From Semantic Ablation to Attentional Ablation

What Damasio Said About the Machine

On February 19, 2026, Alain Damasio was invited on the show C ce soir on France 5. A novelist, often described as a techno-critic, he shared that he works with Claude on his writing and said he was “astounded by its creative capabilities on imaginary worlds.” The machine, he said, was “almost at the same level as (his) craft.” This statement fueled debate and criticism against Damasio, who was accused of having “turned his coat.”

On this first aspect, I believe the reading of his statement is simplistic. Being techno-critical does not mean being technophobic or techno-resistant. In “La Vallee du Silicium” (Silicon Valley), Damasio criticizes technology and its excesses with perspectives that can be compared to Jacques Ellul in certain respects (p.61, Damasio writes: The law of least effort is anthropological, echoing Ellul’s observation about the pursuit of efficiency in all things within the technological system, for example).

However, techno-criticism is not a rejection of technology. Ellul specifies this in his works: technology is ambivalent and all progress comes at a cost (see notably J. Ellul - The Technological Bluff). That said, one should probably add that, beyond these aspects, technology is a pharmakon, both remedy and poison, which captures this ambivalence well and seems to me a (good) reason to neither blindly accept nor wholesale reject technologies.

What the Machine Does: Produce or Create?

I do not think Damasio has turned his coat. He is simply stating an observation, a sensitivity in the face of generative AI outputs. He is even expressing his own sensitivity. The conclusions drawn from this observation can be debated, but the criticism I have seen on the subject regarding a supposed “misunderstanding” by Damasio seems excessive and above all not very relevant, because it relies on a debatable distinction: generative artificial intelligence produces, it does not create. Humans, on the other hand, create. The opposition may seem pertinent insofar as it reassures us, but it does not withstand scrutiny. I can even say that this semantic division is clumsy and absurd, as I will try to explain below.

Let us start with the dictionary.

“Produce”: To bring into existence, naturally or otherwise, what does not yet exist. “Create”: To give existence to something.

The two definitions overlap almost perfectly. The opposition between the two terms therefore rests on a semantic misconception. If I have to reformulate the argument cited, I would write that content generated by AI is the result of statistical calculations, therefore degraded content, an ersatz. There is thus no “real” creation.

But this reasoning demands proof that is never really provided, because to go in this direction, one would need to lay out the characteristics to be considered. What would be the objective characteristics that distinguish a “real” creation from a “mere” output of generative AI? The question remains without a precise answer because the answer is not objective. It is a matter of feeling, and that is why using “produce” and “create” as antonyms reveals an absurdity, in my view.

Let us do an exercise to illustrate my point. Here are five reformulations of the same sentence:

1. What distinguishes this sentence?
2. How does this sentence stand out?
3. What is the peculiarity of this sentence?
4. What makes this sentence singular?
5. What differentiates this sentence from the others?

The first is human. The four that follow come from a language model that reformulated it. Which one betrays its origin?

I believe that what we are trying to detect is not a property of the text. It is the reader’s sensitivity that ultimately qualifies what they “consume.” You can love Picasso or hate him. That does not negate the status of the Demoiselles d’Avignon as a work of art. The Theatre d’opera spatial by Jason Michael Allen, generated by AI and awarded in a digital art competition, is it a work of art or a human production assisted by technology? The answer varies according to individual sensibilities.

The “Damasio case” mentioned above is actually a way of criticizing someone because they were moved by a text produced by AI. It is therefore criticizing their sensitivity. It is turning against oneself exactly what one claims to defend: the human capacity to feel, to be affected, to react. If Damasio is “thrilled” by prose generated by a machine and others are unmoved by it, that is not a misunderstanding of the technology, nor is it stupidity: it is a divergence of taste. I am astonished to see these paragons of ethics and virtue slide so easily onto the terrain of the subjective.

Beautiful, ugly, moving, I like it, I don’t like it: these judgments are by nature subjective and will always be contradicted by divergent sensibilities. The hierarchy we believe we can impose on content generated by generative AI is in reality imposed on the sensibilities of others.

I do not, however, want to consider generative AI as beyond all criticism. This criticism deserves to be conducted, but on more “solid” ground. If we stay on the terrain of “artistic” sensitivity and creation, two aspects seem particularly important to consider.

A Confiscation of Judgment

The first is automation bias. This phenomenon refers to the tendency to trust automated systems at the expense of one’s own judgment.

Applied to creation, it produces a perverse effect: the user who regularly relies on a generative model begins to delegate not only execution but also the very direction of their “creative” thinking.

For example, the user will validate formulations they would not have chosen, adopt structures they would not have thought of or would not have built.

Progressively, and unconsciously, the user integrates the machine’s statistical preferences as their own. These effects are insidious because they slowly and almost imperceptibly reconfigure cognitive and creative faculties.

Average as the Goal

The second is the uniformization of thought. The risk here is linked to the standardization resulting from statistical averaging. The author/creator’s singularity will be less and less expressed through these “normalizing” devices. A language model is trained on billions upon billions of existing content: it produces content that will resemble what has already been created, weighted, where applicable, by what the designers of the generative model deemed satisfactory. In other words, it produces content that tends toward the average. Claudio Nastruzzi proposes the term “semantic ablation” to describe this “average” content.

When millions of people use the same models to write, draft, reformulate, and imagine, a convergence movement can begin without users’ awareness. What constitutes the originality of writing or creation finds itself marginalized by the statistical weight of homogeneous or homogenized mass production. This uniformization results from the erosion by standardization produced by generative models.

The Consequence of Semantic Ablation

This concept of semantic ablation deserves closer attention. It refers to the phenomenon by which a text submitted to a generative model to be “polished” does not improve. On the contrary, it loses its substance, its singularity, and in a certain way its quality. The model will identify zones of “high” semantic density, meaning passages where one finds atypical formulations, imperfections or rough edges that characterize, according to some, human thought.

These imperfections are then replaced by the most probable word sequences. What was raw thought with rough edges becomes, for the reader, a smooth surface without semantic or structural friction. It is in these cases that one “feels” the content was drafted by generative AI. Something clean but emptied of its originality in the legal sense of the term: the expression of free and creative choices that reflect the author’s personality.

This loss is therefore not merely aesthetic and semantic. It becomes attentional. Frege made an observation that we rediscover here in a new sense:

Conscious attention is directed only at what is sensible. By giving thought a sensible form, language allows attention to focus on it, and thought to direct and control itself.

If language and therefore also writing allow us to give thought content a “rough” externalization so that attention can grip onto it, then the semantic ablation of generative models, through this effect, creates an attentional ablation. A text with a regular rhythm, with vocabulary and formulations that are probable and mechanical, will no longer offer this “sensible” grip. Attention passes through it without encountering obstacles, and without these obstacles, there is little or no reflective thought.

Can we consider that semantic ablation merely “degrades” a text by making it “less good”? We can measure this phenomenon. By submitting a text to successive cycles of AI “treatment,” we progressively reduce lexical diversity. Like a photocopy copied ad infinitum that ends up no longer resembling the original. Generative models tend to reduce the “entropy” of the original text.

In any case, I believe that this semantic ablation produces a text that fails to fulfill the very function of language: allowing thought to turn back on itself. By smoothing language, we remove from thought its points of grip.

It is on these grounds that the critical reflection on Damasio’s remarks deserves to be extended. Discussing the sensible, the felt response to a generated text, will always lead to dead ends due to judgments and the subjective character of personal feelings.

On the other hand, properly understanding what repeated use of these tools concretely does to the thinking of those who use them is more than necessary. The modeling choices embedded in the parameters of these systems are not neutral. They steer outcomes, sometimes far beyond what users perceive, and risk progressively reconfiguring the creative imaginaries they claim merely to assist.

This is yet another demonstration of the pharmacological character of this technology. This technology, like all others, is a pharmakon. A pharmakon whose inherent logic is to dissolve, in repeated and imperceptible doses, the roughness upon which thought takes hold. This is not, however, a reason to reject this technology. I still see in it the inextricable ambivalence observed and the indispensable necessity of therapeutic practices (here & there).