On Generative Devices

For some months now, I have been using the notion of technical devices, or more generally the notion of device, as a substitute for the excessive and overused term generative artificial intelligence when speaking of conversational agents. In my last text, I became aware of this usage and chose to adopt the term “generative device” to speak of conversational agents and other devices embedding generative artificial intelligence. I would like to explain here my approach and the reasons behind it. This exercise also allowed me to question the value of this qualification and confirmed me in the idea that it designates more accurately what we are talking about when we evoke these machines using so-called generative artificial intelligence techniques.

A Mere Question of Vocabulary?

First, I think we can start from the observation that there is today an excessive anthropomorphism surrounding these devices. This anthropomorphisation is liable to create consequences that are sometimes deleterious but above all insidious. The words we use to qualify a technique also influence the way we perceive it, the way we interact with it and, ultimately, the trust we place in it.

In her Atlas of AI, Kate Crawford pointed out the existence of two myths on this subject. First, the one according to which non-human systems are analogous to the human mind, and second, the one where intelligence would have an independent existence. Crawford continues:

These mythologies are particularly strong in the field of artificial intelligence, where the idea that human intelligence can be formalised and reproduced by machines has been axiomatic since the middle of the twentieth century

Regarding myth, it seems useful to specify that this term covers a “construction of the mind, fruit of the imagination, having no link with reality, but which gives confidence and incites to action”.

There is therefore a performative side to myth, and the vocabulary we use, (un)consciously, only contributes to the subsistence of this myth.

The tendency to use a vocabulary proper to the individual when speaking of research in artificial intelligence has not diminished with current devices. Today, we continue to refer to hallucinations, to biases, to reasoning or to the effort (Claude Opus 5) and to the reflection effort (in ChatGpt 5.6 Sol for example) of the device. This vocabulary is largely borrowed from psychology or the cognitive sciences and contributes to sustaining this double myth identified by Crawford. To put an end to this contribution to myths, we should use different and far less connoted terms to speak of these devices. My reflection on the vocabulary used and to be used also leads me to a broader analysis of the impact these devices may have on the user and the society surrounding them. A way of widening the focus, in other words, which will allow me to traverse the subject with further reflections.

The False Appearance: An Interactional Anthropomorphism

The conversational interface of these devices reinforces this phenomenon further. We exchange with these devices as with an interlocutor. The device answers us using the first person, writes to us that it “thinks”, “understands” or “proposes”. The interaction resembles the one we might have with a colleague, a collaborator or a friend. This is moreover how some presented these devices (the super-intern, for example).

The notion of “generative artificial intelligence” also has an effect on our perception of generated content. There is an overvaluation, often unconscious, of the quality of the generated content because of the qualification we give to the machine. Here lies a danger. The vocabulary used creates a framing that can favour overconfidence, even an automation bias. By qualifying a machine as intelligent, we pre-qualify the result. There is, here again, a performative effect on our perception of the device and of the generated content. The danger is that this overvaluation ends in a tendency to content ourselves with mediocrity as I already wrote in March 2025.

However, the debate on the contours of the intelligence of these devices is not the one I wish to enter. What intelligence is, is a subject and a concept that have been debated for hundreds of years, and even if we could agree on elements intrinsic to intelligence, we would all potentially have additional elements to put forward in considering what intelligence is.

Besides, do we really need to use this term to designate these machines? The answer seems obvious to me: no. Changing vocabulary allows us, on the contrary, to create a certain distance from the object and, perhaps, to recover a form of epistemic vigilance in our interactions with it, as Albert Moukheiber proposes.

Why a “Device”?

I am not the only one to propose other terms to speak of this “machine”. The notion of device deserves an explanation in order to understand its necessity, and the choice of this term is obviously not innocuous.

In its technical acceptation, a device is an “ensemble of elements arranged with a view to a precise end”. This already says a great deal about what we are talking about. The notion of arrangement is important to me. It also allows us to apprehend more precisely the intentions and the finalities of this arrangement. It also allows us to move away from the habitual representation of technical instruments as being neutral, or whose use would determine their qualities.

In using this concept, I cannot pass over the philosophical heritage and the meaning proposed by Michel Foucault as a point of departure.

Foucault’s definition, which one finds regularly, refers to:

*a resolutely heterogeneous ensemble, comprising discourses, institutions, architectural arrangements, regulatory decisions, laws, administrative measures, scientific statements, philosophical, moral and philanthropic propositions — in short: the said as much as the unsaid, such are the elements of the device. The device itself is the network that one establishes between these elements […] By device, I understand a sort — let us say — of formation which at a given moment has had as its major function to respond to an urgency. The device therefore has a dominant strategic function… I have said that the device was of an essentially strategic nature, which supposes that it is a matter here of a certain manipulation of relations of force, of a rational and concerted intervention in these relations of force, either to develop them in such a direction, or to block them, or to stabilise them, to use them. The device is therefore always inscribed in a play of power, but always linked also to one or several bounds of knowledge, which are born from it but, just as much, condition it. That is the device: strategies of relations of force supporting types of knowledge, and supported by them

What interests me is the notion of a network between heterogeneous elements: the said and the unsaid.

The “system cards”, the Constitution proposed by Anthropic, the terms of use and other documents published by their designers today render visible the discursive, Foucault’s “said”. But they do not expose everything that has configured and continues to configure the device. These texts are drafted to reassure and amount to ethical washing (on this subject, see my text: Ethics and compliance: beyond the hype). The events of the summer of 2026 and the recent departure of Jacob Coxon are also part of the narrative of these pyromaniac firefighters, to borrow the expression used by Yannick Meneceur, who also considers that the narratives and terms installed by the industry are installed for its own needs.

Beyond the visible discursive elements, choices remain unsaid and they reside in the architecture, the economic model, the data, the interface, the rules of operation or again the conditions under which the device is made available: all of this is Foucault’s unsaid.

Alongside these non-discursive elements, I also place the imperceptive, which I defined as: “a model of rationality in which a dominant criterion (often efficiency) imposes itself without being discussed, configuring in advance the conditions of the reasonable. Certain options remain formally possible but cease to appear as relevant.” It is perhaps a proposition that exceeds the Foucauldian concept but which, to my estimation, participates in this “networking” of heterogeneous elements and in this dominant strategic function that “links” these different elements.

Kate Crawford, cited above, also envisages the impact of AI systems placed at the service of dominant interests in these terms:

because of the capital required to produce AI at scale and the ways of seeing that it optimises for, AI systems are ultimately designed to serve dominant interests

She continues:

Fundamentally, AI is made of technical and social practices, institutions and infrastructures, politics and culture.

Crawford offers us here a list of the heterogeneous elements of the device. I deliberately prefer this term because it allows several aspects to be captured. Speaking of a device therefore allows us to depart from a naive representation. It is also a way of leaving behind the illusion of the neutrality of technologies and technical objects. On the subject of the absence of neutrality of technique, I refer to my note of September 2025.

The device is therefore not reducible solely to its technical components nor to the explicit discourse that accompanies it. It is inscribed in a far broader ensemble: (technical) choices, institutions, interests, rules, representations, discourses and finalities that have a strategic function.

This conception of the device was updated by Giorgio Agamben in his essay published in 2006 and translated into French in 2014 (G. Agamben, Qu’est-ce qu’un dispositif? Paris: Payot & Rivages) — find here a synthesis of the essay.

Agamben widens the contours of the concept and gives it the following definition:

everything that has, in one way or another, the capacity to capture, orient, determine, intercept, model, control and ensure the gestures, conducts, opinions and discourses of living beings.

From this angle, the device becomes an ensemble that is, to my mind, far too broad, Agamben considering moreover that language is a device. I do not think it is necessary to cross that threshold. Nevertheless, the reflection Agamben proposes around this notion is interesting because it allows us to consider the influence of the device on the subject. Agamben going so far as to consider that:

“what defines the devices we have to deal with in the current phase of capitalism is that they no longer act through the production of a subject, but rather through processes we may call processes of desubjectivation.”

This process of desubjectivation is to be understood alongside the process of subjectivation proposed by Agamben, which seems to me entirely pertinent in our current relations with these devices. On this point, I am working on a text on the subject insofar as this notion appears to me very close to Dewey’s transactionalism, to Simondon’s individuation or again to Foucault’s techniques of the self.

A final aspect that must be specified concerns relations of power. They are often in tension: neither stable nor unstable. Within the framework of the device having a dominant strategic function, this tension of the system resides in its capacity to influence relations of power. The device therefore also has a broader impact, and not exclusively on the individual who uses it.

Computational Modelling

If the device is an arrangement of elements, notably technical ones, I believe it opportune to specify the contours of certain of these elements by applying them to content-generating machines. The interest here is to observe that computational modelling shows that the orientation of the device begins even before the interface and even before generation, in the operation that renders phenomena calculable. I have already written on this subject and I shall approach it with a more technical analysis on the basis of Meunier’s article.

Modelling rests on the notion of machine and on the role of computers as an epistemic mediation. We speak here of mediation, that is, the fact of serving as an intermediary. The epistemic refers to the notion of knowledge, of learning. This mediation is not, however, transparent or neutral. The computer, the search engine or social networks do not merely receive the real in order then to restore it to us as such. For a (real) phenomenon to be processed by a machine, it must first become calculable. It must therefore be the object of a process of computational modelling.

This modelling is selective by reason of its construction in three stages:

1. First intentional or representational: this is the phase of identifying what one wishes to represent of the real.

2. Then functional: this is the translation into mathematical terms and functions of the results of stage 1).

3. Finally physical or material: this is the realisation and execution of the functions of 2) by the machine.

The first phase is very important. It is a phase of selection that aims to translate the identification of the real on the basis of the understanding and perception of the one who carries out this representation. It therefore implies a selection of the elements considered relevant. It is moreover necessarily partial, constructed from certain elements and consequently leaving others aside.

It is in this sense that I join Marcello Vitali-Rosati when he sets out that:

there are only material models; the real is always the result of a modelling. The real is multiple and always mediated; there is no real outside a mediation and every mediation is material. There are therefore no immaterial ideas: everything is matter. The model is not a selection of the real, it is the real itself, such as it results from a determined mediation; but outside this mediation there is nothing at all.

In my view, this partial representation is a fundamental element that is lacking in our understanding of the functioning of the digital and of the processing of the information that passes through it. We tend to consider that what passes through our feeds or what is generated by devices is an “objective” vision of the world, whereas it is a partial vision.

Certain complex subjects can be processed, but this processing engenders, by essence, a selection that will orient the representation and render it partial, incomplete. Of course, there are also subjects that are simply not processed and therefore excluded from the representation proposed, confirming if need be that this representation of the real can only be partial. Computation therefore supposes a series of operations: selecting, representing, formalising, calculating, and so on. This means that the computational layer itself is not a perfectly neutral mediation of the real. This computational modelling does not, however, allow us to identify and explain all the intentions that traverse a device. It allows us more modestly to observe that a part of its orientations is already inscribed in the conditions of its design.

Modelling and Model

I believe it necessary to make an aside in order to conclude this part, so as to avoid a terminological confusion that can be misleading. The computational modelling of which I have just spoken must not be confused with the “model” when one speaks of a “language model” or Large Language Model.

In the first case, modelling designates a process of representation and formalisation allowing something to be rendered calculable. In the second, “model” is the name given to a certain type of computing system resulting notably from a process of training and of specific operation.

This clarification also allows me to close the first part of my proposition: the substantive device allows us to recall that we are faced with an arrangement that cannot be isolated from its conditions of design, from its finalities and from the choices that constitute it.

I therefore conclude with a definition of the device drawing on the elements proposed by the authors cited above:

A device is an arrangement of technical, social, institutional, economic and symbolic elements whose interactions orient conducts while contributing to producing or configuring relations of power.

It remains to understand why this device must be qualified as generative.

Why “Generative”?

In a commentary on Agamben’s essay, Mathieu Quet observes as do others that Agamben’s definition is too general, that every thing would be a device. Quet indicates:

The problem is precisely that we are dealing with an infinity of devices, that not all are equivalent and that only operations of qualification of these different devices in act would allow their stakes to be grasped.

The substantive “device” is too broad to designate on its own the machines in question here. A search engine, a smartphone or a social network can equally be envisaged as devices. It is therefore necessary to characterise the functionality that distinguishes these devices, to distinguish them in act.

The qualifier generative fulfils this function quite simply. I am stating the obvious in indicating that these devices have as their characteristic the generation of content from the instructions addressed to them. They generate text, images, sound, code or other forms of content and can mobilise this capacity in a very great diversity of tasks.

This simplicity nevertheless allows me to broach something that seems to me essential in current debates on the capacities of these devices and their potential to substitute for or replace the human.

The term “generative” does not determine what the machine is. It describes what it does.

This distinction is important because when one qualifies the machine as “intelligent”, it obliges us to enter into an ontological discussion: what is intelligence? Does the machine understand? Does it think? Does it really reason?

Qualifying it as “generative” displaces the discussion towards a functional characteristic.

This consideration is found in the descriptions of the functionalities of these devices. When we consider that “AI is capable of X”, we give it a role, which seems to me the beginning of the slippage. In attributing a function to a device, this does not allow one to infer from it an autonomy, an intentionality or an understanding, as Kate Crawford explains:

AI systems are neither autonomous, nor rational, nor capable of discerning anything whatsoever without extensive and computationally intensive training, thanks to large data sets, with predefined rules and rewards.

Conclusion

Words are not neutral. I am not proposing merely to change words or to deny the technical usefulness of the expression “generative artificial intelligence”, but I consider that it can no longer constitute a sufficient category for thinking the things to which it refers. This also allows a change in the scale of analysis. This change is indeed necessary in the face of the radically trans-formative discoveries (I shall return later to this spelling) with which we are confronted.

Instead of asking ourselves immediately what the machine is, we can begin by examining the device within which we interact, the manner in which it was constituted, the orientations it carries and what it produces — because the words employed to designate our technical objects also participate in the relation we maintain with them and can favour an anthropomorphic reading. Modifying this vocabulary then allows us to constitute a means of keeping the distance necessary to their examination and their critical use.

These notes are part of Duality, my digital notebook. If you would like to receive them directly in your mailbox, you can subscribe to the newsletter.