Richard Sutton, the Turing Award winner often described as a founder of reinforcement learning, and researcher Banafsheh Rafiee outline an alternative approach in a recent paper they call enactive AI. Rather than building models that ingest labeled data and predict patterns, they argue intelligence should arise from an agent acting in an environment and learning from the consequences of its actions.
The four pillars
The authors identify four central principles for enactive AI:
- real experience over labeled datasets;
- perception and action as a unified loop (the perception–action loop);
- autonomy, meaning a system that sets its own success criteria; and
- embodiment, the role of physical action and interaction with an environment.
They contend that current research closest to this view is reinforcement learning, but that it still falls short on all of these counts.
Diagnosis versus practical roadmaps
The paper delivers a sharp diagnosis of what is missing in mainstream approaches, but its prescriptions are largely directional rather than procedural. Autonomy is highlighted as the hardest problem: building a system that defines its own goals or success metrics, rather than optimizing a reward function specified by humans, remains an open challenge with no established engineering solution.
Embodiment presents a different difficulty. An agent that learns by physically acting gathers data at the pace of the real world — a far slower and more constrained process than the way large language models have been trained by ingesting vast amounts of internet text in a matter of months. The paper makes clear what elements are absent from current systems but does not provide concrete methods for constructing them.
Practical implications
Enactive AI offers a conceptual framework that may better capture aspects of natural intelligence: learning from action, the significance of a body and environment, and the importance of self-generated goals. At the same time, Sutton and Rafiee’s critique underlines that conceptual clarity alone does not replace the need for practical engineering advances. In short, enactive AI may point toward a meaningful destination, but today there is no established path to get there.
Conclusion
The work by Richard Sutton and Banafsheh Rafiee stimulates discussion about the foundations of artificial intelligence and sharply frames the limits of current models. Their four pillars serve as a useful agenda for future research, yet the concrete techniques and engineering practices required to realize enactive AI remain to be developed.
