[YOU] on AI documents, with unusual candor, several moments when this distinction became visceral. The passage where Claude produced text about a Deleuzian concept that “sounded like insight” but broke under examination is the distinction made concrete: the system processed patterns in the training data and generated statistically probable language about philosophy. The author enacted a critical evaluation, bringing an embodied history of engagement with difficult thought to the assessment of whether the generated text illuminated or obscured. The generation was competent. The evaluation was cognitive. And the competence of the generation made the absence of cognitive depth harder to detect, because the very smoothness of the surface was a consequence of the absence of anything to resist it.
The cycle’s most difficult finding—that highly skilled practitioners get more from AI tools than less skilled ones—follows directly from this distinction. The senior engineer’s “expertise” is not a larger database of stored solutions. It is an enacted history of structural coupling with real problems: a felt sense for where systems go wrong, an emotional response to design decisions that violate principles she cannot always articulate, a bodily knowing developed through decades of direct engagement. This enacted expertise is what allows her to make sense of Claude’s outputs—to evaluate them against the demands of a lived situation. The junior engineer lacks not knowledge but enacted understanding, and no amount of AI assistance can substitute for the enactive history that understanding requires.
The distinction originates in Thompson’s synthesis of Merleau-Ponty’s phenomenology with Varela’s biology. Merleau-Ponty had insisted that the body is not a vehicle the mind inhabits but the very medium through which consciousness is constituted: the reaching hand is not executing a motor command from a disembodied processor; the reaching is itself cognitive, discovering the weight and texture of the cup through the act of grasping it. Varela showed that this embodied cognition is continuous with the most primitive biological processes: the autopoietic cell already performs a rudimentary version of the same operation, evaluating its environment in terms of its own needs and acting on that evaluation.
The synthesis produced the enactive approach—the claim that the computational theory of mind gets the order of explanation exactly wrong. The mind does not first represent the world and then act on the representation. The organism acts in the world, and the action constitutes the representation as a secondary product of the engagement. To process a representation is to operate on the output of a prior enactment; it is never itself the enactment. Large language models operate entirely on representations—on the textual residue of billions of prior enactments by living beings—without ever performing the enactive acts from which those representations emerged.
The asymmetry is categorical, not scalar. Processing and enacting are not two ends of a spectrum that scale and architectural sophistication can traverse. They are categorically different kinds of process, distinguished not by degree of complexity but by the presence or absence of a living organism with something at stake. A more sophisticated processor is still a processor. It has not become an enactor by becoming more powerful.
Fluency is evidence of depth only for enactors. When a human expert speaks fluently, the fluency is evidence of understanding—the surface smoothness reflects the depth of enacted engagement beneath it. When a language model generates fluent text, the fluency reflects statistical regularity in training data, and the surface smoothness is entirely independent of any depth of understanding. The sense-making that would validate the fluency is absent. This is why confident AI error is structurally different from confident human error: the human’s confidence is connected, however imperfectly, to an organism that will be affected by being wrong; the model’s confidence is a property of token probability distributions.
Digital structural coupling is real but asymmetric. When a practitioner works intensively with an AI tool, something genuine happens in the relationship—a form of digital structural coupling in which the practitioner’s prompting habits, evaluative dispositions, and cognitive rhythms are shaped by the interaction. But the coupling is asymmetric in the deepest possible way: the practitioner’s side of the coupling is enacted—embodied, stakes-laden, cognitively formative—while the model’s side is processed without any organism being changed by it. The practitioner is genuinely coupled. The model is not.