> Paper: Toward Enactive Artificial Intelligence > Authors: Richard S. Sutton & Banafshee Rafiee > Published: 2026-05-22, arXiv:2605.24238 > Reference translations: 36Kr / Welian Chinese coverage
1. A Harsh Truth
Richard S. Sutton, father of reinforcement learning and Turing Award winner, released a paper in May 2026 with a calm title: *Toward Enactive Artificial Intelligence*.
But the content is anything but calm.
Sutton and his co-author Banafsheh Rafiee start from the foundations of cognitive science to tell the AI industry a harsh truth:
The path we are currently on is fundamentally wrong.
It is not that models are not big enough, data not plentiful enough, or compute not powerful enough. The problem is: we have misunderstood what 'intelligence' is.
From symbolic AI to deep learning, from CNNs to Transformers, from GPT to Claude, all mainstream AI shares the same underlying assumption — Passive Representationalism:
> Intelligence = receive input → internal processing → generate representation → output action
In this picture, the brain (or neural network) is a central processing unit that builds an 'internal model' of the external world and then reasons and decides based on that model. The goal of perception is to 'copy' reality as accurately as possible.
Sutton says: no. The world is not a static object waiting to be encoded. The world is dynamic, open, and inexhaustible. No finite internal model can capture all states of the world. The world is not a set of features but a space of possibilities that unfolds with the agent's actions, context, and interaction history.
So the most reliable, up-to-date, and richest information is not inside the agent but in the world itself.
This is exactly what roboticist Rodney Brooks meant by his famous line: 'The world is its own best model.'
2. What Is Enactive Cognition?
'Enactive Cognition' comes from enactivism in cognitive science. Its core claim:
> Cognition is not an internal copy of a pre-given objective world, but a product generated in the interaction between an embodied subject and its environment.
Perception is not something that 'happens to' an organism; it is something the organism actively does.
Sutton and Rafiee extract four key concepts from this framework that they consider most relevant to AI:
1. Experience
In enactive cognition, experience is not data. Experience is continuous, real-time, mutually influential interaction between agent and environment.
The world is not a fixed feature set, and the agent is not a passive receiver. The agent co-constitutes the world through action. Because the world exceeds any finite description, the agent must continuously interact with it, adjusting actions, calibrating expectations, and refining understanding through real-time feedback.
What does this mean for AI?
| AI Paradigm | View of Experience | Problem | |---|---|---| | Symbolic AI | Almost no concept of experience | Rules disconnected from the real world | | Supervised learning | Experience = human-labeled datasets | Agent does not collect its own data; learning is one-shot | | Reinforcement learning | Experience ≈ ongoing interaction + feedback | Close, but lacking skillfulness, normativity, and embodiment |
The 'Era of Experience' proposed by Silver and Sutton (2025) points in this direction: data must keep improving alongside agent capability, and only the agent's own experience can achieve that.
2. Action-Perception Inseparability
Enactive cognition emphasizes that action and perception are inseparable. Perception neither precedes action nor merely guides it; they unfold together in continuous interaction with the environment.
What is perception? It is mastery of sensorimotor contingencies: understanding how bodily movements produce sensory changes. What you see depends on how your eyes move; what you hear depends on how your head turns; what you feel depends on how your hands move.
Perception is not passive input reception but skillful activity. The agent acts to reveal, stabilize, or make sense of perceptual information — moving the head to resolve ambiguity, tilting the body to hear better.
This is a feedback loop (Merleau-Ponty's 'intentional arc'): the more refined the understanding, the more precise the response; the more precise the response, the more refined the understanding. Agents naturally tend toward 'maximal grip' — a more stable, clearer, more environment-aligned state.
What does this mean for AI?
In mainstream AI, perception is understood as preceding action: extract information first, then decide. Video generation models learn 'intuitive physics' through pure observation, but this is tracking regularities, not understanding. When a traffic light fails and action is needed to change the situation, such systems are helpless.
Early AI explored perception-action coupling — the Pengi system (Agre & Chapman, 1987), Ballard's active vision (1991), Brooks' behavior-based robotics (1991). But modern large models have returned to 'passive encoding' mode.
3. Autonomy
Autonomy is central to enactive cognition. The agent is not passively responding to external stimuli but is a self-organizing system: its perception is shaped by its own goals and needs.
Autonomy stems from autopoiesis: the agent is a self-producing, self-maintaining system that actively sustains its own organization. Perception therefore reflects what is relevant to the agent's ongoing survival, not everything that 'objectively exists'.
The world is not a neutral array of features but a field of significance relative to the agent's goals and needs. The environment is not 'what is' but 'what matters': what supports or threatens the agent's continued self-maintenance.
This brings normativity: agent-environment interactions are not neutral — they can succeed or fail, fit or not fit the situation. The agent continuously adjusts behavior based on such evaluation. And the normative standard is not externally imposed; it emerges from the needs of the agent's own organization.
What does this mean for AI?
| AI System | Self-evaluation | Standard Source | Degree of Autonomy | |---|---|---|---| | Supervised learning | None | External labels | None | | Symbolic planning | No continuous evaluation | External preconditions | None | | Control systems | Continuous deviation evaluation | External goal/cost function | Partial | | Reinforcement learning | Evaluates trajectories via experience | External reward function | Partial | | Enactive ideal | Evaluates via its own organization | Its own maintenance needs | Full |
RL is a major step forward: the agent evaluates behavior through experience, answering 'is this behavior good in the long run?' rather than 'how close am I to the goal now?'. But the evaluation standard is still an externally given reward function.
4. Embodiment
Enactive perception is fundamentally embodied: the body's shape, structure, and capabilities shape how perception works.
Sensorimotor contingencies are not abstract mappings but are grounded in the body's specific capacities: what movements it can make, how it can explore, how its sensory systems are organized. Morphological factors — joint structure, muscle distribution, sensory placement — constrain and determine the space of possible sensorimotor contingencies.
Embodiment also determines what counts as perceptually relevant. Gibson's concept of 'affordances': features in the environment are 'graspable', 'climbable', 'traversable' — only relative to an agent's bodily capabilities. Without a body capable of those actions, these distinctions do not exist.
The body is not an optional add-on but the condition of possibility for perception (Merleau-Ponty).
What does this mean for AI?
Mainstream AI, even in multimodal training, often processes perception in a 'disembodied' way: learning mappings from input to internal representation without sensorimotor engagement or bodily structure. Perception is reduced to pattern recognition on static datasets (Bender et al., 2021).
Embodied RL and robotics often treat the body as an external constraint rather than a constitutive principle of cognition. Modular architectures separate perception, planning, and control, with the body merely an interface executing pre-computed policies. Heavy reliance on simulation and offline training further detaches learning from the full variability and constraint structure of real sensorimotor interaction.
Soft robotics and morphological computation research shows body structure can play an active computational role (Pfeifer & Bongard, 2006; Rus & Tolley, 2015), but these approaches remain marginal in mainstream AI.
3. RL: Closest, But Not Enough
Sutton, as the father of RL, is both affirming and critical:
Structural resonance between RL and enactive cognition:
- Emphasis on action
- Agent-environment interaction
- Feedback-driven adaptation
- Agent-centric evaluation
- No self-evaluation: LLMs do not know whether what they say is correct
- Fully external standards: correctness defined by human feedback (RLHF) or labels
- No ongoing interaction: training is one-shot, not continuous interaction with an environment
- Disembodied: pure text processing, no body, no sensorimotor contingencies
- Bigger models ≠ better understanding
- More data ≠ genuine experience
- Stronger reasoning ≠ autonomous norms
- What constitutes a higher degree of 'action-perception inseparability'?
- What benchmarks can test skillful engagement rather than pattern replication?
- What does 'self-maintenance' mean for an artificial agent? Battery state? Hardware integrity? Acquired capabilities?
- What counts as 'embodiment' in AI? A robotic body? Or a software agent with tools/APIs?
- Paper: https://arxiv.org/abs/2605.24238
- 36Kr coverage: https://eu.36kr.com/en/p/3835601406997641
- Welian coverage: https://www.welian.com/news
But RL still lacks key elements: 1. Evaluation is still externally defined: the reward function is externally given, not derived from the agent's own organization 2. Action-perception inseparability is not fully realized: perception is still usually understood as preceding action 3. Embodiment is treated as an implementation detail: rather than a constitutive condition of perception
Key gaps:
| Dimension | RL Today | Enactive Ideal | |---|---|---| | Experience | Ongoing interaction + feedback | Skillful, normative, embodied | | Perception-action | Perception precedes action | Perception as action, inseparable | | Autonomy | External reward function | Norms from self-maintenance | | Embodiment | Body = interface | Body = condition of perception |
4. The Deep Critique of LLMs
The paper's critique of large language models is especially sharp:
> 'Although LLMs are trained with self-supervised objectives (such as next-token prediction), they effectively learn by imitating patterns in human-generated data and cannot evaluate their own outputs without external signals.'
What is LLM 'learning'? Extracting patterns from static corpora prepared by humans. It makes no judgment of right or wrong; it only predicts 'what a human would say'. It cannot assess whether its own answers succeed — unless someone tells it.
This is the exact opposite of what enactive cognition requires:
5. Why Does This Matter?
Sutton is not writing a philosophy paper. He is pointing to an engineering problem:
If we want true AI — AI that can understand, adapt, and make judgments in novel situations — we must rethink the foundational architecture.
The current bottleneck in AI is not scale, but paradigm.
The direction enactive cognition points toward: 1. Continual online learning: not train-once-deploy, but continuously interacting with the environment and updating understanding 2. Closed-loop perception-action: perception is not an input stage but a process co-constituted with action 3. Self-maintained norms: evaluation standards arising from the agent's own organization, not external rewards 4. Embodied intelligence: the body is not optional but a necessary condition of intelligence
6. Limitations and Future Work
The paper itself acknowledges it is a conceptual paper, not yet 'operationalized':
> 'A key future direction is making these concepts more precise and testable.'
Open questions include:
7. Closing: Sutton's Turn
The most striking thing about this paper is not that it proposes a new algorithm, but that it comes from Sutton himself — one of the most steadfast founders of reinforcement learning.
He is saying: RL is a step in the right direction, but only a step. If we stop at RL, we stop at an incomplete picture of intelligence.
True intelligence is not the ability to predict the next token, nor the ability to maximize external reward, but rather —
> The capacity to generate one's own experience, sustain one's own organization, understand what matters to oneself, and act accordingly, through ongoing interaction with the world.
This is not a matter of tweaking model architectures. It is a matter of redefining what AI is.
---
References: