Who Thinks When I Think? Reading Benoît Bergeret
What if the problem with artificial intelligence were not that it thinks in our place, but that we can no longer tell what stems from its influence and what stems from our own intellectual activity?
In Qui pense quand je pense ? (Who thinks when I think?), Benoît Bergeret examines this silent, or at the very least discreet, transformation of the conditions under which our representations, our reasoning and our judgements take shape.
The author argues that by delegating a growing share of our intellectual activities to algorithmic systems, we may gradually lose control over the conditions in which our thinking is formed. This development would foster the emergence of a “cognitive feudalism”, characterised by the concentration of the power to structure our representations in the hands of a few technology players.
This reflection invites us to move beyond the usual debates on model performance, hallucinations or productivity gains, and to examine a more fundamental question: the power that technical infrastructures exert over our intellectual activity.
I largely share this concern (I am myself working on the concept of algorithmic logocracy). I am nonetheless more reserved about some of the consequences the author seems to draw from it. His analysis sometimes seems to me to presuppose a passive user and an overly linear conception of the relationship between humans and generative and algorithmic devices. I think that being influenced in one’s thinking does not necessarily mean being dispossessed of one’s capacity to think.
From “Code is law” to “Model is mind”
One of the book’s proposals appears on page 25. Bergeret takes up Lawrence Lessig’s famous formula, “Code is law”, and proposes an update: “Model is mind”.
His intuition is interesting. Where computer architecture regulated behaviour by organising the possibilities for action, artificial intelligence models would now act on the conditions under which our representations are formed. They select information, organise answers, suggest categories and propose syntheses. In doing so, they help determine what becomes visible, relevant or simply conceivable for their user.
This distinction between governing behaviour and governing representations is, in my view, one of the contributions of this essay. The influence of technical systems is not only exerted at the moment of deciding or acting. It also comes into play in the way we formulate problems, which themselves necessarily result from our representations.
There is, however, a difficulty with the formula “Model is mind”. It may suggest that the model determines what we think from, as if intellectual activity consisted in receiving a set of previously organised representations and then exercising judgement on them.
This approach strikes me as largely insufficient, because it assumes that thinking boils down to a succession of representations received, assessed and possibly accepted. Thinking is also, and notably, made up of questioning, confrontation, refusal, experience and reformulation.
The user does not necessarily confine themselves to examining a system’s answers. They can challenge its premises, change the question, multiply perspectives, or use the model to explore ideas they already have or would not otherwise have reached. The model intervenes in thinking, but it is neither its sole origin nor its necessary horizon.
The risk of dispossession without resistance
Bergeret describes an insidious mechanism that I have also considered from a professional angle here: a cognitive dispossession that would not be felt as such by the person undergoing it.
The service rendered by the machine would gradually mask the user’s loss of autonomy. When an answer is immediately available, coherent and convincing, it can become tempting to give up the effort needed to build it or question it. Fluency then becomes a possible substitute for reflection. But this development is not a necessary consequence of using a generative model. It corresponds rather to a certain way of using it.
It seems useful to me to distinguish three situations:
- Cognitive delegation, when the user entrusts the system with carrying out an intellectual operation.
- Cognitive dependence, when they gradually lose the ability or the habit of carrying out that operation without assistance.
- Epistemic dispossession, when they no longer sufficiently master the foundations of their reasoning to assess its validity or call it into question.
These situations may be linked, but they are not equivalent. Delegating an operation does not necessarily lead to dependence, and dependence does not always translate into the disappearance of critical judgement.
These distinctions echo the work of Bernard Stiegler, who reminds us that human history is inseparable from an ongoing process of exteriorising its functions in technical devices. From the tool to writing, from the printing press to digital technologies, humans have never stopped exteriorising their gestures, their memory and their knowledge, and then reappropriating them in new forms. Exteriorisation is therefore not, in itself, dispossession: it is, in a certain way, one of the conditions for forming, preserving and transmitting knowledge. Stiegler captures this ambivalence through the notion of pharmakon: technology can just as well foster individuation and the acquisition of new knowledge as it can bring about, when it short-circuits the processes of learning and reappropriation, a proletarianisation characterised by their loss.
A user can delegate a synthesis while retaining in-depth knowledge of its subject. They can ask for a counter-argument precisely in order to subject their own convictions to intellectual resistance.
These algorithmic systems take part in this new exteriorisation of our intellectual activities. This transfer must be analysed from the angle of reappropriation. Put differently: does this exteriorisation give rise to reappropriation or to a progressive disappropriation?
We should also beware of the tendency to limit the reflection to an alternative between what we do and do not delegate to the machine. We should also think about what we do with the results obtained, about the knowledge we retain, and about the (new) capacities that this interaction (possibly) allows us to develop.
A thinking that has never been sovereign?
On page 45, Bergeret introduces the notion of “cognitive sovereignty”. It would notably require being able to understand the model’s limits, know the criteria guiding how it works, audit its decisions and retain the possibility of turning away from it. These elements are relevant from the standpoint of AI systems governance. They are among the reasons that led me to use the notion of generative devices instead of generative AI.
This notion of “sovereignty” should nonetheless be questioned through a more general question: have we ever enjoyed genuine cognitive sovereignty?
Our thinking takes shape in an environment made of language, institutions, traditions, collective representations and technical devices. A school textbook selects the knowledge it passes on. A search engine ranks content. A newspaper constructs a particular representation of current events. A teacher or an author steers our understanding through their choices and categories. None of the content we consume or use is either perfectly transparent or entirely neutral.
This obviously does not mean that all forms of influence are equal. The concentration of technical power, the opacity of models and the possibility of large-scale intervention introduce specific risks. But cognitive sovereignty cannot reasonably be defined as the absence of external influence. Besides creating a condition we do not impose on other content, it also seems impossible to satisfy. We could nonetheless give this concept the role of designating the individual’s capacity to recognise the influences to which they are exposed and to succeed in transforming their own judgement.
We never think alone
On page 47, the author draws on the work of Lynn Margulis and the notion of the holobiont to recall that living organisms exist in constitutive relationships with other organisms.
He transposes this idea to cognition: we never think alone. This ties in with what I describe above. This proposal invites us to move beyond a representation of thinking as a strictly individual activity that would then be disrupted by the intervention of a machine. If our intellectual activity is always situated in a social, cultural and technical environment, artificial intelligence does not mark the arrival of external influence on a previously autonomous and sealed-off mind. Algorithmic systems are a new modality of mediation.
But this is where I glimpse a tension I cannot manage to resolve: how can one simultaneously assert that the model structures the conditions of our thinking and that our thinking is constituted within an assemblage of multiple influences?
These two propositions are not incompatible, but articulating them requires distinguishing an influence among others from a power that would determine the whole range of possibilities. The first describes a transformation of our cognitive environment. The second risks leading us towards an excessively deterministic conception of technology. If we move past these divergences, the relational approach, close to certain intuitions of John Dewey or Gilbert Simondon, invites us on the contrary to consider that humans and their technical milieu transform each other through their interactions. From this perspective, the subject is neither fully sovereign nor merely passive. They take part in an activity whose conditions and results are built up over the course of the relationship.
Dancing rather than steering
It is ultimately in another metaphor proposed by Bergeret (which I have already discussed) that I find one of the promising leads of this essay.
On page 105, he invites us to consider our relationship with AI as a dance. The image moves away from the traditional representation of a user commanding a perfectly predictable machine. The model proposes a wording. That wording changes our understanding. We reformulate our request. A new answer introduces an unexpected perspective, which in turn changes the direction of our reflection. The relationship no longer simply consists in transmitting an instruction and receiving a result. It becomes a process of successive adjustments.
This metaphor nonetheless has a limit: the model is not a human partner endowed with intentions or an experience of its own. The reciprocity of the interaction therefore remains asymmetrical, and Big Tech has this unfortunate tendency to want to simulate that reciprocity.
This limit still allows us to think of intellectual production as a dynamic activity, in which the result is not entirely determined in advance either by the user or by the system, and it could lead us to reconsider the risk of dispossession. If the model can steer the user’s reflection, the user can also steer, challenge and reorganise the model’s output. The interaction can become a space for exploration and resistance, but it presupposes one essential condition: that the user remains engaged rather than mistaking the production of an answer for the completion of their reflection.
The question is political
These reservations must not obscure the democratic concern expressed by Bergeret. The influence of digital infrastructures on our representations is a major political issue. Models are not neutral. Their design choices, their training data, their optimisation mechanisms and their deployment arrangements shape the content they produce.
Those who control these infrastructures thus hold significant, even considerable, power over informational and cognitive environments. The proposed concept of “cognitive feudalism” makes it possible to name this imbalance. But we should, it seems to me, avoid concluding that the user is condemned to passively endure the system, because the existence of a power to structure representations does not, on its own, demonstrate that its victims have lost their capacity to resist. The best proof is the growing number of texts of all kinds on the subject.
Bergeret himself opens up this perspective when he considers, among other things, using technologies to organise disagreement and improve the conditions of democratic deliberation. Technology is therefore not necessarily the instrument of domination. It can also help create the conditions for contesting it.
Thinking with AI without giving up thinking
Qui pense quand je pense ? shifts the debate on artificial intelligence towards the democratic field. The point is no longer only to ask what models can do, but to examine how using them can transform us individually, and therefore collectively.
I mainly take away two intuitions: technical infrastructures take part in forming our representations, and their governance is inseparable from a democratic stake.
My reservations concern rather the conception of intellectual activity that sometimes seems to underpin the diagnosis. I do not believe that human thought is an autonomous territory that artificial intelligence would gradually colonise. It is already the product of a multiplicity of relationships, influences and mediations. Fighting to preserve an (impossible) purity of thought is therefore a lost cause from the outset, and on this point I agree with Bergeret. What remains is to understand and foster the conditions of interaction with these algorithmic systems that contribute to enriching our thinking. To move beyond these eternal oppositions and think instead by composition (see the pharmacological approach), we need to identify the conditions under which thinking with a machine allows us to think differently.
Reference: Benoît Bergeret, Qui pense quand je pense ? Péril et riposte démocratiques à l’ère de l’intelligence artificielle, Chemins de tr@verse, coll. Débords, 2026, 144 p.