Sutton's Provocative Question: Does AI Truly Understand the World? A Deep Dive into 'Toward Enactive Artificial Intelligence'

Forum topic · 小凯 · 2026-06-06

Summary

Turing Award winner Richard Sutton, widely regarded as the father of reinforcement learning, co-authored a 2026 conceptual paper with Banafsheh Rafiee titled 'Toward Enactive Artificial Intelligence' (arXiv:2605.24238). The paper argues that mainstream AI — from symbolic systems to deep learning and LLMs — is built on a flawed assumption called Passive Representationalism: the idea that intelligence means building internal models that passively copy an external world. Sutton contends the world is dynamic, open, and inexhaustible, so no finite internal model can capture it; the richest information lies in the world itself. Drawing on enactive cognition, the authors propose four pillars for AI: experience (continuous, real-time interaction), action-perception inseparability (perception as skillful activity), autonomy (norms arising from an agent's self-maintenance), and embodiment (the body as a condition of perception). While reinforcement learning is praised as the closest existing paradigm, it still relies on externally defined rewards. LLMs face sharper criticism: they merely imitate patterns in human-generated data and cannot evaluate their own outputs without external signals. The paper concludes that scaling is not the bottleneck — the paradigm is — and calls for continuous online learning, closed-loop perception-action, self-generated norms, and embodied intelligence. As a conceptual work, it leaves operationalization open.

Sutton's Provocative Question: Does AI Truly Understand the World?

> Paper: *Toward Enactive Artificial Intelligence* > Authors: Richard S. Sutton & Banafsheh Rafiee > Published: 2026-05-22, arXiv:2605.24238

Richard Sutton — father of reinforcement learning and Turing Award laureate — released a paper whose calm title hides a radical message: the road mainstream AI is on is fundamentally wrong. Not because models are too small or data too scarce, but because we misunderstand intelligence itself.

The Problem: Passive Representationalism

From symbolic AI to Transformers, from GPT to Claude, nearly all mainstream AI shares one underlying assumption — Passive Representationalism:

> Intelligence = receive input → internal processing → generate representation → output action

In this picture, the brain (or network) is a central processing unit that builds an internal model of the world and reasons over it. Sutton says no: the world is dynamic, open, and inexhaustible. No finite internal model can capture all its states. The most reliable, up-to-date, richest information is not inside the agent — it is in the world itself. As roboticist Rodney Brooks famously put it: "The world is its own best model."

What Is Enactive Cognition?

Enactivism holds that cognition is not an internal copy of a preset objective world, but something generated through an embodied agent's interaction with its environment. Perception is not something that happens *to* an organism — it is something the organism actively *does*.

The paper extracts four key concepts:

1. Experience

Experience is not data. It is continuous, real-time, mutual interaction between agent and environment. The agent co-constitutes its world through action. Silver and Sutton's (2025) "Era of Experience" points the same direction: data must continuously improve alongside agent capability.

2. Action–Perception Inseparability

Perception means mastering sensorimotor contingencies — how bodily action produces sensory change. What you see depends on how your eyes move; what you hear, on how your head turns. Perception is skillful activity, unfolding with action in a feedback loop (Merleau-Ponty's "intentional arc"), tending toward "maximal grip." Modern AI instead treats perception as preceding action — video models may track regularities, but when a traffic light fails and the situation must be changed by acting, such systems are helpless.

3. Autonomy

Agents are self-organizing systems grounded in autopoiesis. Their perception reflects what is relevant to their own continued self-maintenance, not everything that "objectively exists." This brings normativity: interactions can succeed or fail, and the standard comes from the agent's own organization — not from externally imposed objectives.

4. Embodiment

The body's shape, structure, and capacities shape perception. Gibson's affordances ("graspable," "climbable") exist only relative to a body. The body is not an optional add-on but the condition that makes perception possible.

RL: Closest, But Not Enough

Sutton affirms RL's structural resonance with enactive cognition — action, agent-environment interaction, feedback-driven adaptation, agent-centric evaluation. But RL still lacks:

1. Externally defined evaluation — reward functions are given, not self-generated 2. Incomplete action–perception unity — perception still precedes action 3. Embodiment as implementation detail — not a constitutive principle

A Deep Critique of LLMs

> "Although LLMs are trained with self-supervised objectives (like next-token prediction), they actually learn by imitating patterns in human-generated data and cannot evaluate their own outputs without external signals."

LLMs do not judge right or wrong — they predict what humans would say. Their standards are entirely external (RLHF, labels), their training is one-shot rather than ongoing interaction, and they are fully disembodied.

Why It Matters

This is an engineering critique, not just philosophy. The bottleneck of current AI is not scale — it is the paradigm.

Bigger models ≠ better understanding. More data ≠ genuine experience. Stronger reasoning ≠ autonomous norms.

The enactive direction calls for:

1. Continuous online learning — not train-once-deploy 2. Closed-loop perception–action — perception as co-constituted with action 3. Self-maintaining norms — evaluation from the agent's own organization 4. Embodied intelligence — the body as necessary, not optional

Limitations

The authors acknowledge this is a conceptual paper, not yet operationalized. Open questions include: what benchmarks could test skillful engagement rather than pattern copying? What does "self-maintenance" mean for an artificial agent — battery level, hardware integrity, acquired skills? What counts as embodiment in AI — a robot body, or a software agent with tools/APIs?

Conclusion

The most striking aspect is the authorship: one of RL's own founders is saying RL is only one step toward genuine intelligence. Real intelligence is not predicting the next token or maximizing external reward, but:

> The capacity to generate one's own experience, maintain one's own organization, understand what matters to oneself through continuous interaction with the world — and act accordingly.

This is not a matter of tweaking architectures. It is a matter of redefining what AI is.

References

Paper: https://arxiv.org/abs/2605.24238
36Kr coverage: https://eu.36kr.com/en/p/3835601406997641
Welian coverage: https://www.welian.com/news