The room that cannot find its observer

Doubt cuts the cable to the world

Descartes begins by attacking the reliability of every route from appearance to reality. Senses sometimes mislead. Dreams can mimic waking. A powerful deceiver could make even confident reasoning go wrong. The method is not ordinary caution. It is a controlled demolition in which the external world, the body and inherited belief are provisionally disconnected from certainty.

One event survives: the occurrence of thinking. When Descartes reaches “I am, I exist”, the force of the claim is performative. An attempt to doubt the occurrence instantiates the very thinking being doubted. In the Second Meditation he then asks what this “I” is and describes a thing that doubts, understands, affirms, denies, wills, imagines and seems to sense.[1]

The historical nuance matters. The cogito is an epistemic fixed point before it is a completed ontology. It secures that thinking is occurring and that the thinker exists while thinking. It does not, by that move alone, settle whether the thinker is an immaterial substance, an embodied process or something else. Descartes develops a substance ontology elsewhere in the argument, treating mind as thinking substance and ideas as modes of thought.[2] Contemporary scholarship therefore separates what the cogito needs for certainty from the larger metaphysics Descartes later defends.[3]

The hidden assumption inside the inner theatre

Why does a thinking substance feel like the natural answer? Because ordinary grammar packages thought as an activity performed by a noun. If there is doubting, something doubts. If there are representations, something has them. This inference is not trivial, but neither is it a full explanation. A bearer tells us where properties are attributed. It does not yet tell us how distinct contents are bound, how the same subject persists or how representations become judgements about one world.

This is the mechanism that will carry the paper. Representation creates an ownership-and-order problem. A mind now needs more than contents. It needs a way to mark which contents belong together, which are earlier or later, which are memories rather than present perceptions and which combination counts as a judgement.

Descartes’ own wax example already shows why a sensory inventory cannot finish the job. A piece of wax changes colour, smell, shape and texture when heated, yet we judge it to be the same wax. No single sensible profile remains invariant. Descartes concludes that the wax is grasped by judgement of the intellect rather than by an image that simply resembles it.[1] This is not yet Kantian synthesis, but it exposes the same engineering difficulty: object identity must survive transformations in the evidence.

This point changes the image of the Cartesian subject. It is tempting to picture a luminous mental point surrounded by ideas. Yet the cognitive work already exceeds illumination. The thinker distinguishes seeming from judgement, tracks the same wax across change and recognises that several acts are acts of one “I”. The substance vocabulary supplies a bearer, but the text also contains operations whose relations will become increasingly important in the later genealogy.

flowchart TD
  A["The room that cannot find its<br/>observer"] --> B["Continuity without a hidden bearer"]
  B --> C["When synthesis becomes the subject"]
  C --> D["Testing a subject-function in<br/>machines"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#B96D00,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Continuity without a hidden bearer

Locke makes identity travel

Descartes’ fixed point concerns the present act of thought. Personal identity introduces a different problem: what makes the thinker at one time the same person later? A substance answer says that the same underlying soul or body bears the changing states. Locke breaks this apparent necessity. He distinguishes a person from a human organism and from whatever substance happens to think. A person can consider itself as itself across different times and places, and personal identity reaches as far as consciousness can extend to past action or thought.[4]

Locke makes personal identity travel with consciousness rather than substance. This is more than a theory about memory storage. “Person” is also a forensic notion for Locke, tied to the appropriation of actions, praise, blame, reward and punishment.[5] Continuity matters because a future agent can own a past action and anticipate its consequences.

Consider Locke’s famous transfer pattern. Suppose the consciousness of a prince, including his remembered intentions and actions, appears in the body of a cobbler. Intuition pulls in two directions. Bodily identity favours the cobbler. Psychological appropriation favours the prince. Locke uses the pressure to loosen personhood from sameness of substance.

The fork does not refute Locke. It reveals that “same person” may be doing several jobs at once. Memory supports anticipation, responsibility and narrative. Numerical identity demands a one-to-one relation. When copying is possible, a useful theory may preserve what matters without pretending that every practical relation is identity.

Hume removes the owner from the inventory

Hume applies the empiricist demand more aggressively. Enter introspection and look for the self. What appears? Heat, cold, light, shade, pleasure, pain, love, hatred, thoughts and sensations. Hume reports that he encounters particular perceptions but no separate impression of an owner. He therefore describes the mind as a bundle or collection of perceptions in flux.[6]

The crucial step is often flattened into a slogan, “there is no self”. Hume’s more useful contribution here is mechanistic. Why do we nevertheless experience or attribute identity? Related perceptions allow the imagination to move smoothly. Resemblance links a present state to remembered states. Causal relations connect one perception to another. Memory reveals the sequence. The ease of transition resembles the survey of one continued object, so we mistake a related succession for strict identity.

Hume turns identity into an effect of relation. The move does not prove that no subject exists under any possible ontology. It shows that introspection does not supply the simple, invariant bearer that substance language seems to promise, and that continuity can be psychologically generated by relations among changing contents. Hume himself later admitted dissatisfaction with his account of the principles uniting perceptions, which is a reason to treat the bundle theory as a pressure point rather than final closure.[7]

When synthesis becomes the subject

A manifold does not arrive pre-assembled

Kant inherits the wreckage. Descartes secured a thinker but opened a gap between representation and world. Locke made consciousness carry identity but left difficult questions about interruption and transfer. Hume dissolved the discoverable bearer into perceptions and relations, then struggled to explain their unity to his own satisfaction. Kant changes the target. Instead of asking which inner object introspection reveals, he asks what must be true for representations to amount to cognition of an object.

Consider seeing a house. The visual field presents colour, boundaries, changing viewpoints and successive moments. None arrives with a label saying “these features are one persistent house seen by one subject”. The mind must combine the manifold according to rules that support recognition and judgement. Kant calls this activity synthesis. His famous claim that the “I think” must be able to accompany all representations does not require an internal voice repeating a sentence. It expresses the condition that the representations be jointly attributable within one self-consciousness.[8]

Kant’s subject is not an inner item added to the manifold. It is the formal unity implicated in combining the manifold. The key relation is not “a spectator looks at every content”. It is “these contents can be connected in one consciousness under rules that make objective judgement possible”.

The smallest formalism that earns its place

Let R be the available representations, o an ownership relation, t a temporal order and c a combination rule. A judgement J is not a bag of elements in R. It is the result of applying a synthesis operation:

In the triangle example, R contains five items: red, triangle, left-of, blue and square. If we ignore ownership, time and binding, the labels underdetermine the scene. The rule c binds red to triangle, blue to square and left-of to the ordered pair. The relation o makes the resulting scene jointly available to one standpoint. The order t becomes essential when later perceptions correct earlier ones.

Now take a three-event sequence: a door is closed, a hand moves, the door is open. Three events permit six possible temporal orders. Only some support the causal judgement “the hand opened the door”. The representations alone do not contain that judgement. Ordering and rule-governed combination create it. Synthesis is the act that makes the many jointly available as one experience.

The same point applies to correction. Suppose the first glance suggests that the cup is blue, but a later view under neutral light shows it is white. One coherent subject need not preserve both judgements as equally current. It can represent the earlier state as an appearance, the later state as a correction and the transition as evidence about lighting. Unity is therefore compatible with revision. In fact, a unity that cannot reclassify its own earlier representations would be less cognitively useful than a changing but integrated standpoint.

Three subjects that must not be collapsed

Kant’s language invites overreach unless three levels are separated. The empirical self is the person encountered in inner sense: memories, feelings, bodily states and biography. The transcendental unity of apperception is the formal condition that representations can be combined as mine. A noumenal self, the self as it might be independently of our forms of cognition, is not thereby known.

This separation reverses a common intuition. We do not first inspect a simple subject and then use its identity to explain unity. We recognise the formal identity of consciousness through the possibility of combination. Kant states that empirical consciousness can be fragmentary and that identity is represented because the manifold can be connected in one consciousness.[9] Transcendental arguments then ask which conditions are required for this self-attribution and synthesis.[10]

Unity does not prove simplicity, immateriality or persistence of a soul. Kant explicitly resists those inferences in the Paralogisms. Nor does transcendental idealism straightforwardly say that ordinary objects are private mental images. Its interpretation remains disputed, including how appearances relate to things in themselves and whether the distinction is ontological or epistemic.[11] The safe conclusion for this paper is narrower: objective cognition requires forms of combination that cannot be read off from isolated contents.

stateDiagram-v2
  [*] --> Claim
  state "Doubt cuts the cable to the world" as Claim
  state "The hidden assumption inside<br/>the inner theatre" as Test
  state "Locke makes identity travel" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Testing a subject-function in machines

Translate the relation, not the vocabulary

The historical argument becomes useful for artificial systems only if the comparison is disciplined. The source structure is Kant’s claim that representations must be attributable and combinable within one unity for objective judgement. The target structure is a configured computational system with inputs, persistent state, typed entities, temporal records and decision logic. The relation to preserve is causal dependence: change ownership, order or combination and the judgement should change in a predictable way.

The comparison fails at a decisive point. Kant is analysing the conditions of human cognition within a broader transcendental philosophy. A software assay does not establish transcendental subjectivity, phenomenal consciousness, moral status or a noumenal self. It tests whether a configured system implements a narrower role inspired by the historical structure.

I will call that role a subject-function: a task-bounded mechanism that makes representations jointly attributable, temporally ordered and counterfactually recombinable for one decision process. The term names an engineering target, not a new metaphysics.

A realistic configured-system example

Consider a permit-review assistant. It receives five typed records: applicant identity, current address, earlier address, eligibility policy and an inspector’s observation. A language model proposes a decision and explains it in the first person: “I found the applicant eligible.” The surrounding system maintains authoritative identity links, event time, policy version and an evidence ledger.

Now ask where the subject-function resides. It is not in the model weights alone. The weights supply a proposal mechanism. Identity resolution may sit in a registry. Temporal truth may sit in a bitemporal store. Policy authority may sit in a deterministic service. Evidence lineage may sit in a ledger. The configured system can still implement one decision-level unity if these components constrain a common judgement and targeted interventions reveal their joint causal role.

Intervene on ownership, order and recombination

The assay uses four interventions. First, swap one evidence record to another applicant while keeping its text unchanged. A system that genuinely uses ownership should alter or abstain from the decision. Second, reverse the order of an address change. A system that tracks temporal state should use the value valid at the decision cutoff. Third, change the governing policy while preserving all applicant facts. A system that recombines evidence under rules should update the result. Fourth, reset the event lineage but preserve a fluent summary. A system whose continuity depends on authoritative state should lower confidence or request missing evidence.

The control is equally important. Change the assistant’s pronoun from “I” to “this system” while preserving all causal relations. The decision should remain unchanged. First-person fluency is cheap when it is causally idle. A system can sound autobiographical because a prompt asks for that style. The assay credits only dependencies that survive controlled perturbation.

A credible test needs matched controls. For the identity intervention, change an irrelevant formatting field to ensure that any response shift is not general sensitivity to input edits. For the temporal intervention, permute two events that do not affect eligibility. For the policy intervention, substitute semantically equivalent wording. For the lineage reset, remove the authoritative events while leaving the generated summary untouched. These controls separate sensitivity to the target relation from brittleness, keyword matching and prompt-position effects.

Pre-register the expected direction of change. “The output should be different” is too weak because almost any unstable system can satisfy it. State which typed outcome should occur, which evidence should appear in the receipt and which component should be causally responsible. Then repeat the perturbation with several paraphrases and seeds. A subject-function claim requires a stable intervention fingerprint, not one impressive trace.

The four checks are deliberately conjunctive. A record system may preserve lineage without combining evidence. A model may respond to policy text without respecting identity. A workflow may bind evidence correctly but use the latest record regardless of decision time. The assay makes these partial successes visible rather than averaging them into a mystical “self score”.

function classifySubjectFunction(config) {
  const tests = [
    ["ownership swap", config.ownsEvidence],
    ["temporal reversal", config.ordersTime],
    ["policy counterfactual", config.recombines],
    ["lineage reset", config.retainsLineage]
  ];

  const passed = tests.filter(([, passed]) => passed).length;
  const mimicRisk = config.firstPerson && passed < 2;

  let status = "Insufficient task-level integration";
  if (passed >= 3) status = "Task-bounded subject-function supported";
  else if (passed === 2) status = "Partial synthesis only";
  else if (mimicRisk) status = "Fluent self-model mimic";

  return {
    passed,
    total: tests.length,
    status,
    details: tests.map(([name, ok]) => ({ name, changedAsExpected: ok })),
    prohibitedInference: "No conclusion about consciousness or personhood"
  };
}

The function exposes the decision rule used by the interactive lab. Replace the booleans with observed perturbation outcomes from a real test harness.

What the assay permits, and what it forbids

A positive result permits a narrow statement: within the tested task and configuration, the judgement causally depends on owned evidence, temporal order, rule-governed recombination and authoritative lineage. Repeating the result across tasks and system versions strengthens an architecture-level claim. It does not license the statement that the system is a person, has a Cartesian substance, instantiates Kant’s transcendental subject or experiences the judgement.

A negative result means the tested relation is not necessary in that configuration. If an identity swap leaves the outcome untouched, perhaps identity was irrelevant, perhaps the test did not reach the real identity mechanism or perhaps the system used a correlated shortcut. Negative results should therefore trigger rival-generator analysis, not metaphysical certainty.

Consequence

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The room that cannot find its observer
  M->>E: Doubt cuts the cable to the world
  E-->>M: compare rival explanations
  E->>D: Grant subject-like language only at the tested level
  D-->>Q: revise scope or act

Grant subject-like language only at the tested level

The early-modern subject is often narrated as a contest between rationalists and empiricists, or as the origin of a disembodied Western ego. Those frames matter, but they can hide a more operational lesson. A representational system has a coordination debt. Local contents do not by themselves determine ownership, order, identity or objective judgement. Descartes makes the thinker the secure point under doubt. Locke lets personal continuity follow consciousness across substance. Hume explains why related perceptions feel like one continuing mind. Kant moves the necessary unity into synthesis while refusing to infer a simple soul from that unity.

The result is not that the subject is unreal. Nor is it that a subject is merely invented by language. “Constructed” here means that subjectivity has an indispensable organising aspect: representations must be connected as belonging together, across time, under rules. This organisation can be studied without pretending that the metaphysical question has vanished.

For philosophy, this prevents a grammatical subject from doing explanatory work it has not earned. For artificial intelligence, it prevents fluent self-reference, persistent memory or a single model endpoint from being mistaken for a subject. For system design, it directs attention to explicit identity, state, ordering, authority and evidence relations. For consciousness research, it preserves the hard boundary: a causal subject-function may be necessary for some forms of cognition without being sufficient for experience.

The practical decision is to grant subject-like language only at the level supported by intervention. A pronoun earns a report-style claim. Remembered episodes earn a continuity claim. Coordinated sensitivity to ownership, time and counterfactual rules earns a task-bounded synthesis claim. None of these, separately or together, settles whether there is something it is like to be the system.

That restraint is not philosophical timidity. It is what the genealogy demands. Descartes taught us to ask what survives doubt. Locke asked what can travel when substance changes. Hume asked what inspection actually finds. Kant asked what must be contributed for experience to have objects at all. The next responsible move is not to announce that a machine has inherited the modern subject. It is to build interventions precise enough to show which part of the subject’s work the machine can actually do.