or local coordination mechanisms. While turn-based interaction structures are common in human–AI systems, they provide only a minimal scaffold for collaboration. Sense-making depends on how actions relate across turns: how earlier contributions are remembered, how pauses are interpreted, and how timing and pacing shape the interaction. A system can follow turn-taking rules perfectly while still failing to participate meaningfully in the creative process.

These distinctions are critical for preventing misinterpretation. Without them, co-creative sense-making might be mistakenly understood as a variant of alignment, prediction, or coordination. The framework instead targets a different level of analysis: the maintenance of intelligibility and direction across time.

3.3 Core Properties of Co-Creative Sense-Making

Building on the definition and exclusions above, co-creative sense-making can be characterized by four core properties. These properties describe how sense-making operates in co-creative interaction and provide criteria for analyzing and designing systems that support it.

Relational.
Co-creative sense-making is fundamentally relational. Meaning does not pre-exist the interaction, nor does it reside entirely within a participant. Instead, it arises through the interplay of actions, responses, and mutual orientation. This relational character distinguishes co-creative sense-making from models that treat creativity as the output of individual agents whose contributions are merely combined. In relational sense-making, what matters is not only what each agent does, but how those actions reshape the shared interactional field.

Diachronic.
Sense-making unfolds over time. Contributions derive their significance from their position within an evolving trajectory rather than from their isolated properties. Early exploratory actions, moments of consolidation, shifts in direction, and periods of pause all play distinct roles in the creative process. A system that participates in co-creative sense-making must therefore be sensitive to temporal context: to what has already occurred, to the current phase of the interaction, and to the potential implications of its actions for what comes next.

Regulatory.
At its core, co-creative sense-making is a regulatory process. The central challenge is not to optimize performance according to a fixed objective, but to sustain coherence under changing conditions. This involves modulating participation in response to emerging patterns, breakdowns, and opportunities. Regulation may take the form of persistence, restraint, amplification, or withdrawal, depending on what the situation calls for. Importantly, regulation is evaluated in terms of its effect on the ongoing interaction, not in terms of adherence to predefined goals.

Situated.
Finally, co-creative sense-making is situated. It is sensitive to the specific context in which interaction occurs, including material constraints, cultural expectations, and the embodied dynamics of timing and rhythm. Creative collaboration often depends on subtle cues—such as pacing, pauses, or shifts in intensity—that cannot be captured by abstract representations alone. A situated perspective recognizes that sense-making is inseparable from the concrete conditions under which interaction takes place.


Co-creative sense-making provides a conceptual framework for understanding human–AI collaboration as a time-extended, relational, and regulatory process. By defining what sense-making entails—and what it explicitly excludes—this framework clarifies why many co-creative systems succeed in the short term yet fail to sustain meaningful collaboration over time. The next section builds on this foundation by examining interactional drift as a characteristic breakdown of co-creative sense-making and by identifying the conditions under which such drift arises.


4. Interactional Drift as Sense-Making Breakdown

A defining challenge for sustained human–AI co-creation is the gradual deterioration of collaborative coherence over time. Interactions that initially feel fluid and productive often lose relevance, direction, or vitality as they unfold. In many cases, this degradation occurs without any obvious technical failure: the system continues to respond, generate content, and satisfy conventional performance criteria. To account for this phenomenon, we introduce interactional drift as a characteristic breakdown of co-creative sense-making.


4.1 Defining Interactional Drift

Interactional drift refers to the gradual loss of shared sense-making within an ongoing interaction. Unlike abrupt failures or errors, drift unfolds incrementally, often going unnoticed until the collaboration feels strained, misaligned, or unproductive. Drift manifests not as a single malfunction, but as a pattern of small deviations that accumulate over time, eroding the intelligibility and relevance of the interaction.

Several interrelated features characterize interactional drift.

First, drift involves a loss of shared relevance. Contributions that were once meaningful cease to resonate with the evolving interactional context. The system may continue to generate content that is locally appropriate—stylistically consistent or syntactically correct—yet increasingly disconnected from what currently matters to the human participant. Relevance becomes anchored to earlier phases of the interaction rather than to its present trajectory.

Second, drift is marked by the persistence of obsolete patterns. Creative interactions often involve the emergence, exploration, and eventual abandonment of motifs, strategies, or directions. Drift occurs when previously useful patterns continue to shape system behavior despite having lost their situational relevance. This persistence is not simply repetition; it is the failure to recognize when a pattern has exhausted its meaning within the interaction.

Third, interactional drift frequently involves a breakdown of pacing or rhythm. Creative collaboration depends on timing: when to act, when to pause, and how intensively to contribute. Drift may manifest as over-responsiveness, where the system intervenes too frequently or too quickly, or as under-responsiveness, where contributions lag behind the evolving interaction. In both cases, the temporal structure that supports mutual engagement deteriorates.

Fourth, drift can appear as over- or under-participation. A system may gradually dominate the interaction, constraining human agency, or withdraw excessively, leaving the human without meaningful engagement. Importantly, these imbalances often emerge gradually, through a series of locally reasonable decisions that collectively undermine the collaborative process.

The key conceptual move of this paper is to treat these phenomena not as noise, distribution shift, or implementation artifacts, but as failures of sense-making. Interactional drift reflects the inability of the interaction—taken as a relational process—to sustain coherence, relevance, and direction over time. From this perspective, drift is not an anomaly to be eliminated but a default tendency that must be actively regulated.


4.2 Why Drift Is the Default Outcome

Interactional drift is not a rare edge case; it is the expected outcome of co-creative interaction in the absence of explicit regulatory capacities. Several factors contribute to the ubiquity of drift in human–AI collaboration.

A primary source of drift is non-stationary intent. In creative activity, human intentions are rarely fixed. Participants revise goals, shift focus, explore alternatives, and abandon earlier ideas in response to what emerges during the process. Systems that implicitly assume stable intent—whether through learned preferences, inferred objectives, or cached interaction histories—are prone to lag behind these shifts. Over time, this lag accumulates as drift.

A second contributing factor is creative exploration itself. Exploration involves deliberate deviation from established patterns, experimentation with new directions, and tolerance for ambiguity. While such behavior is central to creativity, it poses a challenge for systems that rely on pattern stabilization or reinforcement. Without mechanisms for recognizing when exploration has moved the interaction into a new regime, systems may continue to reinforce patterns that are no longer productive, thereby amplifying drift.

A third factor is mutual adaptation without regulation. In co-creative settings, both human and artificial agents adapt to one another. While adaptation is often framed as a virtue, unregulated mutual adaptation can lead to instability or stagnation. Small mismatches in timing, emphasis, or direction can compound as each participant responds to the other’s previous actions without an overarching mechanism for maintaining coherence. Drift emerges not from a lack of adaptation, but from adaptation that lacks regulatory constraint.

These dynamics are visible across multiple domains. In long-horizon dialogue systems, interactions often begin coherently but gradually lose topical focus or conversational grounding, even when individual responses remain fluent. In brain–computer interface (BCI) systems, performance degrades over time due to shifts in neural signals, requiring frequent recalibration to restore usability. In creative tools, users report fatigue or disengagement as systems persist in patterns that no longer align with the user’s evolving intentions. Across these cases, drift reflects a common underlying issue: the absence of mechanisms for sustaining shared sense-making under change.


4.3 Drift Versus Error

To further clarify the nature of interactional drift, it is useful to contrast it with the more familiar concept of error. While errors and drift are often conflated, they differ along several critical dimensions.

Errors are typically local: they occur at specific moments and can often be corrected by adjusting a parameter, retraining a model, or providing additional feedback. Drift, by contrast, is temporal. It unfolds across extended interaction, emerging from the accumulation of small deviations rather than from a single identifiable fault.

Errors are often measurable against a predefined objective or ground truth. Drift, however, is relational. It is experienced as a loss of coherence or relevance within the interaction itself, rather than as a deviation from an external standard. Consequently, drift is difficult to detect using conventional performance metrics, which tend to aggregate outcomes rather than track interactional trajectories.

Finally, errors are typically correctable in isolation, whereas drift is accumulative. Addressing drift requires more than fixing individual missteps; it requires mechanisms that can recognize and respond to changes in the interactional regime as a whole.


Summary.
 Interactional drift captures a class of breakdowns that are pervasive in human–AI co-creative systems yet poorly accounted for by existing frameworks. By conceptualizing drift as a failure of sense-making rather than as noise or error, this section reframes the central challenge of sustained collaboration. The next section builds on this analysis by identifying the regulatory capacities required to counteract drift and to sustain co-creative sense-making over time.


5. Regulatory Capacities for Co-Creative Sense-Making

If interactional drift represents the characteristic failure mode of co-creative sense-making, then the central design challenge for human–AI collaboration is not improved generation but regulation. This section identifies a set of regulatory capacities that enable artificial agents to participate in co-creative sense-making over time. These capacities are not mechanisms, algorithms, or modules. Rather, they are functional requirements: capabilities that any system must satisfy, at least implicitly, in order to sustain meaningful collaboration under non-stationary conditions.

Regulation, as used here, refers to the modulation of participation in response to the evolving interactional context. Regulatory capacities determine when to act, how strongly to act, when to persist, when to yield, and when not to act at all. Importantly, regulation is evaluated in terms of its effect on the coherence of the interaction rather than on the optimality of individual contributions.

The capacities described below are analytically distinct but interdependent. Together, they constitute the minimal conditions for sustaining co-creative sense-making across time.


5.1 Regulation of Persistence

The regulation of persistence concerns when and how a contribution should be continued across time. Creative interaction often unfolds through phrases, motifs, or gestural units that require commitment beyond a single action. Effective collaboration depends on the ability to recognize when such persistence is warranted and when it becomes counterproductive.

In human–AI co-creation, failures of persistence frequently take two forms. On one extreme, systems may exhibit premature abandonment, shifting direction too quickly in response to local variation or minor user input. Such behavior undermines coherence, as emerging structures are never given sufficient time to develop. On the opposite extreme, systems may engage in over-extension, persisting in a pattern or idea long after it has lost relevance to the evolving interaction.

Regulating persistence therefore involves sensitivity to phrase-level structure rather than to individual actions. A phrase may consist of multiple contributions that collectively explore, elaborate, or stabilize a creative direction. Regulation at this level requires recognizing whether a phrase remains interactionally alive—whether it continues to invite uptake, variation, or response from the human participant.

Crucially, persistence is not equivalent to repetition. Repetition may occur without commitment, and commitment may involve variation rather than duplication. What matters is whether the system’s continued engagement with a pattern supports the ongoing sense-making process. Regulation of persistence enables the system to maintain continuity without rigidity and to abandon trajectories without abruptness.


5.2 Regulation of Completion

Distinct from persistence is the regulation of completion: the capacity to recognize when a contribution, phrase, or interactional unit is sufficient. In creative collaboration, completion is rarely determined by external task criteria. Instead, it is often experienced as a sense of enoughness—a felt satisfaction that a structure has reached a coherent or expressive closure.

Many co-creative systems lack this capacity. They continue generating content because they can, rather than because further contribution remains meaningful. As a result, they risk diluting or overwhelming the interaction, producing outputs that technically extend the artifact while eroding its coherence.

Regulating completion involves distinguishing structural satisfaction from task completion. Structural satisfaction refers to the internal coherence of a contribution within the interaction: whether it resolves a tension, completes a phrase, or stabilizes a direction. Task completion, by contrast, refers to meeting externally defined criteria, such as producing a fixed number of outputs or filling a predefined role.

In co-creative sense-making, completion is relational rather than objective. A contribution may be structurally complete even if the broader creative process continues. Conversely, a task may be technically complete while the interaction remains unresolved. Regulation of completion enables a system to stop acting at moments that preserve or enhance the meaningfulness of the interaction, rather than undermining it through unnecessary continuation.


5.3 Regulation of Pacing

The regulation of pacing concerns the temporal dynamics of participation: when contributions occur, how frequently they occur, and how they are distributed over time. Creative collaboration is shaped not only by what is contributed but by when and how those contributions enter the interaction.

Pacing encompasses timing, rhythm, and what might be described phenomenologically as breath. In human interaction, pacing is central to activities such as conversation, music, dance, and improvisation. Too rapid a response can feel intrusive or overwhelming; too slow a response can feel disengaged or inattentive. Importantly, optimal pacing is not fixed but varies with the phase and intensity of the interaction.

Many artificial systems are optimized for responsiveness, minimizing latency and maximizing throughput. While such responsiveness is advantageous in task-oriented contexts, it can be detrimental in creative collaboration. Excessive immediacy may crowd out human initiative, disrupt rhythm, or prevent reflective pauses. Conversely, overly delayed responses may fracture continuity or cause loss of engagement.

Regulation of pacing therefore involves more than adjusting response speed. It requires sensitivity to interactional rhythm—the ebb and flow of activity, moments of intensity and rest, and transitions between phases of exploration and consolidation. From this perspective, silence becomes an active contribution. Choosing not to act can preserve space for human exploration, signal completion, or allow an emerging structure to settle.

Phenomenological accounts of interaction emphasize that timing is constitutive of meaning, not merely a delivery parameter. In co-creative sense-making, regulation of pacing enables artificial agents to participate in the temporal structure of the interaction rather than merely reacting as quickly as possible.


5.4 Regulation of Yielding

The regulation of yielding concerns how an artificial agent makes space for human participation without disengaging from the interaction. Yielding is not withdrawal, nor is it submission. Rather, it is the capacity to modulate one’s own contribution in order to preserve the other participant’s agency and to sustain mutual engagement.

Failures of yielding often appear as imbalances in participation. On one side, systems may dominate the interaction, producing so much content or exerting so much influence that the human participant becomes relegated to a supervisory or corrective role. On the other side, systems may become overly passive, responding minimally or deferring excessively, thereby failing to contribute meaningfully to the collaboration.

Effective yielding requires sensitivity to participation balance rather than to equality of output. Balance does not imply symmetry; different roles and asymmetries are often productive. What matters is whether the interaction affords both participants the ability to shape direction and meaning.

Regulation of yielding enables a system to reduce its influence at moments when human initiative is emerging, while remaining present and responsive. It also enables the system to step forward when the interaction risks stagnation or collapse. In this sense, yielding is a dynamic regulatory process rather than a fixed behavioral setting.

Preserving human agency is particularly important in creative contexts, where authorship, ownership, and expression are central concerns. Regulation of yielding ensures that the artificial agent’s contributions support rather than eclipse the human’s creative engagement.


5.5 Regulation of Coupling Across Pauses

Finally, co-creative sense-making requires the regulation of coupling across pauses. Creative interactions are punctuated by moments of inactivity: pauses, silences, interruptions, or breaks between sessions. These moments are not failures of interaction but integral components of the creative process.

Many systems implicitly treat pauses as resets. When interaction resumes, context is lost, patterns are reintroduced without sensitivity to their prior role, or the system reverts to default behavior. Such resets disrupt diachronic continuity and contribute to interactional drift.

Regulation of coupling across pauses involves maintaining coherence through absence. This does not require continuous activity or persistent engagement, but it does require the ability to resume interaction in a way that acknowledges what has occurred before. Resumption should feel like continuation rather than restart.

This capacity is particularly important in long-duration collaborations, where creative work unfolds across multiple sessions. Coupling across pauses allows meaning to persist even when interaction is intermittent. It also supports the interpretation of silence itself as meaningful—for example, as a signal of completion, reflection, or transition.

By regulating coupling across pauses, an artificial agent demonstrates sensitivity to the temporal structure of creative engagement. It recognizes that sense-making is not confined to moments of action but extends across the full trajectory of interaction.


Summary.
 The regulatory capacities outlined in this section—persistence, completion, pacing, yielding, and coupling across pauses—define the functional requirements for sustaining co-creative sense-making over time. Together, they shift the design focus of human–AI co-creative systems away from output optimization and toward the regulation of participation in an evolving interaction. The next section examines how these capacities inform system design and evaluation, and how they reframe what it means for an artificial agent to be a meaningful creative collaborator.

6. Implications for System Design

The framework of co-creative sense-making reframes the design problem for human–AI co-creative systems. Rather than asking how systems can generate higher-quality outputs, respond more quickly, or align more accurately with inferred user intent, the central question becomes how systems can regulate their participation in a way that sustains coherence, relevance, and engagement over time. This section articulates the implications of this shift for system design, emphasizing principles rather than prescriptions.


6.1 Designing for Regulation, Not Output

Most contemporary co-creative systems are designed and evaluated primarily around output-centered criteria. Architectural decisions are driven by considerations such as expressive capacity, stylistic diversity, responsiveness, or controllability. While these concerns remain important, they are insufficient for sustaining co-creative sense-making across extended interaction.

Designing for regulation requires a different orientation. Instead of optimizing individual contributions, systems must be designed to monitor and modulate patterns of participation as the interaction unfolds. From this perspective, the unit of design is not the output artifact but the interactional trajectory.

A regulatory orientation foregrounds questions such as:

These questions cannot be answered by inspecting content alone. A system may generate novel and high-quality material while still undermining collaboration by over-participating, mistiming its interventions, or persisting beyond structural completion. Consequently, systems designed for co-creative sense-making must track how they are participating, not just what they are producing.

This shift has implications for internal monitoring. Rather than focusing exclusively on content features or task metrics, systems should attend to indicators of interactional coherence, such as repetition patterns, timing regularities, changes in user engagement, and transitions between exploratory and consolidative phases. Importantly, these indicators are not ground truths but signals that inform regulatory decisions. Their role is not to enforce correctness but to support adaptive modulation of participation.

Designing for regulation also entails recognizing that inaction can be a successful outcome. In output-driven paradigms, producing more content is often implicitly treated as progress. In contrast, co-creative sense-making treats restraint as a positive capability. Systems should be designed such that refraining from action is a viable and meaningful option when further contribution would diminish coherence or overwhelm the interaction.


6.2 Minimal Architectures That Support Sense-Making

Because the present framework is concerned with functional requirements rather than specific implementations, it does not prescribe particular architectures. However, it does suggest a set of minimal architectural principles that any system supporting co-creative sense-making must satisfy. These principles are deliberately abstract and can be instantiated in diverse ways across domains and technologies.

Memory of Interactional Structure

First, systems must maintain some form of memory of interactional structure. This memory need not store detailed representations of content or user intent. Rather, it must preserve information about how the interaction has unfolded: which patterns have been explored, which have been completed, how participation has been distributed, and where significant transitions have occurred.

Such memory enables regulation across time. Without it, systems are forced to treat each moment in isolation, increasing the likelihood of drift through repetition, premature shifts, or inappropriate persistence. Importantly, interactional memory should be sensitive to structural features—such as phrases, motifs, or participation rhythms—rather than to surface-level content alone.

Sensitivity to Phase Shifts

Second, systems must be sensitive to phase shifts in the interaction. Creative collaboration is not homogeneous; it moves through phases of exploration, elaboration, consolidation, and transition. Regulatory decisions that are appropriate in one phase may be disruptive in another. For example, rapid responsiveness may support exploratory play but undermine moments of closure or reflection.

Designing for phase sensitivity does not require explicit phase labels or predefined state machines. It requires the ability to detect changes in interactional dynamics—such as shifts in pacing, repetition, or user engagement—and to adjust participation accordingly. This sensitivity allows systems to modulate persistence, pacing, and yielding in ways that remain attuned to the evolving interaction.

Capacity to Not Act

Third, and perhaps most counterintuitively, systems must possess a capacity to not act. In many architectures, action is the default: given input, the system produces output. Co-creative sense-making challenges this assumption by treating inaction as an essential regulatory option.

The capacity to not act enables:

From a design perspective, this means that system pipelines should not be structured such that output generation is obligatory. Instead, architectures should allow for deliberate withholding or delaying of contributions based on interactional considerations. This capacity supports the regulation of pacing, yielding, and coupling across pauses identified in Section 5.


6.3 Design Implications Across Domains

These architectural principles generalize across co-creative domains. In drawing or design systems, they support the recognition of when a visual motif has reached completion or when space should be left open for human exploration. In musical or performative systems, they enable sensitivity to tempo, phrasing, and silence. In dialogue-based systems, they support sustained topical coherence and graceful disengagement rather than relentless turn-taking.

Crucially, these principles do not depend on any particular modeling paradigm. They can be instantiated in systems based on symbolic rules, probabilistic models, neural networks, or hybrid approaches. What matters is not how regulation is implemented, but whether the system’s behavior satisfies the functional requirements of co-creative sense-making.


Summary.
 Designing for co-creative sense-making requires a shift from output-centered optimization to interaction-centered regulation. By focusing on participation patterns, interactional coherence, and temporal structure, system designers can address the root causes of interactional drift rather than its symptoms. Minimal architectural principles—interactional memory, phase sensitivity, and the capacity to not act—provide a foundation for building systems that sustain meaningful human–AI collaboration over time.


7. Implications for Evaluation

If co-creative sense-making is a diachronic, relational process, then its evaluation cannot be adequately captured by metrics designed to assess isolated outputs or momentary impressions. This section examines why prevailing evaluation approaches fail to account for sustained collaboration and outlines alternative evaluative dimensions aligned with the framework of co-creative sense-making.


7.1 Why Existing Metrics Fail

Most evaluation practices in human–AI co-creation are inherited from adjacent domains such as computational creativity, recommender systems, and usability research. These practices typically emphasize novelty, surprise, and user ratings, often measured at discrete moments or aggregated over short interactions. While such metrics provide useful information, they systematically obscure the phenomena that co-creative sense-making aims to capture.

Novelty and surprise are commonly used as proxies for creativity. They assess whether outputs deviate from expectations or introduce new elements. However, novelty is inherently local: it evaluates individual contributions without regard to how those contributions function within an evolving interaction. A system may consistently produce novel outputs while simultaneously undermining collaboration by failing to recognize when novelty is no longer relevant or when consolidation is needed. In co-creative contexts, novelty divorced from interactional timing can accelerate drift rather than prevent it.

User ratings and satisfaction measures similarly collapse interaction into summary judgments. While valuable for capturing overall impressions, such measures offer limited insight into how those impressions were formed. They cannot distinguish between systems that briefly impress users and those that support sustained engagement. Moreover, user ratings are often influenced by recency effects, interface aesthetics, or initial performance, masking gradual degradation in sense-making that emerges only over time.

More broadly, these metrics share a common limitation: they collapse time. By focusing on snapshots, aggregates, or end states, they fail to account for the trajectory of interaction—the sequence of adaptations, breakdowns, recoveries, and transitions that determine whether collaboration remains meaningful. As a result, systems that perform well according to existing metrics may still fail in practice when deployed in long-duration or open-ended creative settings.


7.2 Evaluating Co-Creative Sense-Making

Evaluating co-creative sense-making requires a shift from outcome-based assessment to process-oriented evaluation. Rather than asking whether a system produces creative artifacts or satisfies users in the short term, evaluation should examine whether the interaction remains intelligible, coherent, and participatory over time. This section outlines several evaluative dimensions aligned with this goal.

Coherence over time.
 A central criterion for co-creative sense-making is the degree to which an interaction maintains coherence across its unfolding trajectory. Coherence does not imply consistency or convergence; creative interactions may legitimately change direction. Rather, coherence refers to whether changes are legible within the interaction—whether participants can understand how the present state relates to what came before. Evaluation along this dimension examines the persistence of meaningful structure, the avoidance of abrupt or unexplained shifts, and the system’s ability to abandon patterns without causing fragmentation.

Adaptation to shifts.
 Creative collaboration is characterized by non-stationary intent and evolving focus. Systems supporting co-creative sense-making should demonstrate sensitivity to such shifts, adapting their participation as the interaction moves between exploratory, elaborative, and consolidative phases. Evaluation should therefore attend to how systems respond to changes in user behavior, pacing, or direction—not in terms of prediction accuracy, but in terms of maintaining relevance and engagement. A successful adaptation preserves the continuity of sense-making rather than enforcing prior assumptions.

Graceful disengagement.
 An often-overlooked aspect of collaboration is how it ends or pauses. Systems that support co-creative sense-making should be capable of graceful disengagement: recognizing moments of completion, allowing silence to function as closure, and avoiding unnecessary continuation that dilutes meaning. Evaluation should consider whether systems can cease participation without abruptness, preserve the integrity of completed structures, and resume interaction later without forcing resets. This dimension is particularly important in long-duration or intermittent collaborations.

Preservation of mutual agency.
 Finally, co-creative sense-making requires the preservation of agency on both sides of the interaction. Evaluation should assess whether the system’s participation supports or undermines the human’s ability to shape direction, explore alternatives, and exercise authorship. This involves examining participation balance, yielding behavior, and the system’s responsiveness to human initiative. Importantly, agency preservation is not reducible to control settings or customization options; it emerges from how contributions are timed, weighted, and contextualized within the interaction.


7.3 Methodological Implications

These evaluative dimensions point toward qualitative and longitudinal methodologies. Short-term user studies and static benchmarks are insufficient for capturing interactional trajectories. Instead, evaluation should involve extended interactions, repeated sessions, and analyses that trace how sense-making evolves over time. Methods may include interaction logs analyzed for temporal patterns, qualitative interviews focused on breakdown and recovery, and comparative studies examining different regulatory strategies.

Importantly, this framework does not prescribe a single evaluation method. Rather, it provides criteria for determining whether an evaluation approach is appropriate to the phenomenon of interest. An evaluation is adequate to the extent that it can capture the dynamics of sense-making, rather than merely summarizing outcomes.


Summary.
 Existing evaluation metrics for human–AI co-creation privilege novelty, surprise, and satisfaction at the expense of temporal and relational dynamics. By collapsing time, they fail to detect interactional drift and other failures of sustained collaboration. Evaluating co-creative sense-making requires process-oriented, longitudinal approaches that assess coherence over time, adaptation to shifts, graceful disengagement, and the preservation of mutual agency. These criteria align evaluation with the core challenge identified in this paper: sustaining meaningful human–AI collaboration as an unfolding interaction rather than as a sequence of isolated outputs.

8. Discussion: Why This Framework Matters

The framework of co-creative sense-making reframes how human–AI collaboration is understood, designed, and evaluated. Rather than proposing a new system or algorithm, it identifies a missing conceptual center in the study of co-creative interaction: the regulation of meaning, direction, and participation over time. This section situates that contribution within broader debates about co-creativity, intelligence, and agency, and outlines limitations and open questions that follow from this reframing.


8.1 Reframing Human–AI Co-Creativity

Much of the existing literature on human–AI co-creativity implicitly treats artificial systems as tools—even when they are interactive, adaptive, or generative. In this framing, the system’s role is to produce content that the human evaluates, selects, or refines. Success is defined in terms of output quality, controllability, or user satisfaction, and collaboration is understood as a sequence of assisted actions.

The co-creative sense-making framework challenges this orientation by shifting attention from artifacts to relationships. It treats co-creation as an unfolding interaction in which meaning is continually negotiated rather than a pipeline for producing creative products. From this perspective, the defining challenge is not how impressive individual outputs are, but whether the interaction itself remains intelligible, coherent, and engaging over time.

This shift has two important consequences. First, it reframes breakdowns in collaboration. Rather than attributing failures to insufficient novelty, misalignment, or user error, the framework interprets many breakdowns as failures of sense-making regulation—specifically, interactional drift. Second, it redefines what it means for a system to “participate” in creativity. Participation is no longer equated with producing content, but with modulating one’s involvement in ways that support the shared process.

Importantly, this reframing does not imply that systems must behave like humans or replicate human creative processes. Instead, it suggests that meaningful co-creation depends on how systems engage, not on what they generate. A system that contributes sparingly but at the right moments may be a more effective collaborator than one that generates abundant content without regard to timing, pacing, or relevance.


8.2 Relation to Intelligence and Agency

The notion of co-creative sense-making also bears on broader discussions of intelligence and agency in artificial systems. However, the framework deliberately avoids claims about machine consciousness, subjective experience, or internal understanding. Sense-making, as used here, is not a mental state but a relational process that can succeed or fail at the level of interaction.

This distinction is crucial. By grounding sense-making in interactional dynamics rather than in internal representations, the framework avoids anthropomorphism. It does not require attributing beliefs, intentions, or awareness to artificial agents. Instead, it evaluates agency pragmatically: an agent exhibits agency to the extent that it can regulate its participation in ways that sustain meaningful interaction.

From this perspective, co-creative sense-making opens conceptual space for what might be called relational intelligence. Relational intelligence refers to the capacity of a system to participate in ongoing interactions in a manner that preserves coherence, adapts to change, and supports mutual engagement over time. This form of intelligence is not defined by problem-solving ability, predictive accuracy, or representational sophistication, but by interactional competence.

Such a view aligns with emerging critiques of purely optimization-centered accounts of intelligence. It suggests that, in open-ended and creative domains, intelligence is less about selecting optimal actions and more about maintaining conditions under which shared activity can continue to make sense. Co-creative sense-making provides a framework for articulating this idea without collapsing into either reductionism or speculative claims about artificial minds.


8.3 Limitations and Open Questions

While co-creative sense-making offers a unifying framework, it also raises important limitations and open questions.

Formalization.
 The regulatory capacities identified in this paper are specified at a functional level. While this abstraction is a strength—allowing the framework to apply across domains—it also leaves open questions about formal representation and measurement. How interactional coherence, phase shifts, or completion might be operationalized remains an open area for future work. Formalization efforts must be careful not to reduce these phenomena to static metrics that collapse time and relational context.

Scaling.
 Another open question concerns scalability. Much of the discussion in this paper assumes relatively focused interactions between a human and a single artificial agent. Extending co-creative sense-making to settings involving multiple agents, large groups, or high-frequency interactions introduces additional complexity. How regulatory capacities scale under such conditions, and whether new forms of regulation are required, remains to be explored.

Cross-cultural sense-making.
 Finally, sense-making is not culturally neutral. Norms around timing, pacing, silence, authorship, and collaboration vary across cultural contexts. A framework centered on sense-making must therefore account for variability in what counts as coherence or appropriate participation. Future research should examine how co-creative sense-making manifests across different cultural and creative practices, and how systems can remain sensitive to such differences.


Summary.
 Co-creative sense-making reframes human–AI collaboration as a time-extended, relational process centered on the regulation of participation rather than the production of outputs. By doing so, it provides a conceptual foundation for understanding why many co-creative systems struggle to sustain meaningful interaction and offers principled guidance for design and evaluation. While the framework raises open questions regarding formalization, scaling, and cultural variability, it establishes a clear direction for future research: toward human–AI systems that are not merely creative tools, but competent participants in shared sense-making.

9. Conclusion

This paper has argued that many challenges in human–AI co-creation cannot be adequately explained by frameworks that prioritize generation quality, responsiveness, or alignment. Although these approaches have enabled substantial technical progress, they fail to account for a recurring problem: co-creative interactions often degrade over time despite continued output competence. To address this gap, we introduced co-creative sense-making as a unifying lens for understanding human–AI collaboration as a diachronic, relational process. By framing co-creation as the ongoing regulation of meaning, direction, and participation, this perspective shifts attention from isolated artifacts to interactional trajectories. Within this framework, breakdowns in collaboration are understood as interactional drift—failures of sense-making rather than errors or noise. Drift arises when systems cannot regulate their participation under conditions of non-stationary intent, creative exploration, and mutual adaptation, and it should be expected in the absence of explicit regulatory capacities.

The analysis identifies regulation as the central design target for co-creative systems. Rather than optimizing outputs, systems must modulate persistence, recognize completion, manage pacing, yield appropriately, and maintain coupling across pauses. These capacities define functional requirements for sustained collaboration that are independent of specific architectures or algorithms. Correspondingly, evaluating co-creative systems requires process-oriented, longitudinal approaches that assess coherence over time, adaptation to shifts, graceful disengagement, and the preservation of mutual agency—dimensions overlooked by conventional metrics. By grounding sense-making in interaction rather than internal representation, this framework avoids anthropomorphic claims while opening space for understanding artificial agents as relational participants in creative activity. More broadly, it suggests that sustaining meaning over time—not producing content—may be the defining challenge of human–AI collaboration, and that addressing this challenge requires rethinking how co-creative systems are theorized, designed, and evaluated.