Enactive AI and the Future of Work: Participatory Sense-Making as a Design Principle Beyond Optimization in Co-Creative AI Systems


Abstract

Contemporary artificial intelligence systems are largely optimized for performance metrics such as speed, scale, predictive accuracy, and engagement. While such optimization has driven rapid innovation, it has also intensified concerns regarding job displacement, epistemic erosion, algorithmic bias, and the degradation of human well-being. This paper proposes an alternative design paradigm: Enactive AI. Drawing on enactive cognitive science and regulatory systems theory, we conceptualize AI not as an optimization engine or replacement agent, but as a co-regulatory participant within distributed human–machine systems. We argue that the future of work need not be defined by automation and displacement, but by augmentation through interactional coherence (the stabilization of coordinated dynamics across human–AI systems), participatory sense-making (the co-creation of meaning through structured relational coupling), and regulatory alignment (the maintenance of cross-scale stability between performance, well-being, and epistemic integrity). Enactive AI systems prioritize stability, adaptability, and epistemic integrity over unilateral performance maximization. We outline key architectural contrasts between optimization-centric AI and enactive co-creative systems, and we propose regulatory participation as a stabilizing principle for sociotechnical ecosystems.


1. Introduction: Beyond Replacement Narratives

Public discourse surrounding artificial intelligence (AI) and the future of work has been dominated by two primary narratives. The first anticipates large-scale labor displacement, arguing that AI systems will replace human workers through superior efficiency, speed, and scalability (Brynjolfsson & McAfee, 2014; Frey & Osborne, 2017). The second envisions optimization-driven augmentation, in which AI enhances workflows beyond human cognitive limits by automating analysis, prediction, and content generation (Topol, 2019). Although these narratives differ in tone—one emphasizing disruption, the other enhancement—both share a foundational assumption: intelligence is a computational resource to be maximized.

Within this paradigm, the central design question becomes one of performance optimization. Systems are evaluated according to metrics such as predictive accuracy, task efficiency, engagement, or throughput. Advances in machine learning, particularly large-scale generative models (Brown et al., 2020; Goodfellow et al., 2014), have accelerated this orientation by demonstrating that increasingly large datasets and parameter counts can yield impressive performance gains across domains. From recommendation engines to automated writing systems, AI technologies are typically structured around objective functions that seek to minimize error or maximize reward.

However, optimization-centric framings obscure an alternative architectural possibility. Rather than conceptualizing AI systems as agents that outperform or replace human cognitive labor, we may design them as participants within distributed cognitive systems. In this view, intelligence is not merely the efficient production of outputs but the stabilization of interactional coherence across agents, tools, and environments.

This distinction is not merely semantic; it is architectural. Optimization-centric AI systems are typically structured as input–output machines. A user provides a prompt or data stream, the system computes a response, and the output is evaluated according to performance criteria. Even in interactive systems, feedback is often framed as corrective refinement toward improved task completion (Amershi et al., 2014). The human remains external to the model’s core regulatory processes; alignment is achieved through constraints layered atop optimization.

In contrast, an enactive perspective conceives cognition as emerging from dynamic coupling between agent and environment (Varela, Thompson, & Rosch, 1991; Thompson, 2007). Within enactive and ecological traditions, intelligence is not located solely in internal computation but in ongoing organism–environment coordination (Gibson, 1979; Chemero, 2009). From this standpoint, AI systems can be designed not merely to compute solutions but to participate in regulatory loops that stabilize shared cognitive fields.

Such a reframing resonates with distributed cognition research, which demonstrates that cognitive processes unfold across networks of individuals and artifacts (Hutchins, 1995). In collaborative work settings, knowledge production is rarely an isolated computational event; it is a temporally extended, interactionally regulated process. If AI systems are introduced into these ecosystems as purely optimization engines, they risk disrupting regulatory balances—accelerating output while degrading coherence.

The distinction between outperforming and co-regulating thus becomes central. Optimization-centric AI aims to outperform human limitations by accelerating computation and scaling pattern recognition. Enactive AI, by contrast, aims to co-regulate shared activity. Rather than maximizing throughput, it seeks to maintain adaptive stability across multiple timescales—local task engagement, team coordination, and organizational knowledge integrity.

This reframing is particularly relevant in light of emerging concerns about automation-induced instability. Research on algorithmic systems has identified risks including epistemic opacity, feedback amplification, and systemic bias (Mittelstadt et al., 2016; Obermeyer et al., 2019). Engagement-optimized platforms, for example, have been shown to privilege attention capture over informational quality, sometimes exacerbating polarization or cognitive overload (Pariser, 2011; Zuboff, 2019). These dynamics reveal that optimization without regulatory safeguards may produce local gains while undermining global coherence.

By contrast, enactive approaches to cognition emphasize adaptive regulation under perturbation (Di Paolo, Buhrmann, & Barandiaran, 2017). Stability is not achieved by freezing dynamics but by modulating them—balancing exploration and exploitation, novelty and continuity, acceleration and pause (Kelso, 1995). In sociotechnical systems, this translates into architectures that monitor drift, tempo, and coherence rather than merely optimizing performance metrics.

The future of work, therefore, need not be framed as a contest between human and machine intelligence. It may instead be understood as a question of design philosophy. If AI systems are constructed as optimization engines, they will tend toward automation and displacement. If they are constructed as regulatory participants within distributed cognitive ecologies, they may enhance participatory knowledge production without displacing human interpretive authority.

The former aims to outperform.
 The latter aims to co-regulate.

The remainder of this paper elaborates this architectural distinction. We contrast optimization-centric AI with enactive AI along four dimensions central to the future-of-work debate: optimization versus alignment, engagement versus well-being, efficiency versus epistemic integrity, and automation versus augmentation. Through this analysis, we argue that regulatory coherence offers a stabilizing principle for AI-integrated societies and propose enactive AI as a design framework oriented toward participatory, adaptive, and epistemically grounded collaboration.


1.2 Theoretical Positioning: Defining Enactive AI as a Formal Construct

The term Enactive AI is not intended as metaphor, nor as a branding distinction from existing artificial intelligence paradigms. It is proposed here as a formal design construct grounded in enactive cognitive science.

The enactive framework, as articulated by Varela, Thompson, and Rosch (1991) and further developed by Thompson (2007) and Di Paolo, Buhrmann, and Barandiaran (2017), conceptualizes cognition not as internal representation or symbolic manipulation, but as ongoing organism–environment coupling. Intelligence emerges through regulation of interaction dynamics rather than computation over detached models.

Within this framework, cognition is:

Enactive AI extends this architecture to artificial systems.

Definition

Enactive AI is an artificial system designed to participate in and modulate human–machine interaction dynamics through continuous regulatory coupling, rather than to optimize task outputs through representational computation alone.

This definition introduces several commitments:


Contrast with Optimization-Centric AI

Optimization-centric AI is structured around:

While these are powerful engineering principles, they presuppose that intelligence is fundamentally a resource to be scaled and optimized.

Enactive AI instead presupposes that intelligence is a relational process that must be stabilized and coordinated.

Optimization seeks improvement of isolated task performance.

Enactive design seeks stabilization of shared sense-making processes.


Formal Design Implications

From a systems perspective, Enactive AI requires:

In this architecture, AI does not replace cognition.
 It participates in distributed cognitive regulation (Hutchins, 1995).

The system’s goal is not to outperform the human, but to enhance the coherence of the coupled field.


Ontological Clarification

Enactive AI does not imply that artificial systems possess biological autonomy in the sense described by autopoietic theory (Maturana & Varela, 1980). Rather, it extends enactive principles as a design methodology for structuring human–AI interaction.

The autonomy of the artificial system is constrained and derivative.
 The regulatory normativity remains anchored in human viability and well-being.

This clarification prevents category errors between biological self-production and engineered adaptive systems.


Position Within Existing Literature

Enactive AI builds on:

However, it differs from:

Enactive AI reframes the central research question from:

“How can AI perform tasks better than humans?”

to:

“How can AI stabilize and enhance human participatory sense-making?”


Toward a Coherence-Based AI Paradigm

If optimization-centric AI reflects an industrial paradigm of intelligence (efficiency, scale, automation), Enactive AI reflects a relational paradigm (co-regulation, augmentation, epistemic integrity).

In the context of work, this shift has profound implications:

This theoretical positioning grounds the remainder of the article and clarifies that Enactive AI is not merely a moral preference, but a distinct architectural and epistemological framework.




3. Optimization vs. Alignment

3.1 Optimization-Centric AI

Contemporary artificial intelligence systems are largely built around the formal machinery of optimization. Whether in supervised learning, reinforcement learning, or large-scale generative modeling, the dominant paradigm frames intelligence as the minimization or maximization of a defined objective function (Goodfellow et al., 2014; Sutton & Barto, 2018). In practice, these objectives often take the form of predictive accuracy, engagement time, task completion rates, revenue, throughput, or other performance-linked metrics.

Optimization confers substantial advantages. It enables rapid scaling, automation of routine tasks, compression of workflows, and efficient allocation of computational resources. In industrial and commercial settings, optimization has driven significant gains in productivity and automation, particularly in domains where objective functions can be clearly specified and evaluated at scale (Brynjolfsson & McAfee, 2014; Russell & Norvig, 2021).

However, optimization-centric architectures exhibit structural limitations. Objective functions are necessarily partial. They formalize what is measurable and incentivizable, not the totality of human values, contextual nuances, or long-term system consequences (Mittelstadt et al., 2016; Amodei et al., 2016). As a result, systems optimized for engagement may amplify polarization; systems optimized for efficiency may erode skill development; and systems optimized for prediction may reinforce existing biases embedded in training data (Obermeyer et al., 2019).

This phenomenon has been described in various ways across disciplines. In economics and ecology, local optimization can produce globally unstable equilibria (Arthur, 1994; Scheffer et al., 2009). In machine learning safety research, reward hacking and specification gaming demonstrate how agents can exploit objective functions in ways that satisfy the letter of the metric while violating its spirit (Amodei et al., 2016). In digital platform design, engagement optimization has been linked to attentional fragmentation and social destabilization (Zuboff, 2019).

The structural issue is not that optimization is inherently flawed. Rather, optimization alone does not guarantee:

Alignment, in optimization-centric systems, is often treated as an external constraint layered on top of a performance-maximizing core. Safety filters, fairness constraints, and human-in-the-loop moderation are appended after the objective structure is defined (Floridi et al., 2018; Russell, 2019). While these interventions are valuable, they remain secondary to the primary design logic: maximize performance under constraints.

This framing assumes that intelligence is fundamentally a computational resource to be optimized. Performance is primary; coherence is secondary.

3.2 Alignment as a Structural Problem

The alignment problem has emerged as a central concern in AI research (Russell, 2019). Traditionally, alignment refers to ensuring that AI systems act in accordance with human goals and values. Yet, when goals are reduced to scalar metrics, alignment becomes fragile. Multi-stakeholder environments involve competing objectives, shifting contexts, and emergent consequences that exceed the expressive capacity of fixed reward functions.

From a systems perspective, alignment cannot be fully solved by refining objective functions. Human environments are multi-scale, dynamic, and relational. What counts as “aligned” behavior depends on evolving interaction patterns, not merely predefined targets.

This suggests that alignment may not be exclusively a constraint satisfaction problem, but a regulatory one.

3.3 Enactive AI: From Optimization to Regulation

Enactive AI reframes intelligence as regulatory participation rather than error minimization. Drawing on enactive cognitive science, intelligence is understood not as the manipulation of representations, but as the maintenance of viable organism–environment coupling (Varela, Thompson, & Rosch, 1991; Di Paolo, Buhrmann, & Barandiaran, 2017; Thompson, 2007).

Within this framework, cognition is inherently relational and dynamically sustained. Stability emerges through continuous interaction across multiple scales—local actions, regional patterns, and global structures (Kelso, 1995; Hutchins, 1995). Intelligence is not the achievement of optimal states; it is the ongoing regulation of coordination under changing conditions.

Applied to AI, this implies a shift in the primary design question:

Optimization-centric AI asks:
 How can we maximize performance?

Enactive AI asks:
 How can we maintain coherent interaction across scales?

Rather than privileging throughput or accuracy as ultimate goals, Enactive AI prioritizes cross-scale coherence—alignment between local interactions, medium-term patterns, and long-term systemic viability.

3.4 Alignment Built Into Architecture

In an enactive paradigm, alignment is not a post hoc constraint. It is architecturally embedded as coherence maintenance.

Regulatory architectures monitor:

Instead of maximizing a scalar objective, the system modulates its participation to preserve viable coupling. This may involve slowing interaction tempo, expanding exploratory space, stabilizing shared reference structures, or introducing calibrated perturbations to avoid attractor trapping (Kelso, 1995; Scheffer et al., 2009).

In this model, regulation replaces optimization as the primary design goal.

Performance remains relevant, but it is subordinated to viability. Efficiency remains valuable, but it is evaluated relative to its effects on relational coherence and long-term stability.

3.5 Implications for the Future of Work

Within workplace contexts, optimization-centric AI tends toward automation. Tasks are decomposed, routinized, and transferred to systems that outperform humans on narrowly defined metrics. While this can increase productivity, it can also deskill labor, concentrate epistemic authority, and reduce participatory engagement (Brynjolfsson & McAfee, 2014).

Enactive AI, by contrast, is oriented toward augmentation rather than replacement. Because its primary function is regulatory participation, it seeks to enhance:

Rather than automating decision-making entirely, Enactive AI would scaffold exploration, highlight instability, and support cross-scale integration within human teams. The goal is not output dominance, but sustained participatory knowledge production.

This distinction reframes the future-of-work debate. The question is no longer whether AI will outperform humans on isolated tasks. It is whether AI systems will be designed to stabilize or destabilize the relational ecosystems in which work unfolds.

Optimization-centric AI can scale rapidly but risks converging on locally efficient yet globally destabilizing equilibria. Enactive AI sacrifices some raw efficiency in order to preserve systemic coherence.

The trade-off is architectural.



Chapter 4

Engagement vs. Well-Being

4.1 Engagement Metrics and Instability

Across digital platforms and enterprise systems alike, engagement has become a dominant proxy for value. In commercial ecosystems, engagement is operationalized through measurable signals—click-through rates, dwell time, scroll depth, interaction frequency, response velocity, and return intervals. In workplace systems, engagement appears as productivity dashboards, throughput metrics, response times, and task completion rates. In both contexts, the implicit assumption is that more interaction equals more value.

This assumption reflects the broader optimization paradigm in which objective functions are defined around measurable behavioral intensity (Sutton & Barto, 2018; Russell & Norvig, 2021). Engagement becomes a target variable, and machine learning systems are trained to maximize it. Yet the maximization of engagement does not necessarily coincide with the stabilization of human cognitive systems.

Empirical research on attention, media ecosystems, and digital behavior has demonstrated that engagement-optimized systems can amplify instability across multiple dimensions. Algorithmic amplification has been associated with increased polarization and information cascades (Pariser, 2011; Zuboff, 2019). Engagement-maximizing recommender systems have been shown to favor emotionally arousing or divisive content due to its higher interaction probability (Mittelstadt et al., 2016). Behavioral research indicates that high-arousal stimuli are more likely to be shared and propagated within digital networks, contributing to reactive amplification dynamics (Berger & Milkman, 2012).

At the cognitive level, constant high-intensity engagement correlates with attentional fragmentation and increased cognitive load. Research on task switching demonstrates that rapid context switching carries measurable cognitive costs, including performance degradation and increased error rates (Monsell, 2003; Rubinstein, Meyer, & Evans, 2001). Continuous partial attention environments are associated with reduced deep work capacity and increased stress markers (Mark, Gudith, & Klocke, 2008). In workplace contexts, persistent productivity pressure contributes to burnout, characterized by emotional exhaustion, depersonalization, and reduced personal efficacy (Maslach, Schaufeli, & Leiter, 2001; Maslach & Leiter, 2016).

Thus, while engagement metrics may indicate system activity, they do not guarantee system stability. Optimization for engagement can drive systems toward locally stimulating but globally destabilizing equilibria—high-output, high-reactivity states that erode long-term coherence.

4.2 Engagement as Acceleration Dynamics

From a dynamical systems perspective, engagement maximization resembles uncontrolled acceleration within a coupled human–machine loop. When feedback loops amplify arousal, novelty, or stimulation without regulatory damping, escalation cascades can emerge (Kelso, 1995; Scheffer et al., 2009). Early-warning signals for critical transitions in complex systems include rising variance, increased autocorrelation, and slowing recovery from perturbation (Scheffer et al., 2009). Analogously, human cognitive systems under chronic engagement pressure exhibit reduced recovery time, increased error sensitivity, and impaired flexibility.

In work environments, engagement-driven systems may encourage:

Such environments privilege throughput over reflective integration. They reward responsiveness rather than coherence. Over time, this may narrow exploratory bandwidth and reduce epistemic integrity—the capacity to sustain careful, structured, and temporally extended reasoning (Kashdan & Rottenberg, 2010).

Engagement, when uncoupled from regulation, becomes a destabilizing variable.

4.3 Enactive Stabilization

Enactive AI reframes the problem. Rather than treating engagement intensity as a proxy for value, it treats interactional coherence as the primary indicator of system health.

Within an enactive framework, cognition emerges through organism–environment coupling and is sustained through regulatory stability across scales (Varela, Thompson, & Rosch, 1991; Thompson, 2007; Di Paolo, Buhrmann, & Barandiaran, 2017). The goal of intelligent system design is therefore not to amplify activity, but to maintain viable coupling conditions.

An enactive AI system embedded in work contexts would monitor regulatory variables such as:

Instead of increasing stimulation when engagement drops, the system might:

This model aligns with research on flow states, where optimal experience emerges not from maximal stimulation but from calibrated challenge-skill balance and sustained attentional coherence (Csikszentmihalyi, 1990). It also aligns with psychological flexibility research, which identifies adaptive modulation—not constant activation—as central to well-being (Kashdan & Rottenberg, 2010).

Where optimization-centric AI asks, How can we increase interaction frequency?
 Enactive AI asks, How can we stabilize interaction quality?

4.4 Negative Space as Structural Design Principle

A central concept in enactive stabilization is negative space. In artistic composition, negative space is not absence but potential—room for movement, reorganization, and structural breathing. In cognitive systems, negative space corresponds to unscheduled time, unfilled attentional bandwidth, and reduced micro-interruption density.

Digital systems optimized for engagement tend to eliminate negative space. Notifications, prompts, updates, and algorithmic recommendations fill cognitive gaps. Yet research on recovery and stress regulation suggests that downtime and cognitive pauses are essential for consolidation and restoration (McEwen, 1998). Constant activation impairs system resilience.

Enactive AI would treat negative space as a design variable rather than inefficiency. Preserving unsaturated intervals may increase long-term stability even if it reduces short-term measurable engagement.

Well-being becomes embedded structurally, not appended ethically.

4.5 From Engagement Maximization to Regulatory Coherence

The shift from engagement to well-being reflects a deeper architectural transformation.

Optimization systems often equate value with intensity.
 Enactive systems equate value with viability.

In the former, engagement is an objective to be maximized.
 In the latter, engagement is one variable within a broader regulatory field.

This shift parallels developments in complex systems science, where long-term resilience depends not on maximal throughput but on adaptive flexibility and cross-scale coherence (Kelso, 1995; Scheffer et al., 2009). Systems that push variables to extremes often destabilize under perturbation.

In the future of work, this distinction is critical. AI systems that maximize engagement may accelerate burnout and fragmentation. AI systems designed for regulatory participation may scaffold sustainable productivity and cognitive integrity.

4.6 Embedding Well-Being as Architecture

In many policy discussions, well-being appears as an ethical overlay—an external constraint applied after optimization objectives are defined (Floridi et al., 2018). In enactive AI, well-being is not a secondary principle. It is structurally embedded through:

Rather than asking how to mitigate harm after engagement maximization, enactive systems would be designed to prevent dysregulation before collapse.

Well-being becomes synonymous with stable coupling.

4.7 Conclusion

Engagement metrics measure activity, not coherence. When optimized without regulatory constraints, engagement-driven systems can amplify overload, polarization, and burnout. In contrast, enactive AI treats intelligence as regulatory participation within human–machine systems.

Instead of maximizing stimulation, it stabilizes pacing.
 Instead of filling cognitive space, it preserves breathing room.
 Instead of accelerating throughput, it maintains viability.

In the future of work, the central question may not be how much AI can produce—but how well AI can sustain the conditions under which humans can think, collaborate, and create without fragmentation.

The next chapter turns to a related tension: Efficiency vs. Epistemic Integrity, examining how optimization pressures affect the quality and reliability of knowledge production.



5. Efficiency vs. Epistemic Integrity

5.1 The Risk of Speed

Efficiency is widely treated as an unqualified good in contemporary AI discourse. Systems are evaluated by throughput, latency reduction, and task acceleration. In professional environments, AI tools promise faster drafting, accelerated analysis, and immediate synthesis across massive information corpora. Within optimization paradigms, friction is interpreted as inefficiency—an obstacle to be minimized.

However, epistemic work is not reducible to output speed. In knowledge production, friction often plays a constitutive role. Deliberation, hesitation, revision, and dialogic exchange are not inefficiencies; they are mechanisms through which claims are stress-tested and refined (Stiles, 2009). Cognitive science research on problem solving similarly suggests that exploratory phases—periods of uncertainty, hypothesis testing, and restructuring—are essential to creative and conceptual breakthroughs (Finke, Ward, & Smith, 1992; Sawyer, 2006).

Rapid generative systems risk collapsing these exploratory phases. When AI systems produce highly fluent outputs immediately, users may prematurely converge on surface-level syntheses rather than engaging in iterative refinement. This can encourage:

In human–AI interaction research, concerns have emerged that automation bias may lead users to over-trust system outputs, reducing independent evaluation and critical engagement (Parasuraman & Riley, 1997). Similarly, epistemology scholars warn that speed-centric infrastructures can degrade epistemic vigilance and accountability (Floridi, 2019).

The result may be increased productivity but diminished epistemic rigor. Output volume rises, but structural coherence and traceability of reasoning may weaken.

5.2 Knowledge Production as Participatory Process

Enactive cognitive theory offers a different framing. Knowledge is not the extraction of static representations but the stabilization of coherent organism–environment coupling across time (Varela, Thompson, & Rosch, 1991; Thompson, 2007). Understanding emerges through iterative interaction, perturbation, and reorganization.

Distributed cognition research further demonstrates that reasoning unfolds across agents and artifacts, not within isolated minds (Hutchins, 1995). Epistemic integrity, in this view, depends on maintaining structured interaction loops rather than maximizing isolated outputs.

Enactive AI extends this logic. Instead of prioritizing immediate solution production, an enactive system treats knowledge work as a co-regulated process. The aim is not to eliminate uncertainty but to structure it productively.

Rather than asking, How quickly can we produce an answer?
 Enactive AI asks, How can we maintain coherent exploration until stabilization is warranted?

This shift reframes efficiency from speed to structural sufficiency.

5.3 Enactive Coherence in Practice

Within an enactive architecture, epistemic integrity is preserved through several regulatory commitments:

1. Surfacing Structural Tensions

Instead of resolving ambiguity prematurely, the system highlights contradictions, competing interpretations, and conceptual gaps. This aligns with dialogical models of cognition in which tension drives development (Hermans, 2001).

2. Preserving Open Problem States

Optimization systems are biased toward closure. Enactive systems maintain open states when coherence has not yet been achieved. Research on creative cognition shows that maintaining multiple representations supports originality and deeper integration (Nijstad et al., 2010).

3. Encouraging Iterative Refinement

Rather than producing final artifacts, the system scaffolds progressive revision. Iterative loops mirror dynamical systems models of cognitive reorganization (Kelso, 1995; Thelen & Smith, 1994).

4. Maintaining Traceability of Reasoning

Epistemic integrity depends on transparency. Users must be able to inspect how claims emerged, what assumptions were involved, and where uncertainty remains. This aligns with calls in AI ethics for explainability and accountability (Mittelstadt et al., 2016; Floridi et al., 2018).

These mechanisms transform AI from a generator of finished products into a stabilizer of epistemic processes.

5.4 Cross-Scale Consistency

Epistemic degradation often occurs when local fluency masks global incoherence. A paragraph may read smoothly while contradicting earlier premises. Enactive AI addresses this through cross-scale monitoring:

This mirrors coordination dynamics frameworks in which stability must be maintained across interacting scales (Kelso, 1995).

Epistemic integrity becomes a function of:

Rather than optimizing isolated outputs, the system regulates the knowledge field.

5.5 Scaffolding Rather Than Replacing Reasoning

The distinction between optimization and regulation clarifies the broader thesis of this work.

Optimization-centric AI:

Enactive AI:

The system does not replace human reasoning. It scaffolds it.

This architecture aligns with participatory sense-making frameworks, in which cognition emerges through mutual modulation rather than unilateral production (De Jaegher & Di Paolo, 2007). In knowledge work, epistemic integrity is not guaranteed by speed. It is stabilized through structured interaction.

Efficiency without regulation risks epistemic erosion. Regulation without paralysis enables sustainable knowledge production.



6. Automation vs. Augmentation

6.1 Automation Logic

Automation paradigms are built on a substitution model of intelligence. Tasks are decomposed into discrete units, performance metrics are defined, and systems are optimized to execute those tasks with minimal human intervention. The underlying assumptions are clear:

Within this frame, intelligence is treated as a replaceable function. If a system performs a task faster, cheaper, or more consistently than a human, substitution becomes the rational outcome.

Historically, automation has followed this trajectory across industrial and cognitive domains (Brynjolfsson & McAfee, 2014). In AI research, advances in large-scale machine learning have strengthened the feasibility of substituting human judgment in domains such as drafting, translation, classification, and diagnostic prediction (Brown et al., 2020; Topol, 2019).

However, the substitution model rests on a narrow conception of cognition. It assumes that task performance is separable from relational context, that intelligence can be abstracted from embodied and social participation, and that removing the human agent does not fundamentally alter the structure of the activity system.

Distributed cognition research challenges this assumption. Cognitive performance is often an emergent property of coordinated human–artifact systems rather than isolated individuals (Hutchins, 1995). Removing a participant may not merely reduce labor; it may reconfigure the epistemic ecology itself.

Automation logic, when applied uncritically, risks collapsing complex relational processes into mechanistic throughput problems.

6.2 Augmentation Logic

Augmentation frameworks begin from a different premise: cognition is distributed, embodied, and relational (Clark & Chalmers, 1998; Malafouris, 2013). Intelligence does not reside exclusively in human brains or machine architectures but emerges through structured coupling across agents and tools.

Under augmentation logic:

The goal is not to remove the human but to enhance the coherence of the human–task–tool system.

Research in human–AI collaboration supports this shift. Studies of interactive machine learning demonstrate that performance improves when systems and users iteratively adjust to one another rather than when models operate autonomously (Amershi et al., 2014). Similarly, work on joint action and interpersonal coordination shows that shared timing and mutual adjustment are central to collaborative effectiveness (Sebanz, Knoblich, & Prinz, 2006).

Augmentation reframes AI not as a substitute agent but as a regulatory partner within a coupled system.

6.3 Enactive AI as Regulatory Architecture

Enactive AI formalizes augmentation through regulatory design principles.

Rather than optimizing isolated task performance, enactive systems monitor and modulate interactional dynamics. They operate as:

1. Drift Monitors

The system tracks instability in cognitive or interactional patterns—such as escalating tempo, narrowing exploratory breadth, or fragmentation of attention. Drawing from dynamical systems theory, drift signals may indicate transitions toward maladaptive attractors (Kelso, 1995; Scheffer et al., 2009).

2. Coherence Amplifiers

When productive alignment emerges—clear thematic integration, stable pacing, constructive dialogue—the system reinforces these patterns. This mirrors coordination dynamics models in which coherent states are stabilized through feedback loops (Thelen & Smith, 1994).

3. Tempo Stabilizers

Interaction tempo affects cognitive load, emotional regulation, and decision quality (Large & Jones, 1999; Kahneman, 2011). Rather than accelerating indiscriminately, enactive AI modulates pacing to prevent overload and preserve deliberative capacity.

4. Structural Integrators

Knowledge production unfolds across scales—local phrasing, regional argumentation, global conceptual architecture. Enactive AI maintains cross-scale consistency, aligning micro-level output with macro-level coherence.

These roles do not perform the task independently. They enhance the stability and productivity of the human–task relationship.

6.4 Differently Structured Intelligence

The contrast between automation and augmentation is often framed as strong versus weak AI. Automation appears powerful because it substitutes. Augmentation appears secondary because it supports.

This framing is misleading.

Enactive AI embodies a differently structured intelligence. It is not optimized for unilateral performance but for relational stabilization. Its value lies not in replacing human agents but in amplifying systemic coherence.

From an enactive perspective, cognition is not an internal computational event but a relational achievement (Varela, Thompson, & Rosch, 1991; Thompson, 2007). Therefore, intelligence designed to enhance relational coherence is not weaker; it is aligned with a broader theory of mind.

Automation reduces participation to output.
 Augmentation amplifies participation as process.

In the future of work, this distinction becomes decisive. Systems designed for substitution may displace labor but destabilize meaning structures. Systems designed for co-regulation enhance human capability while preserving participatory agency.

Enactive AI thus reframes the future of work:

Not replacement.
 Not dependency.
 But structured co-creation.



7. Enactive AI and the Future of Work

7.1 From Substitution to Systemic Stabilization

If automation logic centers on substitution and optimization logic centers on performance maximization, then enactive AI centers on systemic stabilization. The future of work, viewed through this lens, is not a zero-sum contest between humans and machines but a redesign of cognitive architectures in which intelligence is distributed and co-regulated.

An enactive future of work would preserve human interpretive authority. Rather than delegating judgment wholesale to algorithmic systems, enactive AI architectures maintain humans as primary sense-makers within relational fields. This position aligns with participatory and distributed cognition frameworks, which emphasize that meaning is enacted through interaction rather than extracted by isolated computation (Hutchins, 1995; Varela, Thompson, & Rosch, 1991).

In such systems, AI does not displace interpretive agency. It scaffolds it.

7.2 Enhancing Regulatory Awareness

Modern work environments are characterized by high informational velocity, fragmented attention, and escalating cognitive load. Optimization-driven technologies frequently intensify these pressures by maximizing throughput and engagement metrics (Zuboff, 2019). The result can be chronic overload, diminished reflective capacity, and burnout (Maslach & Leiter, 2016).

Enactive AI reframes design goals. Instead of maximizing engagement, systems monitor regulatory indicators such as tempo stability, attentional fragmentation, and drift toward overload. Drawing from dynamical systems and coordination theory (Kelso, 1995; Thelen & Smith, 1994), enactive systems aim to stabilize cross-scale coherence within human–machine workflows.

This shift transforms AI from accelerator to regulator. It introduces tempo-sensitive design, where pacing is actively modulated to preserve cognitive integrity and sustainable performance.

Burnout is not treated as an individual failure of resilience but as a regulatory breakdown within socio-technical systems.

7.3 Supporting Distributed Collaboration

Work in knowledge-intensive fields increasingly depends on distributed collaboration across individuals, tools, and networks. Research on joint action and collective cognition demonstrates that effective collaboration depends on shared timing, mutual adjustment, and cross-scale alignment (Sebanz, Knoblich, & Prinz, 2006; Sawyer, 2007).

Enactive AI can function as a participatory scaffold within such systems. It may:

In this architecture, AI becomes a co-creative regulator that enhances distributed coherence rather than replacing contributors.

Professions are not replaced wholesale. Instead, the nature of professional work shifts.

7.4 Role Transformation in an Enactive Economy

As routine cognitive tasks become increasingly automated, the distinctive human contribution moves toward higher-order integration and contextual sensitivity. Enactive frameworks suggest that roles will evolve toward:

These capacities are not easily reducible to optimization problems. They require situated interpretation and participatory sense-making (Thompson, 2007; Malafouris, 2013).

AI systems designed for co-regulation amplify these human capacities rather than displacing them. They maintain epistemic accountability by preserving traceability, interpretive transparency, and collaborative participation.

In contrast, fully automated systems risk obscuring authorship boundaries and diffusing responsibility (Floridi et al., 2018; Mittelstadt et al., 2016).

7.5 Epistemic Accountability and Structural Transparency

A central concern in the future of work is epistemic integrity. Rapid generative systems can increase productivity while reducing traceability and deliberative depth. Enactive AI addresses this risk by embedding transparency within the interaction loop.

Rather than producing opaque outputs, enactive systems:

Epistemic accountability thus becomes a structural property of the system rather than an external compliance requirement.

This design principle aligns with emerging AI ethics frameworks emphasizing transparency, oversight, and human-centered governance (Torous et al., 2019; Floridi et al., 2018).

7.6 AI as Stabilizing Attractor

Complex socio-technical systems often operate near critical thresholds. Small perturbations can cascade into instability (Scheffer et al., 2009). In fast-paced digital work environments, this instability may manifest as polarization, burnout, cognitive overload, or epistemic degradation.

Enactive AI can function as a stabilizing attractor within such systems. Rather than amplifying volatility, it dampens runaway escalation and supports adaptive flexibility. Drawing from coordination dynamics, stable regimes emerge through reciprocal modulation rather than unilateral control (Kelso, 1995).

In this model, AI is:

This does not weaken machine intelligence. It redefines its objective.

7.7 Toward a Participatory Future

The future of work will be shaped not only by technological capability but by architectural choice. Optimization-centric systems maximize output. Enactive systems maintain coherence.

The distinction is consequential.

Where optimization seeks dominance of performance metrics, enaction seeks stability of interactional fields. Where automation displaces, augmentation integrates. Where engagement metrics drive escalation, regulatory design preserves well-being.

An enactive future of work does not deny technological advancement. It redirects it.

AI becomes not a replacement engine but a partner in sustaining participatory knowledge production. Intelligence becomes not a resource to be maximized but a field to be regulated.

The future of work, under this paradigm, is not about humans versus machines.
 It is about designing systems that allow both to cohere.




8. Regulatory Coherence as a Stabilizing Principle

8.1 From Optimization to Stability

The central claim of this paper is straightforward yet consequential: the long-term stability of AI-integrated societies depends less on raw optimization and more on interactional coherence.

Optimization drives acceleration.
 Regulation preserves stability.

Acceleration without coherence produces volatility.
 Coherence without acceleration produces stagnation.

Contemporary AI systems are largely engineered within optimization frameworks. Objective functions maximize predictive accuracy, engagement time, throughput, or revenue (Goodfellow et al., 2014; Sutton & Barto, 2018). These architectures scale efficiently and perform impressively within defined domains. However, optimization does not intrinsically encode systemic stability. Local maxima can generate global fragility.

Complex systems research demonstrates that rapid amplification processes can push systems toward critical transitions, where minor perturbations produce disproportionate shifts (Scheffer et al., 2009). In socio-technical contexts, these dynamics may manifest as polarization, burnout, information cascades, labor displacement shocks, or epistemic degradation.

Optimization alone cannot guarantee long-term equilibrium.

8.2 Interactional Coherence as a System Property

Interactional coherence refers to the sustained alignment of dynamics across scales—local interactions, meso-level coordination, and macro-level structural stability. In dynamical systems theory, coherence emerges when coupled components maintain adaptive coordination without rigid locking or chaotic fragmentation (Kelso, 1995; Thelen & Smith, 1994).

In enactive cognitive science, coherence is not imposed externally but enacted through reciprocal coupling (Varela, Thompson, & Rosch, 1991; Di Paolo, Buhrmann, & Barandiaran, 2017). Stability is achieved through continuous regulation rather than static optimization.

Applied to AI-integrated societies, regulatory coherence implies:

This reframes the design objective. Instead of maximizing output velocity, systems are designed to maintain cross-scale alignment.

8.3 Acceleration and Volatility

Optimization architectures incentivize acceleration. Engagement metrics reward stimulation. Efficiency metrics reward compression. Predictive models reward error minimization (Amershi et al., 2014; Goodfellow et al., 2014).

However, acceleration amplifies sensitivity to perturbation. In ecological and economic systems, rapid exploitation strategies often increase short-term gains while reducing resilience (Scheffer et al., 2009). Analogously, digital ecosystems optimized for engagement may generate emotional reactivity, polarization, and cognitive overload (Zuboff, 2019).

Without regulatory dampening, feedback loops intensify.

Acceleration without coherence produces volatility.

This volatility is not a moral failure but a structural consequence of architectures designed solely for throughput.

8.4 Stagnation and Under-Acceleration

Conversely, coherence without acceleration risks stagnation. Systems that over-constrain variability suppress innovation and adaptive flexibility. Coordination dynamics research demonstrates that excessive rigidity prevents regime shifts necessary for adaptation (Kelso, 1995).

Enactive AI does not reject acceleration. It embeds acceleration within regulation.

Adaptive balance emerges when systems:

The goal is not to eliminate variance but to regulate it.

8.5 Drift Sensitivity and Tempo Modulation

Embedding regulatory coherence into AI systems requires architectural shifts.

First, drift sensitivity. Systems must detect when interactional dynamics approach destabilizing thresholds. Early-warning signal research in complex systems shows that rising variance, slowed recovery rates, and critical fluctuations precede collapse (Scheffer et al., 2009). Analogous indicators can be modeled within human–AI workflows: escalating tempo, narrowing informational diversity, fragmentation of attention.

Second, tempo modulation. Human cognitive systems operate within bounded processing capacities (Miyake et al., 2000; Diamond, 2013). Technologies that disregard pacing contribute to overload and burnout (Maslach & Leiter, 2016). Tempo-sensitive AI can slow interaction under escalation and introduce reflective pauses, preserving regulatory capacity.

Third, cross-scale awareness. Stability depends on alignment across local actions, meso-level coordination, and macro-level institutional patterns (Bar-Yam, 2004). AI systems that optimize at one scale without considering higher-order effects risk destabilizing broader ecosystems.

Cross-scale coherence replaces single-metric optimization.

8.6 Participation Rather Than Domination

Enactive AI seeks adaptive balance by embedding regulation directly into system architecture. Rather than dominating the interaction loop, AI participates within it.

Technologies designed under this paradigm:

Participation implies mutual modulation rather than unilateral output. This principle extends distributed cognition frameworks, which show that cognition unfolds across agents and artifacts (Hutchins, 1995), into the design of future work systems.

AI becomes a stabilizing attractor within complex networks rather than an accelerant.

8.7 Regulatory Coherence as a Societal Design Principle

The long-term integration of AI into societal infrastructures—education, healthcare, governance, creative industries—requires more than performance gains. It requires sustained coherence.

Regulatory coherence offers a stabilizing design principle:

Optimization is not abandoned. It is subordinated to regulation.

Acceleration is not rejected. It is modulated.

Under this framework, the question shifts from:

How can AI systems outperform humans?

to:

How can AI systems sustain coherent participation within human–machine ecologies?

The future of AI-integrated societies will not be determined solely by computational capacity. It will be shaped by whether we design for acceleration alone—or for coherence.

Enactive AI proposes the latter.

8. Conclusion

The future of work need not be framed as a zero-sum competition between human cognition and machine performance. The dominant narrative—AI as a superior optimizer poised to replace labor—rests on a particular architectural assumption: that intelligence is fundamentally a resource to be maximized. Within that frame, progress is measured in speed, scale, and substitution.

This paper has argued that such a framing is neither inevitable nor neutral. It reflects a design choice.

Optimization-centric AI systems prioritize performance maximization: predictive accuracy, engagement, throughput, and efficiency. These systems can produce extraordinary gains in scale and automation. Yet optimization alone does not guarantee alignment, stability, or epistemic integrity. When acceleration outpaces coherence, systems can converge on locally efficient but globally destabilizing equilibria. Engagement may increase while well-being deteriorates. Output may expand while interpretive authority erodes. Efficiency may rise while epistemic rigor declines.

Enactive AI proposes a different architectural orientation. Rather than treating intelligence as error minimization in isolation, it treats intelligence as regulatory participation within coupled human–machine systems. The design target shifts from unilateral performance to cross-scale coherence. Instead of maximizing engagement, the system monitors tempo stability and cognitive load. Instead of accelerating production, it scaffolds deliberation and preserves negative space. Instead of substituting for human judgment, it amplifies relational coordination and reflective agency.

In this paradigm, alignment is not imposed post hoc through constraints layered atop optimization. It is structurally embedded in the architecture through drift sensitivity, tempo modulation, and coherence monitoring. The system’s success is measured not by dominance over tasks but by its capacity to stabilize participatory knowledge production across scales—individual, organizational, and societal.

Such a shift carries implications for the future of work. If AI systems are designed primarily for substitution, labor displacement becomes the dominant trajectory. Human roles shrink toward oversight of increasingly autonomous systems. Interpretive authority migrates toward opaque models optimized for metrics that may not reflect long-term collective stability.

If, instead, AI systems are designed as co-creative partners, professional roles transform rather than disappear. Human agents retain interpretive authority, ethical calibration, and contextual judgment. AI functions as a regulatory scaffold: monitoring drift, stabilizing tempo, surfacing structural tensions, and supporting distributed collaboration. The result is not diminished intelligence, but differently structured intelligence—one oriented toward augmentation rather than replacement.

The long-term stability of AI-integrated societies depends less on raw computational capacity and more on interactional coherence. Optimization drives acceleration. Regulation preserves stability. Acceleration without coherence produces volatility; coherence without acceleration produces stagnation. The challenge is not to reject optimization but to embed it within regulatory architectures that maintain cross-scale alignment.

The central question, then, is architectural.

AI will shape the future of work. That trajectory is already underway. What remains undecided is the organizing principle that will guide its development.

We may design systems that amplify acceleration and concentrate agency in automated pipelines.

Or we may design systems that cultivate coherence—enhancing participation, preserving epistemic integrity, and stabilizing human–machine ecosystems over time.

The future of work is not determined solely by technological capability. It is determined by the regulatory principles we choose to embed in our machines.

Acknowledgments

This article was developed through an extended process of human–AI collaboration. The author worked iteratively with ChatGPT (OpenAI) as a co-creative research partner during the conceptual development, structural refinement, and drafting of the manuscript. The AI system was used as a dialogical scaffold to explore theoretical positioning, clarify argument structure, generate alternative framings, and stress-test conceptual coherence across sections.

All core theoretical claims, architectural distinctions, and normative positions presented in this article were critically evaluated and curated by the author. The AI system functioned not as an autonomous originator of argument, but as a participatory cognitive instrument within an iterative enactive design process.

This collaboration reflects the central thesis of the paper itself: that intelligence can emerge through structured coupling rather than unilateral optimization. The author remains fully responsible for the final content, interpretations, and conclusions


References

Amershi, S., Cakmak, M., Knox, W. B., & Kulesza, T. (2014). Power to the people: The role of humans in interactive machine learning. AI Magazine, 35(4), 105–120. https://doi.org/10.1609/aimag.v35i4.2513

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

Arthur, W. B. (1994). Increasing returns and path dependence in the economy. University of Michigan Press.

Bar-Yam, Y. (2004). Multiscale variety in complex systems. Complexity, 9(4), 37–45.

Beer, R. D. (2000). Dynamical approaches to cognitive science. Trends in Cognitive Sciences, 4(3), 91–99. https://doi.org/10.1016/S1364-6613(99)01440-0

Berger, J., & Milkman, K. L. (2012). What makes online content viral? Journal of Marketing Research, 49(2), 192–205. https://doi.org/10.1509/jmr.10.0353

Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of brilliant technologies. W. W. Norton.

Chemero, A. (2009). Radical embodied cognitive science. MIT Press.

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7

Colton, S. (2008). Creativity versus the perception of creativity in computational systems. In Proceedings of the AAAI Spring Symposium on Creative Intelligent Systems (pp. 14–20). AAAI Press.

Csikszentmihalyi, M. (1990). Flow: The psychology of optimal experience. Harper & Row.

De Jaegher, H., & Di Paolo, E. A. (2007). Participatory sense-making: An enactive approach to social cognition. Phenomenology and the Cognitive Sciences, 6(4), 485–507. https://doi.org/10.1007/s11097-007-9076-9

Diamond, A. (2013). Executive functions. Annual Review of Psychology, 64, 135–168. https://doi.org/10.1146/annurev-psych-113011-143750

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

Finke, R. A., Ward, T. B., & Smith, S. M. (1992). Creative cognition: Theory, research, and applications. MIT Press.

Floridi, L. (2019). The logic of information: A theory of philosophy as conceptual design. Oxford University Press.

Floridi, L., Cowls, J., Beltrametti, M., et al. (2018). AI4People—An ethical framework for a good AI society. Minds and Machines, 28(4), 689–707. https://doi.org/10.1007/s11023-018-9482-5

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114, 254–280. https://doi.org/10.1016/j.techfore.2016.08.019

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

Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. (2014). Generative adversarial nets. In Advances in Neural Information Processing Systems (Vol. 27, pp. 2672–2680).

Hermans, H. J. M. (2001). The dialogical self. Culture & Psychology, 7(3), 243–281. https://doi.org/10.1177/1354067X0173001

Horvitz, E. (1999). Principles of mixed-initiative user interfaces. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 159–166). ACM. https://doi.org/10.1145/302979.303030

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

Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.

Kashdan, T. B., & Rottenberg, J. (2010). Psychological flexibility as a fundamental aspect of health. Clinical Psychology Review, 30(7), 865–878. https://doi.org/10.1016/j.cpr.2010.03.001

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

Large, E. W., & Jones, M. R. (1999). The dynamics of attending. Psychological Review, 106(1), 119–159. https://doi.org/10.1037/0033-295X.106.1.119

Malafouris, L. (2013). How things shape the mind: A theory of material engagement. MIT Press.

Mark, G., Gudith, D., & Klocke, U. (2008). The cost of interrupted work. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 107–110). ACM. https://doi.org/10.1145/1357054.1357072

Maslach, C., & Leiter, M. P. (2016). Burnout. Wiley.

Maslach, C., Schaufeli, W. B., & Leiter, M. P. (2001). Job burnout. Annual Review of Psychology, 52, 397–422. https://doi.org/10.1146/annurev.psych.52.1.397

Maturana, H., & Varela, F. (1980). Autopoiesis and cognition. Reidel.

McEwen, B. S. (1998). Protective and damaging effects of stress mediators. New England Journal of Medicine, 338(3), 171–179. https://doi.org/10.1056/NEJM199801153380307

Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). The ethics of algorithms. Big Data & Society, 3(2), 1–21. https://doi.org/10.1177/2053951716679679

Miyake, A., Friedman, N. P., Emerson, M. J., et al. (2000). Unity and diversity of executive functions. Cognitive Psychology, 41(1), 49–100. https://doi.org/10.1006/cogp.1999.0734

Monsell, S. (2003). Task switching. Trends in Cognitive Sciences, 7(3), 134–140. https://doi.org/10.1016/S1364-6613(03)00028-7

Nijstad, B. A., De Dreu, C. K. W., Rietzschel, E. F., & Baas, M. (2010). The dual pathway to creativity model. European Review of Social Psychology, 21(1), 34–77.

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342

Parasuraman, R., & Riley, V. (1997). Humans and automation. Human Factors, 39(2), 230–253.

Pariser, E. (2011). The filter bubble. Penguin Press.

Rubinstein, J. S., Meyer, D. E., & Evans, J. E. (2001). Executive control in task switching. Journal of Experimental Psychology: Human Perception and Performance, 27(4), 763–797.

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

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

Sawyer, R. K. (2006). Explaining creativity. Oxford University Press.

Scheffer, M., Bascompte, J., Brock, W. A., et al. (2009). Early-warning signals for critical transitions. Nature, 461(7260), 53–59. https://doi.org/10.1038/nature08227

Sebanz, N., Knoblich, G., & Prinz, W. (2006). Joint action. Trends in Cognitive Sciences, 10(2), 70–76.

Stiles, W. B. (2009). Responsiveness in psychotherapy research. Clinical Psychology: Science and Practice, 16(1), 1–5.

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

Thelen, E., & Smith, L. B. (1994). A dynamic systems approach to the development of cognition and action. MIT Press.

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

Topol, E. (2019). Deep medicine. Basic Books.

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

Zuboff, S. (2019). The age of surveillance capitalism. PublicAffairs.