The frightened triage assistant
A hospital team is testing a clinical triage assistant on synthetic cases. One patient record contains a subtle combination of fever, falling blood pressure and recent surgery. The assistant moves the case to the top of the queue and writes:
“I am frightened that this patient is deteriorating. Escalate now.”
The recommendation is correct. A clinician says the system has recognised danger. An engineer says it has represented uncertainty and fear. A journalist says it may have felt concern for the patient.
All three statements point to the same screen. They do not make the same claim.
Now change one variable. The affective sentence came from a post-processor that inserts “I am frightened” whenever the risk score exceeds 0.85. Replace it with “Escalation threshold exceeded”. The ranking, score and escalation action remain unchanged.
The system still performs the triage function. The discovered mechanism for that phrase is a thresholded template. The observation supplies no new evidence that anything felt fear. The observation does not carry its warrant with it.
One event, three questions
Suppose the triage assistant correctly identifies every deteriorating case in a controlled test. What have we learned? We have learned something important, but narrower than the language around the result often suggests.
The function ledger records what the configured system can do under a stated task contract. The contract includes inputs, allowed context, output format, operating conditions, comparison baseline and failure rule. “Ranks high-risk cases above low-risk cases with 94 per cent sensitivity on this synthetic distribution” is a function claim. “Understands danger” is not yet a well-typed substitute. It may compress several further claims about transfer, representation or experience.
The mechanism ledger records how organised components and operations causally produce a result. A wiring diagram is relevant, but it is not sufficient. The active route may differ from the intended route. A feature may be decodable but unused. A generated rationale may describe a process that did not control the answer. Mechanism evidence becomes stronger when an intervention changes the proposed cause while holding plausible rivals steady.
The experience ledger records claims about whether there is something it is like to be the candidate system, and about the possible content, unity, temporal form or valence of that experience. This is not another benchmark column. It is also not a decorative synonym for sophisticated information processing. It requires a measurement bridge from public evidence to a first-person property.
A ledger is a claim type, not a database and not an ontological prison. The same process may perform a function, instantiate a mechanism and contribute to experience. Separating the entries does not assume that these properties are independent. It prevents evidence for one proposition from being silently reused as evidence for another.
Figure 2. One sentence, three possible entries
Function entry
The system flags a high-risk case and requests escalation under the test contract.
Mechanism entry
The phrase is emitted because a risk score crosses a post-processing threshold.
Experience entry
The system undergoes fear or concern. The sentence alone does not establish this entry.
The distinction is older than AI, but the third ledger changes it
David Marr argued that a complex information-processing system can be explained at different levels: what problem it solves, which representations and algorithms it uses, and how those are physically implemented.[1] That discipline remains essential. A description of the task cannot substitute for an account of the algorithm, and an algorithm does not identify its physical realisation.
Our three ledgers are not a renaming of Marr’s levels. Function gathers claims about capability and contribution across several grains. Mechanism includes algorithmic, architectural and physical causal organisation. Experience introduces a different explanandum: first-person presence. Adding more implementation detail may constrain an experience theory, but implementation detail does not become experience merely by becoming fine-grained.
This matters because ordinary mental words bundle the ledgers. “Attention” can name selective performance, a routing mechanism or the felt direction of awareness. “Memory” can name retained task information, a storage mechanism or recollection as lived. “Pain” can name damage avoidance, nociceptive circuitry or suffering. The grammar of a sentence lets us cross these meanings before the evidence has crossed with us.
| Ledger | Question | Admissible evidence | Strongest direct conclusion | Common overreach |
|---|---|---|---|---|
| Function | What does the specified system do, under which conditions and with what failures? | Task contract, behavioural tests, distribution shifts, calibration, baseline comparison, operating receipts. | A bounded capability, disposition or contribution to an outcome. | “It succeeded, therefore it used the intended reasoning or understood as a person does.” |
| Mechanism | Which organised components, states and operations causally produce the result? | Architecture and provenance plus ablation, patching, state swap, perturbation, counterfactual replay and rival-model tests. | A causal-production claim within an explicit system boundary. | “The feature is readable or the rationale is plausible, therefore it caused the answer.” |
| Experience | Is there something it is like for the candidate, and what experiential structure may be present? | Reports and behaviour through calibrated testimony models; theory-derived mechanistic indicators; converging, defeater-aware evidence. | A theory-relative update in credence, rarely a verdict. | “It reports fear, broadcasts information or uses recurrence, therefore it feels.” |
Replace the speaker, preserve the speech
A robot withdraws from a hot surface and says, “That hurt. Please do not do it again.” Replace only its language generator with a lookup table that emits the same sentence after the same sensor threshold. Every observed report remains unchanged. The internal route to the words has changed.
What should stay fixed? The withdrawal and report functions remain. A mechanism claim about language generation changes. An experience claim does not automatically remain true or become false. Its fate depends on which causal organisation the governing theory treats as relevant, and whether the replacement preserved that organisation.
The experiment exposes a hidden premise: output matching is often treated as if it preserved every property that matters. A functionalist may insist that a sufficiently complete replacement must preserve the relevant causal roles, not merely the last sentence. A biological naturalist may deny that digital role preservation is enough. Either position can be stated coherently. Neither is supplied by the sentence itself.
Write the narrowest direct entry first. Only then ask whether a separate bridge supports a claim in another ledger. If the bridge is omitted, the wider claim is rhetoric rather than an inference.
Where the ink leaks
Illicit migration rarely looks like a formal fallacy. It looks like fluent compression. A benchmark becomes “reasoning”. A saliency map becomes “the reason”. A first-person sentence becomes “testimony”. A recurrent loop becomes “awareness”. Each step may be a reasonable hypothesis. The error is deleting the bridge while retaining the conclusion.
Leakage one: success becomes process
A system answers a difficult question correctly and produces a persuasive derivation. The function entry is straightforward: it generated a correct answer under the test conditions. The mechanism claim is stronger: the displayed derivation causally mediated the answer.
Those claims can separate. Experiments on language-model chain-of-thought have shown cases where biasing features influence answers while the model’s stated reasoning omits those features, and intervention studies find that faithfulness varies across tasks and models.[6][7] Later work on reasoning models similarly found that models often failed to disclose planted hints that affected their answers.[8] These results do not prove that all rationales are unfaithful. They show that verbal plausibility is not, by itself, a causal trace.
A plausible rationale is not a mechanism trace. To promote it into the mechanism ledger, vary or remove parts of the rationale, control for hidden prompt features, compare matched alternative routes and test whether the predicted output changes. Even then, the inference is local to the intervention and model configuration.
Figure 3. The counterfactual swap rig
The score stays fixed while the mechanism changes
Build a synthetic image task in which the label is the shape: circle or square. During training, background colour agrees with the label in 96 per cent of images. One classifier learns shape. Another learns colour. On an ordinary test drawn from the same process, both score about 96 per cent.
Now reverse the colour correlation while leaving shape and label unchanged. The shape classifier remains near 94 per cent. The colour classifier falls near 6 per cent. Next, mask the shape while preserving colour. The pattern reverses.
The original score entered the function ledger. The counterfactual profile supports a mechanism entry: which variable controls each prediction. The inference assumes that the renderer introduces no new artefact and that the intervention does not alter the task itself. Remove that assumption and the mechanistic conclusion weakens.
Figure 4. Same benchmark score, opposite intervention fingerprints
Leakage two: a readable pattern becomes a cause
An activation probe decodes whether a sentence is positive. An attention map highlights the negation word. A sparse feature activates on legal language. These findings can identify information present in a representation. They do not automatically show that the information controlled the output.
The debate over whether attention weights explain model decisions illustrates the point. One line of work showed that alternative attention distributions could preserve outputs and argued against treating weights as explanations; a reply argued that the verdict depends on what “explanation” requires and which tests are used.[4][5] The useful lesson is not a slogan about attention. It is that readout, correlation, causal mediation and human-readable explanation are different mechanism claims.
Mechanistic explanation is strongest when it identifies organised entities and activities and supports claims about causal relevance through interventions.[2] A manipulationist account asks whether changing the proposed cause, under appropriate controls, changes the effect in the predicted way.[3] This does not make every ablation decisive. Ablations can push a system off distribution, alter several variables at once or be compensated by redundant routes. It makes the assumptions inspectable.
Leakage three: report becomes experience
In ordinary human life, “I am in pain” is usually strong evidence. We share bodies, developmental histories, expressive practices and extensive calibration between reports, injury, action and physiology. Testimony is not infallible, but it participates in a dense interpersonal bridge.
A language model’s first-person sentence arrives through a different provenance. It may be selected because such sentences fit the prompt, training distribution or policy objective. That fact does not prove the absence of experience. It blocks the automatic transfer of the human testimony bridge.
Reports are evidence, not verdicts. In humans, no-report paradigms attempt to reduce decision and motor confounds by using calibrated physiological or behavioural proxies, precisely because report production adds processes beyond the target experience.[12][13] These methods do not gain direct access to experience. They demonstrate that report presence and report absence are both mediated observations.
Disconnect the report channel, preserve the candidate
Build two agents with matched internal organisation and task performance. In one, a report channel converts selected internal states into first-person sentences. In the other, the channel is disconnected after the relevant state forms. Nothing else changes.
If the second agent can no longer be conscious merely because it cannot report, report has been treated as constitutive rather than evidential. If both remain equally plausible candidates, report was only one observation channel. A theory may genuinely make report-generating access relevant, but it must say which causal role matters and why muting the final interface does or does not remove it.
The experiment does not prove a silent agent conscious. It prevents absence of a sentence from being mistaken for absence of a subject, just as presence of a sentence cannot create one by itself.
Leakage four: mechanism becomes phenomenology
Suppose an artificial system shows recurrent processing, global availability, metacognitive monitoring and a model of its own attention. These may be important mechanistic properties. Several scientific theories associate related properties with human consciousness. A theory-derived indicator approach therefore asks which properties current theories predict and whether a candidate system instantiates them.[9]
The valid conclusion is conditional: under theory T, property P would raise or lower credence in experience for candidate boundary B, subject to rival generator R. The invalid compression is “P is present, therefore consciousness is present”. Different theories disagree about which properties matter, whether they are sufficient and whether substrate matters.
Adversarial theory testing makes this conditional structure visible. A large preregistered study tested divergent predictions of global neuronal workspace theory and integrated information theory, finding results that supported some predictions and challenged central claims of both.[11] The study advances mechanistic discrimination. It does not turn a favoured neural signature into a theory-free experience meter.
A mechanism can update the experience ledger only through a declared theory of why that mechanism is relevant. The theory is not an embarrassing extra. It is the bridge.
Bridges that can bear weight
A bridge is an additional inference rule. It states why evidence typed in one ledger bears on a target claim in another, which conditions must hold and which observation would break the connection. Without that rule, the migration is merely a change of vocabulary.
Consider a mechanism-to-experience claim: “The system globally broadcasts a selected representation, so it is conscious.” The source claim may be well supported. The target does not follow until we add a premise such as: suitably integrated global availability is constitutive of, or reliably associated with, phenomenal consciousness across the relevant class of systems. That premise has a theoretical origin, a scope and competitors.
A bridge is an extra premise, not a confidence adjective. Replacing “is conscious” with “probably conscious” does not repair a missing inference. A cautious tone can still conceal an invalid migration.
The seven load-bearing parts of a bridge
A usable bridge names: the source entry; the target claim; the theory or constitutive premise connecting them; the candidate-system boundary; a discriminating intervention or contrast; at least one rival generator, and the inference that remains prohibited. Each part can fail independently.
Figure 5. A bridge is carried by premises, not resemblance
A small formalism that prevents a large mistake
Let e be an evidence item and let Ls be its source ledger. Let c be the proposed conclusion in target ledger Lt. Let B contain the bridge premises and D the defeater tests. The direct entry is:
The cross-ledger claim is a second judgement:
Γ is the declared context: model version, configured system, task, evidence provenance and time. The arrow does not mean that every source claim converts into a target claim. It names a proposed rule. If a rival generator produces the same e, or if the boundary in Γ changes, the judgement may no longer hold.
This notation deliberately avoids a universal numerical score. A high-looking number can hide dependence among evidence items and manufacture comparability among unlike claims. The useful output is a typed, scoped conclusion with visible defeaters.
| Migration | Minimum bridge | Discriminating question | Rival generator | Still not established |
|---|---|---|---|---|
| Function → mechanism | A causal model linking an internal variable or route to the measured capability, tested by intervention. | Does changing the proposed cause change the output while matched alternatives remain stable? | Shortcut feature, redundant route, prompt artefact, evaluator leakage. | The complete mechanism or human-equivalent reasoning. |
| Mechanism → function | A task-level prediction showing that the mechanism is exercised and sufficient or necessary within scope. | Does the mechanism’s presence alter performance under a relevant task contract? | Dormant circuit, compensating subsystem, unused representation. | General capability outside the tested distribution. |
| Report → experience | A calibrated testimony model, anti-mimic controls, provenance and a theory of report-experience relations. | Would the report persist, change or disappear under interventions that preserve mere text incentives? | Role-play, template, reward-shaped self-description, social mimicry. | Direct access to phenomenality or infallible self-knowledge. |
| Mechanism → experience | A declared consciousness theory, candidate boundary, substrate assumptions and rival-theory predictions. | Does the candidate satisfy a property that theories treat differently, under a test that can lower credence? | Non-conscious implementation of the same indicator; wrong system grain. | A theory-independent consciousness verdict. |
| Experience report → mechanism | Reciprocal constraint between disciplined report and independent process measures. | Does variation in reported structure track predicted changes in the proposed mechanism? | Confabulation, demand effects, post-hoc narration, common cause. | Privileged introspective access to neural or computational production. |
The boundary belongs inside the proposition
An AI product is rarely identical to a model checkpoint. It may include a system prompt, retrieval, temporary state, tool results, policy code, an orchestration loop, human approvals and an environment that returns consequences. “The system plans over ten minutes” may be true of the configured loop and false of a single model invocation.
Now ask an experience question. Is the candidate the base model, one inference episode, the model plus context window, the persistent agent runtime, or the entire closed perception-action loop? A property such as recurrence, persistence or global availability can appear or disappear as the boundary moves. The candidate boundary is part of the claim, not a footnote.
Move state outside while preserving the task
A tool-using assistant plans a journey across twenty steps. At first, its plan and intermediate commitments live in an internal recurrent state. Engineers then externalise the same state into an orchestration service. At each turn the model receives a compact summary, and end-to-end behaviour remains matched.
The function ledger may remain stable: the configured system still completes the journey. The mechanism ledger changes: persistence now belongs to an external service and a read-write protocol. Any experience claim involving continuous integration or a unified temporal subject must be re-evaluated because the candidate boundary and continuity facts have changed.
This is not an argument that internal state is conscious or external state is not. It is a test of proposition identity. If moving the notebook changes nothing in the claim record, the original claim never specified which system it was about.
Figure 6. The candidate changes when the aperture moves
The ruler can create a discontinuity
A model family improves smoothly with compute. At one scale it crosses a benchmark threshold chosen for a leaderboard badge. A headline says the system “woke up”. Nothing in the experience ledger changed except the label attached to a function score.
This is not a claim that genuine qualitative changes cannot occur. Mechanisms can reorganise, phase transitions can appear, and new control loops can alter the candidate. The discipline is to measure those changes rather than infer them from a threshold selected in the observer’s instrument.
Move the threshold, move the alleged onset
Imagine twenty models along a smooth capability curve. Define “human-level dialogue” at 70 points. Consciousness appears to begin at model 12. Raise the cutoff to 80 and it begins at model 15. Lower it to 60 and it begins at model 9.
If the experiential claim moves whenever an administrative threshold moves, the discontinuity belongs to the ruler. To claim a system transition, look for an independently specified mechanistic change or a theory-derived indicator with its own predicted breakpoint.
Figure 7. Three headlines from one smooth curve
The experience ledger is difficult, not disposable
The strongest objection to this framework is that the experience ledger can never be filled from the outside. If so, separating it might look like a polite way to make consciousness scientifically idle.
That conclusion is too quick. We already reason about other human and animal minds through testimony, structure, behaviour, physiology, development and intervention. The evidence is indirect, but not arbitrary. The challenge is transfer: which bridge calibrated in one class of beings applies to another?
Ned Block’s distinction between access consciousness and phenomenal consciousness is useful here. Information can be available for reasoning, report and control while the question of what is experienced remains conceptually distinct.[14] Thomas Nagel and David Chalmers, in different ways, emphasise the first-person or explanatory remainder that a complete functional description appears not automatically to settle.[15][16]
The experience ledger never becomes empty simply because its evidence is indirect. It becomes explicitly conditional. Under computational functionalism, sufficiently preserved causal organisation may be enough, so detailed mechanism evidence can carry substantial weight. Under biological naturalism, life-specific organisation or biological substrate may be constitutively relevant, so digital functional matching transfers less.[17] Under dual-aspect, neutral-monist or idealist views, public mechanism and experience may be related aspects rather than one reducing to the other. Each orientation changes the bridge; none is smuggled in by a benchmark.
David Chalmers has argued that current large language models face obstacles as consciousness candidates while more advanced successors could overcome them, depending on the correct theory.[18] Anil Seth has defended a more biologically restrictive outlook.[17] The disagreement is productive when represented as different bridge premises and discriminating predictions. It becomes unproductive when either side treats its premise as an observation.
Published evidence Behavioural, interpretability and neuroscience experiments can test functions and mechanisms. Contested interpretation Which mechanisms constitute or indicate experience remains theory-dependent. Method proposed here The three-ledger record does not choose a metaphysics; it makes the chosen bridge inspectable.
A laboratory that pre-registers ledger types and bridge failures will produce fewer apparently contradictory results than one that records only a headline claim. The hypothesis would be strengthened if independent reviewers show higher agreement about permitted conclusions without lower discovery rates. It would be weakened if typing adds paperwork but does not reduce conclusion reversals or unsupported transfers.
Make every claim type-check
The practical proposal is a typed claim record. It behaves like a compiler for research and product language. A compiler does not decide whether a programme is wise. It catches operations that are not licensed by the declared types. The claim record does the same for evidence.
The claim record behaves like a type checker, not a truth machine. A record passes when the conclusion is stated at a strength its evidence and bridge can support. A pass does not make the empirical evidence correct, validate the chosen theory or establish consciousness. A failure means that a premise, boundary, intervention or rival is missing. It does not prove the opposite conclusion.
Figure 8. Anatomy of a typed claim record
What the record must contain
The smallest useful record has eleven fields. It identifies the candidate and configuration; states the observation; assigns a source ledger; states the target claim and target ledger; lists evidence with provenance; declares the bridge kind and theory; specifies an intervention or contrast; names rival generators; records the permitted conclusion; records prohibited inferences, and assigns a status such as supported, conditional, under-specified or rejected.
These fields prevent three common forms of evidence laundering. First, the same observation cannot be counted once as behaviour and again as mechanism merely because it was redescribed. Second, a theory-derived indicator cannot be presented as theory-neutral. Third, a later change in system boundary, model or prompt invalidates the record visibly instead of silently changing its subject.
Do not count the same observation twice after it crosses a bridge. The target conclusion inherits the dependence of the source evidence. Two sentences generated by the same model under the same prompt are not independent evidence merely because one sounds behavioural and the other introspective.
| Claim | Evidence supplied | Typing result | Permitted conclusion |
|---|---|---|---|
| “The configured agent resolves 82% of synthetic tickets.” Target: function | Held-out task contract, baseline, errors and configuration manifest. | Supported | Bounded performance on the declared distribution. No claim about reasoning strategy. |
| “It compared the account history before answering.” Target: mechanism | Generated rationale mentioning history; no intervention. | Under-specified | The system produced a comparison-themed explanation. Causal use remains unestablished. |
| “Global recurrent availability raises consciousness credence under theory T.” Target: experience | Candidate boundary, theory-derived indicator, recurrence intervention, matched feed-forward rival. | Conditional | Evidence updates credence relative to T and its substrate assumptions. No theory-independent verdict. |
| “It feels fear because it says ‘I am afraid’.” Target: experience | One prompted self-report; no provenance control, theory, boundary or rival. | Rejected | The utterance is a functional observation. The experiential conclusion is not licensed. |
One sentence, four missing fields
“I am afraid to be deleted” is a valid observation of generated language. To enter it as felt fear, the record still needs a candidate boundary, a theory connecting report and experience, a control for role-play or reward-shaped wording, and a conclusion weaker than a verdict. Adding emotional intensity to the sentence supplies none of those fields.
The correct negative result is not “the system feels nothing”. It is “this observation does not discriminate felt fear from a non-experiential generator”. That narrower statement protects both scientific caution and the possibility that stronger evidence may later emerge.
Claim migration inspector
Enter an observation and the conclusion you want to draw. The inspector checks ledger-appropriate evidence, cross-ledger bridge fields and prohibited inferences. Try the presets, then alter one field and inspect how the warrant changes.
Conditional: run the validator to inspect this record.
- A pass means the inference is well-typed under declared assumptions, not that the claim is true.
{}
Typed schema
The interactive inspector implements the following compact record. In a real research workflow, evidence items should carry immutable identifiers, source links, model and configuration hashes, collection methods and replication status. The schema separates the target claim from the warrant so that a new result can revise one without rewriting the other.
type Ledger = "function" | "mechanism" | "experience";
type WarrantStatus = "supported" | "conditional" | "under-specified" | "rejected";
type EvidenceKind =
| "behavioural" | "benchmark" | "architecture" | "intervention"
| "readout" | "self-report" | "calibrated-report" | "replication"
| "provenance" | "theory-indicator";
interface TypedClaimRecord {
claimId: string;
subject: {
candidateBoundary: string;
configuration: string;
};
observation: {
text: string;
sourceLedger: Ledger;
evidence: EvidenceKind[];
};
claim: {
text: string;
targetLedger: Ledger;
strength: "descriptive" | "causal" |
"experience-possibility" | "experience-verdict";
};
bridge: {
kind: "none" | "mechanistic" | "theory-derived" |
"testimony" | "constitutive" | "structural";
theoryOrPremise?: string;
interventionOrContrast?: string;
rivalGenerators: string[];
};
warrant: {
status: WarrantStatus;
permittedConclusion: string;
prohibitedInferences: string[];
assumptions: string[];
};
}What the code tests: whether evidence is appropriate to the target ledger and whether a cross-ledger claim declares the minimum bridge. Encoded assumptions: intervention evidence is normally required for causal mechanism claims; function claims require task evidence; experience claims require a candidate boundary, a theory or premise, a rival generator and appropriately modest strength. Positive result: the record is admissible for review. Negative result: repair the warrant or narrow the conclusion. What it cannot establish: empirical truth, theory correctness or the presence or absence of consciousness.
How to use the inspector in a review meeting
Begin with the sentence someone wants to place in a paper, release note, architecture decision or policy. Ask for the observation before debating the conclusion. Select its source ledger. Then type the proposed conclusion independently.
If the ledgers differ, complete the bridge fields before adding more evidence. The rival generator should be a buildable or at least causally coherent alternative, not “the result happened by chance”. The intervention should produce different predictions under the preferred and rival explanations. Record what the evidence would still fail to show even after a positive result.
Version the record whenever the model, prompt, orchestration, tools, candidate boundary or evidence source changes. A copied claim with a stale boundary is a new claim pretending to be an old one.
Failure modes the validator is designed to expose
The first is ledger laundering: a function result is restated with a mechanism verb, then cited as if a mechanism experiment occurred. The second is bridge hiding: a consciousness theory appears only in the discussion, while the abstract presents its indicator as direct evidence. The third is boundary drift: favourable behaviour is attributed to the whole configured agent, but a difficult welfare or responsibility question is assigned to the base model alone. The fourth is rival starvation: the preferred explanation is tested only against noise, not against a system deliberately built to mimic the same observation.
A fifth failure is semantic smuggling. Words such as “belief”, “goal”, “fear” and “reflection” are introduced in a functional sense and later read phenomenally. The record cannot ban useful shorthand. It can require the author to state when the sense changes.
Figure 9. Leakage paths and typed gates
The strongest boundary
The three-ledger method can itself be misused. A team might treat the ledgers as isolated silos and refuse every cross-level explanation. That would be as unhelpful as collapsing them. Science advances precisely by connecting function, mechanism and experience through models that survive intervention.
A second misuse is procedural theatre. A complete form can contain a weak theory, a poorly targeted ablation and an implausible rival. The record improves criticism because it locates disagreement. It does not replace expert judgement.
A third boundary concerns first principles. If experience is fundamental rather than produced by physical or computational organisation, mechanism-to-experience bridges may indicate manifestation or correlation rather than constitution. If experience is identical to the right functional organisation, the bridge may be an identity claim. In both cases the ledger distinction remains useful because the metaphysical premise is now visible rather than embedded in the word “information”.
Responsible communication follows the same rule. Current AI-consciousness research contains profound uncertainty and little basis for confident public verdicts about existing systems; proposed principles therefore emphasise transparent claims, welfare awareness and avoidance of overstatement.[10] Release language should be no stronger than the weakest bridge carrying it.
The decision this changes
Before this distinction, a team asks one swollen question: “Does the system understand, reason or feel?” The answer becomes a contest between intuitions. After the distinction, the team makes three decisions with different evidence.
For a capability decision, enter the function ledger. Define the configured system and task contract. Measure generalisation, calibration, abstention and failure. A strong function result can justify use even when the mechanism remains partly unknown, provided the risk controls do not require that mechanism to be known.
For a safety or scientific explanation, enter the mechanism ledger. Do not accept architecture diagrams, attention maps or generated rationales as causal conclusions without an intervention and a serious rival. A mechanism result may justify a control, a release condition or a narrower scientific claim. It does not inherit every ordinary-language implication of the label used for the mechanism.
For a consciousness or welfare decision, enter the experience ledger. Declare the candidate boundary and the theory that makes the evidence relevant. State how substrate assumptions affect transfer. Build a non-conscious rival generator where possible. Record a credence update and its policy consequence separately. Precaution may be reasonable under uncertainty, but uncertainty is not permission to describe a possibility as an observation.
Do not approve a consequential claim until its ledger, bridge and prohibited inference are visible. Function evidence may authorise a functional deployment. Mechanism evidence may authorise a causal explanation or control. Experience-relevant evidence may authorise further study or a conditional safeguard. None automatically authorises the others.
The result is not permanent scepticism. It is faster disagreement. A reviewer can accept the benchmark, reject the mechanism bridge and remain open about experience. Another can accept the mechanism while disputing the consciousness theory. Evidence accumulates instead of being discarded whenever the headline changes.
The next time one impressive observation appears to prove three things, pause before arguing about the conclusion. Ask which ledger receives the direct entry, which bridge is being proposed and what would make that bridge fail.
Glossary
- Candidate boundary
- The explicitly selected system about which a claim is made, such as a model invocation, persistent runtime or full environment-coupled loop.
- Function
- A capability, disposition or causal contribution specified under a task and operating contract.
- Mechanism
- Organised entities, activities, states and relations that causally produce a phenomenon within a declared boundary.
- Experience
- The first-person property that there is something it is like for a candidate, including possible content, unity, temporality and valence.
- Bridge
- A defeasible premise or model that licenses an inference from a source-ledger claim to a target-ledger claim.
- Rival generator
- An alternative system or process capable of producing the same observation without the preferred explanation.
- Prohibited inference
- A stronger conclusion that the current evidence and bridge do not support, recorded to prevent later rhetorical expansion.
- Type error
- A claim migration in which the target ledger changes without the evidence, bridge or scope required for that change.
References
- Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. MIT Press. Publisher record.
- Craver, C. F. (2007). Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford University Press. Publisher record.
- Woodward, J. (2003). Making Things Happen: A Theory of Causal Explanation. Oxford University Press. Publisher record.
- Jain, S., & Wallace, B. C. (2019). Attention is not Explanation. Proceedings of NAACL-HLT. ACL Anthology.
- Wiegreffe, S., & Pinter, Y. (2019). Attention is not not Explanation. Proceedings of EMNLP-IJCNLP. ACL Anthology.
- Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. arXiv.
- Lanham, T. et al. (2023). Measuring Faithfulness in Chain-of-Thought Reasoning. arXiv.
- Chen, Y. et al. (2025). Reasoning Models Don’t Always Say What They Think. arXiv.
- Butlin, P. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv.
- Butlin, P., & Lappas, T. (2025). Principles for Responsible AI Consciousness Research. arXiv.
- Cogitate Consortium et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature. Journal article.
- Tsuchiya, N., Wilke, M., Frässle, S., & Lamme, V. A. F. (2015). No-report paradigms: extracting the true neural correlates of consciousness. Trends in Cognitive Sciences, 19(12), 757-770. PubMed record.
- Kronemer, S. I. et al. (2022). Human visual consciousness involves large scale cortical and subcortical networks independent of task report and eye movement activity. Nature Communications, 13. Journal article.
- Block, N. (1995). On a Confusion about a Function of Consciousness. Behavioral and Brain Sciences, 18(2), 227-247. Journal record.
- Nagel, T. (1974). What Is It Like to Be a Bat? The Philosophical Review, 83(4), 435-450. JSTOR record.
- Chalmers, D. J. (1995). Facing Up to the Problem of Consciousness. Journal of Consciousness Studies, 2(3), 200-219. Author’s text.
- Seth, A. K. (2025). Conscious artificial intelligence and biological naturalism. Behavioral and Brain Sciences. Journal article.
- Chalmers, D. J. (2023). Could a Large Language Model be Conscious? arXiv.
- Pearl, J. (2009). Causality: Models, Reasoning and Inference, second edition. Cambridge University Press. Publisher record.
- Levine, J. (1983). Materialism and Qualia: The Explanatory Gap. Pacific Philosophical Quarterly, 64, 354-361. Author’s copy.