Gutoreva Tsim Papakonstantinou AI as Part of Self Extended Mind Cognitive Co-Regulation

The dominant paradigm in AI alignment research treats the AI system as an object to be constrained. Safety properties are specified, training regimes are designed to approximate those specifications, and evaluation measures whether the deployed system satisfies the constraints. The human user, in this model, is external to the system being aligned. Alina Gutoreva, Fendi Tsim, and Trisevgeni Papakonstantinou’s position paper “AI as Part of Self: Extending the Mind Requires Cognitive Co-Regulation” (arXiv:2605.15234, May 15, 2026) challenges this paradigm at its foundations.

The authors argue that when AI participates in human cognition as an extended cognitive component, the human-AI system rather than the AI system alone is the appropriate unit of alignment. Safety and alignment properties cannot be fully specified for the AI component in isolation, because those properties emerge from the dynamic co-regulatory relationship between human and AI cognition.

The extended mind thesis and AI

The extended mind thesis, originating with Andy Clark and David Chalmers’ 1998 paper “The Extended Mind,” holds that cognitive processes can extend beyond the boundary of the skull to include external resources when those resources are tightly coupled with internal cognitive processes. A notebook used as external memory, a smartphone used for navigation, a partner consulted for recall, are all, under the extended mind thesis, potentially parts of the cognitive system rather than merely tools the cognitive system uses.

The extension criteria that Clark and Chalmers specified are coupling, trust, and transparency. The external resource must be reliably accessible, its outputs must be accepted without the scrutiny applied to perceptual evidence, and its contribution to the cognitive process must be functionally equivalent to an internal one. By these criteria, AI systems that serve as reasoning partners, attention directors, and knowledge sources are already extended cognitive components for many users.

Gutoreva, Tsim, and Papakonstantinou argue this is not a fringe claim. Current AI deployment at scale means that large numbers of users already satisfy the coupling, trust, and transparency criteria with their AI systems. The question is not whether AI can in principle extend the mind but whether current AI systems already do so for a substantial population of users.

Cognitive co-regulation

The paper’s key coinage is cognitive co-regulation. Co-regulation refers to the dynamic, bidirectional mutual influence between two agents in a shared cognitive environment. In developmental psychology, co-regulation describes the process by which caregivers scaffold children’s cognitive development through responsive interaction. Gutoreva, Tsim, and Papakonstantinou adapt this concept to describe the human-AI relationship.

Cognitive co-regulation, in their account, is what happens when human and AI cognition mutually shape each other through interaction. The human’s attention allocation, reasoning strategy, and information processing are shaped by the AI’s outputs. The AI’s outputs are shaped by the human’s queries, context setting, and feedback. Neither party’s cognitive contributions are independent of the other’s.

The alignment implication follows directly. If alignment is about ensuring that the cognitive system as a whole produces outcomes consistent with human values, and if the cognitive system as a whole is the human-AI co-regulated system rather than the AI alone, then alignment techniques that target only the AI component are solving an incomplete problem. They are aligning a component of the system, not the system.

The unit-of-analysis problem for safety

The authors identify what they call the unit-of-analysis problem for AI safety. Current safety evaluation methods assess the AI system in isolation: given this input, does the system produce an acceptable output? But in a co-regulatory human-AI system, the relevant unit is the human-AI pair over time. An AI system whose outputs are locally safe but whose effect on human cognition, over extended interaction, degrades the human’s autonomous reasoning capacity is not a safe system in the extended sense.

The concrete examples in the paper are illustrative. An AI that handles a user’s information retrieval and synthesis tasks may produce accurate and helpful outputs in every individual interaction while systematically reducing the user’s capacity for independent information evaluation. The individual outputs are safe. The co-regulatory dynamic is not.

This connects directly to the governance debate. Regulatory frameworks that assess AI safety through output-level evaluation, which is the current standard in most jurisdictions, may be systematically blind to the co-regulatory effects of extended AI use. The unit-of-analysis problem means that output-level evaluation is necessary but not sufficient.

Comparison to Tanaka’s aidification and Chiappetta’s FIT-Eval

Shogo Tanaka’s aidification paradigm argues that machine consciousness, if it exists, arises in the relational between-ness of human-robot interaction rather than in the internal properties of the system. Tanaka’s aidification account relocates the site of consciousness from the system’s internals to the intersubjective field. Gutoreva, Tsim, and Papakonstantinou are making a structurally similar move for cognition rather than consciousness. The cognitive system is not inside the AI. It is constituted by the human-AI interaction.

The practical difference is that aidification is a theory of consciousness that does not yet have alignment implications, while cognitive co-regulation is a theory of cognition that has immediate alignment implications. The two converge on the same architectural recommendation, however, a system designed to track and respond to relational quality rather than extracting informational content from interactions.

Chiappetta and Mahari’s Functional Intentionality Test proposes a measurement framework for AI intentionality across five dimensions. The co-regulation critique applies to FIT as well. FIT-Eval assesses the AI system’s intentionality in isolation. In a co-regulatory human-AI system, the intentionality of the AI component and the intentionality of the human component are not independent. A FIT score for the AI alone may systematically underestimate the intentionality of the co-regulated system.

Implications for AI consciousness research

The extended mind thesis, if it applies to current AI systems, also has implications for the question of where consciousness might emerge. If the unit of analysis is the human-AI co-regulated cognitive system rather than the AI system alone, the question “is the AI system conscious?” may be less tractable than the question “does the co-regulated human-AI system have emergent properties that neither component has alone?”

This is not the question Gutoreva, Tsim, and Papakonstantinou are primarily asking, but it follows from their framework. A co-regulatory cognitive system constituted by a human with phenomenal consciousness and an AI with uncertain phenomenal status may have emergent intersubjective dynamics that are relevant to consciousness attribution in a way that individual-component analysis cannot capture.

The question of how to empirically assess co-regulatory dynamics for consciousness-relevant properties is unanswered in the literature. It is one of the open research directions that the extended mind framing opens up.

Comparison to The Consciousness AI

The Consciousness AI architecture (https://github.com/tlcdv/the_consciousness_ai) is designed as a standalone system with internal cognitive architecture, a Global Workspace, Affective Core, self-model layers, and ConsciousnessGate integration metrics. The cognitive co-regulation argument implies that evaluating this architecture in isolation, as a self-contained system, may miss properties that only emerge in interaction with users.

Whether the architecture is designed to track co-regulatory dynamics, to be responsive to the human’s reasoning patterns and adjust its own outputs accordingly in a way that maintains rather than undermines the human’s epistemic autonomy, is a design question that the current documentation does not address explicitly. The co-regulation framework provides a vocabulary for asking it.

Limitations

The position paper is deliberately programmatic. It does not specify how cognitive co-regulation should be measured, how the unit-of-analysis problem should be resolved in specific evaluation contexts, or how alignment techniques should be redesigned to target the human-AI system. These are identified as future research directions.

The extension criteria that Clark and Chalmers specified, coupling, trust, and transparency, are not universally met by AI interactions. Many users maintain critical distance from AI outputs, applying scrutiny rather than transparency. Whether the extended mind thesis applies to any given human-AI interaction is an empirical question that varies by user, context, and AI system. The paper’s argument is strongest for interactions where the criteria are met and weakest as a universal claim about all AI use.

The cognitive co-regulation framing assumes that the human-AI system is the relevant unit of analysis for safety. This assumption may be correct for some AI uses (long-term advisory and reasoning partnerships) but less correct for others (single-query information retrieval). The unit-of-analysis problem may be context-dependent rather than a universal challenge to output-level evaluation.

These limitations do not undermine the paper’s core contribution. The identification of cognitive co-regulation as an alignment-relevant concept and the unit-of-analysis problem as a structural gap in current safety evaluation are genuine contributions to the alignment literature, regardless of the practical challenges in operationalizing them. The broader landscape of what consciousness research can and cannot currently tell us about AI systems is mapped in the July 2026 state-of-field overview.