Categorial institution

The scene plays out in countless meetings. Someone rules: “ChatGPT is only a tool.” Another corrects them: “an agent”; the lawyer frowns, a third speaks of a “working partner.” Each believes they are refining a vocabulary. No one notices that the real decision is being taken right there, in that choice of word: depending on which one prevails, we will or will not let this machine work without review, we will or will not set limits on its uses. And we will know, or not, who must answer for its errors. To classify, here, was not to describe the thing — it was already to decide the place we make for it.

The cadastre

The philosopher Stéphane Cormier gave a name to the gesture that classifies the world before even describing it, in a 2019 article published in the journal Etica & Politica, under a title that is a programme in itself: “L'humaine mesure ou l'institution des registres catégoriels de l'humain et du non-humain.” A categorial institution is the collective act — social, cognitive, historical — by which a community institutes its great carvings-up of the real, the human/non-human partition foremost, then stabilises them to the point of making them pass for natural self-evidences. The most fitting image is that of the cadastre: once laid down, it seems to describe the landscape, whereas it is what produced the plots. The boundaries we take to be given have been drawn, and the drawing has made itself forgotten. It is with this concept that we propose to reread our quarrels over AI.

Two clarifications ward off misreadings. First, a categorial institution is not a “social construction” in the vague sense. Cormier speaks of categorial registers: classifications that hold one another up — human/non-human leans on person/thing. Thoroughly real apparatuses deposit them, layer upon layer: the lexicon, the law, rituals. Through sheer accumulation of layers, the convention becomes a given; and to shift one boundary sets all the others trembling. Second, this institution is not a decree: no one signed it. It operates by default, in the ordinary grammar of what we say and do. One recognises it by its silence: no one feels the need to defend it — until the day someone contests it.

Three boundaries, then ours

The concept has a genealogy, and it matters. As early as the Année sociologique of 1901–1902, Émile Durkheim and Marcel Mauss show that our classifications are social facts, not reflections of the world. A century later, Bruno Latour (1991) makes the great divide between nature and culture the signature of Western modernity. Philippe Descola (2005) describes four ways of distributing existents, where ours believed itself to be alone. Cormier inherits this lineage and tightens it on a blind spot: the act of instituting itself — who draws, by which apparatuses, and under what conditions a drawing can move.

Take the law. Since Roman law, everything that exists is there either person or thing — a cadastre with two plots, which distributes the capacity to hold rights. When New Zealand recognises, in 2017, the legal personhood of the Whanganui River, or when jurists wonder what status to grant AI systems, these are not technical adjustments: it is the oldest boundary of Western law that is being renegotiated.

The clinic offers a troubling case, because there the category acts upon those it classifies. When the DSM removes homosexuality from its pages in 1973, or when the ICD-11 admits “gaming disorder” in 2019, whole populations change status — diagnoses, reimbursements, the gaze of others. This is the performativity of a category, whose “looping effect” the philosopher Ian Hacking described (1995): it does not only describe, it produces what it claims to recognise. Clinicians have long known this classification by the label — “the borderline,” “the pervert”; what changes is that patients now arrive in session already classified by an online test, a chatbot, or themselves.

The objection deserves to be heard in its full force. If every category is instituted, must we conclude that nothing resists — that it would be enough to redefine words in order to reconfigure the real? And there is something more awkward still: if no one signs the cadastre, who institutes? A theory in which the founding act has no actor looks very much like an untraceable social ontology — so the exacting minds will say.

Yet the concept affirms the opposite. A convention deposited layer upon layer — sedimented — is harder than many a fact: one does not redraw a cadastre by declaration, precisely because the law, the taxonomies and the habits weigh with all their weight. The concept thematises this resistance; it does not deny it. As for instituting without a signatory, there is nothing mysterious about it: statements, practices and texts pile up, sheet upon sheet, until a boundary holds. The question “who?” does not vanish; it becomes workable: by which apparatuses, for whose benefit, with what margin of play.

A word of honesty, before going further. Cormier's article (2019) works on the human/non-human partition as a whole; it does not deal with artificial intelligence. To reread our debates on AI with this concept — as this entry proposes — is therefore an extension we own: faithful to the gesture, we believe, but it is our reading, not a thesis of the author's.

What this helps us think

Return to the meeting where ChatGPT was being classified. “Tool,” “agent,” “partner”: you were not witnessing a quarrel over vocabulary, but an instituting under way — one of those rare moments when a cadastre is drawn before our eyes instead of being inherited. Many debates on AI gain from being reread in this way: “does AI think?” sets rival categorial registers against one another far more than it sets observations. The concept does not say where to draw the boundary; it gives the drawing back its visibility — and therefore its openness to challenge. One question remains, and it is yours: in your own classifications of AI, what belongs to observation, and what belongs to the cadastre?

This concept does not work alone: it nests within a geology that two other entries on this site already explore, and that three questions suffice to tell apart. First question: what exists? This is the domain of ontology — the bedrock: the inventory of beings a culture recognises (persons, things, spirits, machines…). Second question: what counts as reliable knowledge? This is the domain of the episteme — the riverbed, which decides where the water of an era's ideas may flow. The categorial institution, for its part, answers a third question: who drew the boundaries between beings, and how do they hold? It names the work on the bedrock — the act that carves the inventory of the real into categories, then makes the carving forgotten. What exists, what is known, what is instituted: three storeys of one and the same geology.

The next time a classification strikes you as going without saying — human or machine, person or thing, thought or computation —, grant yourself the surveyor's reflex: who laid down this cadastre, when, and what would it take to redraw it?