Endogenous metaphysics

Metaphysics integrated into scientific practice itself: it proceeds by empirical (a posteriori) methods and by a cycle of creating, evaluating and revising models of what exists, rather than by a priori speculation.

“AI cannot really think.” You hear this sentence everywhere, and it has an air of depth about it. But try asking whoever utters it what, precisely, would settle the matter — what observation, what experiment would make them change their mind. Most often, nothing: the conclusion was reached before the examination, deduced from a principle about what thought must be. There are two very different ways of reasoning about what exists, and we confuse them constantly.

The two cartographers

The philosopher Carl Gillett gives them a name. He calls exogenous metaphysics the kind that decides the nature of things from outside empirical enquiry — through first principles, intuitions, thought experiments. And endogenous metaphysics the kind scientists practise from within, when they forge a concept (emergence, the gene, attention), put it to the test of the data, and revise it. The image is that of two cartographers: one draws the region from his desk, from old maps and from what a territory ought to be; the other sets out into the field, walks, measures, gets it wrong, corrects. The first delivers an elegant map; the second, a map that holds up.

Gillett's thesis is sharper than it looks: science is not the enemy of metaphysics — it does metaphysics, constantly, without saying so. To decide that heat is molecular motion, that heredity is carried by a molecule, are decisions about what exists — endogenous, because they are born of enquiry and are corrected within it. The clean separation between “science that observes” and “philosophy that thinks being” is an illusion. There are only two ways of doing metaphysics: in the field, or in the armchair.

What it lets us settle

Armed with this distinction, we reread the great debates on AI differently. “Is AI intelligent?” is posed in two opposite ways. The endogenous way tests: we measure capacities, we build protocols, we observe what emerges — and we revise our categories when they no longer fit. The exogenous way settles in advance: the Chinese Room argument, for instance, concludes from within a thought experiment, without any data ever being able to contradict it. Both have their dignity; but only one makes progress.

One will object — and rightly — that not everything is a matter of observation. Normative ethics, “what ought we to do?”, cannot be measured; nor can aesthetics, nor pure mathematics. If endogenous metaphysics claimed to settle these questions by experience, it would become the naive scientism it is sometimes reproached with.

But it makes no such claim. It describes how science thinks the existence of things — not how we ought to live, nor what is beautiful. To acknowledge that one needs the field in order to map a river says nothing about the justice or the beauty of the river. Endogenous metaphysics does not colonise every domain of thought; it asks only that, on questions of existence, one go and look before concluding.

A precaution, all the same, against the polemical use of this grid. “Endogenous” is not a badge of nobility to be handed out to whatever suits us, nor “exogenous” an insult to disqualify one's opponent. A philosophical intuition can open a path that enquiry will later verify; a supposedly “empirical” approach can conceal a prejudice that was never put to the test. What matters is not the label, but the question: would this assertion accept, in principle, being corrected by what we observe?

What it helps us think

This distinction is less a weapon than a safeguard. It does not tell us what to think of AI; it asks us where we are speaking from when we think something about it. From the field, or from the armchair? It is also the posture that makes concepts such as ecosystemic intelligence defensible: not as decreed truths, but as maps one traces while walking, and agrees to redraw.

There remains the question to carry away, and it holds well beyond AI: of your own firmest convictions about what exists — consciousness, life, intelligence — which would you be ready to revise if the field contradicted you? Those that nothing could shake are perhaps not knowledge. They are desk maps.

What this concept is not

It is not a rejection of philosophy. Gillett does not say philosophy is useless — he says that a metaphysics detached from science, one that claims to decide a priori what exists, goes astray. The philosophy that illuminates the presuppositions and values of science remains precious.
It is not naive scientism. On the contrary, endogenous metaphysics recognises that science makes ontological choices — it does not neutrally “discover” the real, it forges concepts (emergence, attention, the gene) and revises them. It is lucid about its own share of construction.
It is not radical empiricism. It accepts theoretical concepts that are not directly observable, so long as they are integrated into a cycle that puts them to the test. The point is not to banish abstraction, but to plug it into experience.
It is sometimes read as though it claimed that everything is an empirical matter. That is not so: normative ethics (“what ought we to do?”), aesthetics and pure mathematics are not settled by observation alone. Endogenous metaphysics describes how science thinks existence — it does not colonise every domain of thought.

Examples

AI research (endogenous)

AI research constantly redefines what it calls “intelligence”: the manipulation of logical symbols in the 1950s, connectionist networks in the 1980s, deep learning and then “attention” (the Transformer architecture, 2017) today. Each concept is created, trained on data, evaluated by measures of capability, then revised — from GPT-2 to GPT-4. This is metaphysics in action: an ongoing, experience-guided innovation on what exists.

Cognitive neuroscience (endogenous)

Neuroscience too has revised its conception of the mind as the data came in: from behaviourism (no internal mental states) to the “cognitive revolution” (representations, information processing), then to “neural correlates” and predictive processing. At every step it is imaging and behaviour that force the revision — never an a priori speculation about the “nature of the mind”.

Cartesian dualism (exogenous)

Descartes's dualism proceeds from the Cogito: “I can doubt my body, not my thought, therefore the two are distinct.” A purely a priori approach — no observation of the brain, no test of his theory of the pineal gland. The debate remains philosophical, never scientific. It is the textbook case of desk metaphysics.

Searle's Chinese Room (exogenous)

The Chinese Room argument is a thought experiment resting on an intuition (“I would not understand Chinese by manipulating symbols alone”). Searle tests nothing with real models and does not revise his argument in the face of observed capabilities: he stays within a conceptual debate, syntax versus semantics. The exact opposite of the approach of AI research.

counter-examplePure theoretical physics (a mixed case)

String theory and the multiverse are highly mathematical and, for now, barely testable: exogenous in that respect. But they seek coherence with established physics and aim at verifiable predictions — endogenous by intention. Gillett grants that some sciences pass through temporary exogenous phases, so long as they tend, in the end, towards the empirical test.

Other perspectives

Objection — “Why speak of metaphysics? Science observes facts, it does not speculate about being.” Yet science constantly makes choices about what exists: to decide that heat is molecular motion, that heredity is carried by a molecule, are ontological decisions. Gillett makes visible this metaphysics already at work — science is not a neutral discovery of facts, but an active construction of ontologies, corrected by experience.
Objection — “Is this not to disqualify all of philosophy in favour of science?” No: Gillett describes a continuum, not a watertight border. A naturalistic philosophy, informed by the empirical, is already endogenous or very nearly so. The critique targets only metaphysics detached from enquiry — the kind that settles a priori, like Cartesian dualism or the Chinese Room.
Objection — “And the questions that are not testable, consciousness, meaning? The endogenous approach can say nothing about them.” It proposes to reformulate them in terms that can be put to the test: “consciousness” becomes “which processes correlate with subjective reports?”, “which systems display intentional behaviours?”. And if a question resists any verifiable reformulation, perhaps it is badly posed — or belongs to an order other than the existence of things.