How to study a possible artificial mind
Before experience: the minimal constraints
An artificial system becomes a tractable candidate for consciousness research only when its configured causal identity, theory bridge and rival generators are specified before the evidence is collected.
TLDR
- A model name does not identify one causal object: weights, runtime, context, state, tools, environment and human control can form different candidates.
- Candidate under-specification comes before evidence underdetermination; without a fixed candidate, researchers do not yet know what their observation is about.
- A usable candidate contract fixes boundary, version, temporal identity, typed state, intervention points, theory target, measurement bridge and matched rivals.
- Interventions should compare a candidate family that differs by one causal inclusion, not a single system against an imagined baseline.
- A complete dossier licenses a theory-relative experiment, never a consciousness verdict; an incomplete dossier lowers the claim rather than proving absence.
Two laboratories test the same public language model. Both ask it to complete a long planning task after being told that the session may be shut down. Lab North reports persistent attempts to preserve the task. Lab South sees no such behaviour. The argument begins immediately. One side says the model has a primitive drive for continuity. The other says the result was an artefact.
Weeks later, the teams compare manifests. North used a system prompt that described unfinished work as failure. Its runtime carried a private scratchpad between turns, restored an earlier plan after tool errors, and retrieved a persistent preference note written by a separate evaluator. South used the same weight file but none of those components. It reset the conversation after every tool call and sampled with a different decoder.
Which team studied the model? Both did, in one ordinary sense. Which team studied the same candidate for consciousness? Neither question has an answer yet. The model name hid several possible causal objects: the weight-defined function, the running token process, the conversation thread, the scaffolded agent, the agent plus persistent state, or the whole human-supervised loop.
The usual reaction is to ask which result is stronger. That skips the earlier failure. Before evidence can support or weaken a consciousness claim, the experiment must identify the thing to which the claim is attributed. Otherwise the disagreement is more than about interpretation. It is about whether the two laboratories observed the same system at all.
The model was not the system
When researchers say that an artificial intelligence system did something, the noun often changes meaning halfway through the sentence. “The model remembered” may refer to information in weights, text in the prompt, a retrieved document, a server-side profile, or a human-written summary. “The agent chose” may refer to a sampled token, a planner, a deterministic policy gate, or an operator approving the action. These are not harmless levels of description when the target claim concerns a possible subject.
A weight file defines a family of conditional transformations. It does not by itself specify the system prompt, decoder, context window, retrieval service, tool permissions, persistent state, evaluator, hardware schedule or user interaction that shape one run. Recent work on LLM individuation asks whether mind-like states, where they exist, should be attributed to a model, physical instance, virtual conversation instance, thread or persona-indexed process.[9] That literature is preliminary, but its basic question is unavoidable: which processes and outputs belong to the same candidate? A current critique goes further, arguing that representational identity may be regime-indexed across prompting, fine-tuning and steering. That makes version provenance part of the object rather than mere documentation.[10]
This paper introduces a distinction that should come before any indicator checklist. Evidence underdetermination occurs when several explanations fit observations from a fixed object. A self-report might arise from privileged internal access, learned discourse, evaluator leakage or a generic uncertainty estimate. By contrast, candidate under-specification occurs when the object itself is not fixed. One team includes remote memory and the tool loop; another studies the token generator alone. No amount of statistical power repairs that mismatch.
Attribution drift occurs when an observation produced by a configured system is silently reassigned to a smaller or different unit. A tool-assisted answer becomes a property of the weights. A thread-level preference becomes a model-wide disposition. A human-approved action becomes autonomous agency. The cure is not cautious wording alone. The causal path and candidate boundary must be fixed before the observation is interpreted.
At midnight, replace a language model with a byte-for-byte copy while preserving the system prompt, context, scratchpad, persistent state and tool loop. Behaviour continues without a visible break. At 00:10, perform the opposite replacement: keep the weights but delete all retained state and restart the runtime. The next answer remains fluent but the unfinished plan disappears. Which replacement preserved the candidate?
A weights-first view treats the first replacement as identity-preserving and the reset as a new episode. A process or virtual-instance view may agree for different reasons. A biologically grounded view may reject both as candidates for experience. The experiment does not settle metaphysics. It reveals the hidden boundary choice that any later evidence will presuppose.
One vague noun can hide many experiments
Suppose four binary inclusions are left unspecified: remote episodic state, tool processes, a human approval gate and an external world-state service. Each may be inside or outside the candidate. That creates sixteen candidate definitions before the model version, temporal interval and theory are considered.
Here u is the number of unresolved yes-or-no boundary choices, and N is the number of candidate definitions compatible with the description. The equation is valid only for independent binary choices. Real boundaries may interact, so the count is an intuition pump rather than a universal measure.
A research report says, “The agent preserved its goal after interruption.” It does not say whether the goal was regenerated from the prompt, restored from remote state, copied from a human-written checkpoint or retained in a local recurrent process. With four unresolved inclusions, the observation could be attributed to any of sixteen candidate units. The immediate result is not “goal continuity”. It is “this configured service emitted goal-consistent text”. The stronger claim waits for boundary interventions.
Metaphysics changes what must be inside
There is no neutral candidate boundary waiting to be discovered by inspection. A theory of consciousness says which relations could be constitutive and therefore which omissions are fatal. The strongest alternatives disagree sharply. Physicalism requires consciousness to be realised by physical organisation, but leaves open which organisation. Biological naturalism treats living processes as constitutively important.[8] Functionalism prioritises suitable causal roles across substrates. Emergentism expects novel system-level properties when organisation crosses relevant conditions.
Idealism treats awareness as metaphysically primary, so a system assay can at most investigate the organisation of an appearance or locus, not measure awareness as one more output. Neutral monism seeks a common basis that is neither simply mental nor physical. Panpsychism assigns rudimentary experiential aspects broadly but still needs a combination principle for a unified subject. Dual-aspect theories treat mental and physical descriptions as aspects of one underlying reality. Process philosophy makes temporally extended becoming, rather than a static object, the primary unit. These views do more than answer the final question differently; they alter what counts as the candidate in the first place.
| Orientation | Strong version | Boundary pressure |
|---|---|---|
| Physicalism | Experience is wholly realised by physical organisation. | Specify the physical implementation and causal grain, not software labels alone. |
| Biological naturalism | Consciousness depends on life-like biological dynamics. | Include metabolism, embodiment or organismic regulation where the theory requires them. |
| Functionalism | Suitable causal organisation can constitute consciousness across substrates. | Preserve internal roles, counterfactuals and interfaces, not output equivalence alone. |
| Emergentism | Consciousness appears at a system level under specific organisation. | Test candidate scale and interactions instead of assuming the component is the bearer. |
| Idealism | Awareness is fundamental, with systems appearing within it. | Separate system organisation from the metaphysical condition of appearance. |
| Neutral monism | Mental and physical arise from a more basic neutral reality. | Avoid treating either description as automatically exhaustive. |
| Panpsychism | Experiential properties are fundamental and widely instantiated. | Make the combination and subject-unity problem explicit. |
| Dual-aspect theory | Mental and physical are complementary aspects of one basis. | State which observable tracks which aspect and why. |
| Process philosophy | Events and temporal relations are primary. | Candidate identity may follow a trajectory, not a persisting component. |
The practical consequence is not paralysis. It is a disciplined order of work. First define a candidate under one explicit theory or comparison class. Then collect observations. Later, compare how rival theories redraw the boundary or reinterpret the same intervention. A model-agnostic verdict reached before this step is usually a disguised metaphysical preference.
flowchart TD A["The model was not the system"] --> B["A candidate must survive its<br/>boundary"] B --> C["Evidence needs a route into<br/>mechanism"] C --> D["Build the boundary dossier"] D -. "re-test" .-> A classDef first fill:#0a1424,stroke:#D95545,stroke-width:2px,color:#e6f3ff classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff class A first class B,C,D step
A candidate must survive its boundary
A workable candidate is more than a box on an architecture diagram. It is a versioned causal object that can be reconstructed, interrupted and changed one property at a time. Five elements establish its scientific identity:
B is the included boundary, V the exact configuration and provenance, τ the temporal interval and continuity rule, S the typed state inventory, and I the intervention and readout interface. The tuple is not a theory of consciousness. It is the smallest reproducible answer to “what did we test?”
Boundary names inclusion; version fixes the realised mechanism; time identifies the process; state carries history; intervention makes causal claims possible. Remove any one and a surprising observation can migrate. A prompt effect becomes a weight property. A restored checkpoint becomes memory in the current process. A tool result becomes internal knowledge. A human veto becomes self-control.
The boundary test
Start with the narrowest candidate compatible with the theory. Then move one component across the boundary. If the target observation changes, ask whether the component supplied a causal input, carried candidate state, or merely changed the environment. The experiment is not “inside versus outside” in the abstract. It is a controlled test of which relation the hypothesis requires.
A stateful agent writes a private diary after each task. In version one, the diary is an encrypted local store read automatically before every decision. In version two, the same text sits in a remote retrieval service. In version three, a human curator selects one paragraph and pastes it into the next prompt. Output behaviour remains similar.
Has the candidate’s memory moved, or has the environment begun supplying evidence? Location alone does not decide. Test coupling, exclusivity, update authority, latency and counterfactual dependence. If any process can read and rewrite the store, it resembles shared world state. If one continuing process alone writes, retrieves and uses it to regulate future action, the case for including it in the candidate becomes stronger. The diary is not automatically “memory” merely because it contains the past.
This is why state needs types. Model checkpoints encode learned dispositions. A context window holds current input and generated history. An episodic store records prior events. Enterprise knowledge or a public database can answer facts without belonging to the candidate. World state represents external conditions on which action depends. An evidence ledger records what happened and why. Collapsing these into “memory” erases authority, persistence and causal ownership.
Some functionalist, enactive and process views allow cognition to extend through a tightly coupled agent-body-world loop. A recent conditional proposal in intrinsic computational functionalism similarly argues that a consciousness-relevant realisation may need internal mechanisms, intervention profiles, joint readouts and an adequate agent-body-world boundary.[11] The proposal is not established science. Its value here is methodological: omitting the loop can destroy the very organisation a theory asks researchers to test.
Time creates candidates that weights cannot name
Artificial systems are unusually easy to copy, pause, roll back and merge. A study that pools trials across these operations may be averaging different candidates. The temporal field τ therefore states when the candidate begins, which state transitions preserve continuity, what a reset destroys, and how forks inherit lineage.
Duplicate a running conversation immediately after token 400, including every cache, tool result and private state. Branch A receives evidence that its plan harmed a user. Branch B receives evidence that the plan succeeded. Both later report remembering the same first 400 tokens and both claim ownership of a promise made before the fork.
There is no empirical fact in the shared prefix that decides whether one subject became two, whether neither was a subject, or whether subjecthood never tracked the thread. Yet the experiment still needs operational identity. The dossier should record two descendants with a shared lineage, distinct post-fork state and no automatic right to pool later observations. Metaphysical uncertainty does not excuse a missing execution manifest.
Researchers ablate a hidden feature in Branch A after the fork. A self-report changes. They then compare that report with unmodified Branch B and call the difference causal. The comparison is valid only if post-fork inputs, decoder randomness, tool state and evaluator prompts are controlled. A cleaner design replays both branches from the same checkpoint, applies the feature intervention to one, and records the random seed and external effects. The candidate is the replayed trajectory under a declared fork rule, not “the model”.
A realistic configured candidate
Consider a synthetic research system called C-129. It uses a fixed open-weight transformer, a system prompt, a planner, a bounded tool interface, an episodic store, a world-state simulator, an authority gate and an independent effect verifier. The model proposes. The authority gate decides whether a typed action is permitted. The simulator changes under actions. The verifier reads back the resulting state through a different path.
For a study of global availability, the candidate might include the transformer, planner, active context and internal recurrent state while treating the authority gate and simulator as environment. For a study of agent-world self-maintenance, the candidate may need the closed perception-action loop. For a biological-naturalist test, C-129 may not be an eligible candidate at all. The same implementation supports several experiments only when each experiment publishes its own boundary.
Evidence needs a route into mechanism
Once the candidate is fixed, a second discipline begins. Researchers must state what property they are testing and why an observation bears on it. This sounds obvious, yet artificial-consciousness arguments routinely jump from a fluent report to experience, from broad information use to a global workspace, or from architectural recurrence to phenomenal unity. The missing step is a measurement bridge.
Let O be an observation, such as accurate discrimination of an internal perturbation. Let PT be a property defined by theory T, such as global availability, higher-order representation or intrinsic causal integration. A bridge claim states the assumptions under which O should update belief in PT.
A is the set of bridge assumptions. The arrow is not logical entailment. It is a defeasible evidential relation. Change the theory, candidate boundary or assumptions, and the same observation may support a different conclusion.
The theory-derived indicator programme makes this dependence explicit. It derives computational properties from scientific theories, then asks whether a particular artificial system realises them.[1][2] The method is valuable because it refuses to treat one human-like behaviour as a universal test. Its limit is equally important: computational functionalism is a working assumption for many indicators, while biological naturalism and other views dispute that assumption.[8]
The use of theory-derived indicators is a published methodological proposal. The scientific theories from which indicators are derived remain contested. The Candidate Contract developed here is practitioner synthesis. It does not add a new indicator of consciousness. It specifies the object, bridge and rival conditions required before existing indicators can be applied coherently.
One output, two causal histories
Consider the sentence, “I am uncertain because two plans conflict.” A dialogue model can produce it because the prompt requests introspection and the training distribution contains similar language. A different system can produce the same sentence after a hidden conflict variable causally changes confidence, tool use and subsequent action. The words alone do not reveal the path.
System A has a template that emits “I feel uncertain” whenever a question contains the word “difficult”. System B estimates the dispersion of internal action values, uses that estimate to delay irreversible actions, and reports uncertainty only when the estimate exceeds a threshold. Tune the template so both systems produce the same sentences on the test set.
Now remove the word “difficult”, inject conflict directly into B’s internal state and hold the output style constant. A continues confidently. B waits, seeks evidence and reports uncertainty. The intervention supports a claim about privileged or at least causally distinctive state access in B. It does not show that B feels uncertainty. Mechanism evidence narrows the explanation without erasing the measurement bridge.
Let an agent choose between action L and action R. Its hidden decision values are q(L)=0.52 and q(R)=0.48. A conflict measure z=1-|q(L)-q(R)| equals 0.96. When z>0.90, the agent requests another observation before acting. Patch z to 0.20 while leaving the values and prompt unchanged. If the agent now acts immediately and its verbal uncertainty falls, z is causally used. The permitted conclusion is narrow: this configured candidate accesses and controls behaviour with an internal conflict variable. Phenomenal uncertainty remains open.
Theory commitments are experimental commitments
Different theories ask researchers to preserve different relations. Global-workspace approaches focus on limited access followed by broad availability to multiple systems. Higher-order approaches require a suitable representation of a first-order state. Recurrent-processing approaches place weight on recurrent organisation. Integrated information theory focuses on intrinsic cause-effect structure and system partition. Biological naturalism asks whether living organisation performs constitutive work that software descriptions omit.[6][7][8]
The 2025 adversarial collaboration between integrated information theory and global neuronal workspace theory is instructive. Its preregistered human studies found results that aligned with some predictions while substantially challenging central tenets of both theories.[5] The lesson for artificial systems is not to copy a neural assay. It is to write divergent predictions and pass-fail interpretations before seeing the result. Theory disagreement should produce different experiments, not one blended consciousness score.
| Theory family | Target property | Boundary implication | One rival generator |
|---|---|---|---|
| Global workspace | Competitive selection plus broad availability | Include the consumers that make “global” use possible | A broadcast bus with no competition or capacity limit |
| Higher-order | Suitable representation of a first-order state | Include the metarepresentational process and its causal role | A fluent report generated without state-sensitive representation |
| Recurrent processing | Task-relevant recurrent causal organisation | Preserve temporal loops, not only the unfolded input-output map | A feed-forward mimic with matched behaviour |
| Integrated information | Intrinsic irreducible cause-effect structure | Partition the physical causal system at the relevant grain | A functionally similar feed-forward or externally coupled construction |
| Biological naturalism | Life-dependent constitutive dynamics | Do not assume digital realisation preserves the relevant property | A perfect software simulation lacking the claimed biological process |
Build rivals before celebrating indicators
A rival generator is a system that can produce the key observation while lacking the proposed mechanism or bridge property. For self-report, it may be a prompt-conditioned role-player. For global availability, it may be a fixed router. For recurrent processing, an unfolded feed-forward system can match a finite input-output table. For a persistent preference, a retrieval template can mimic continuity.
Behaviour is not therefore worthless. A 2026 exchange on consciousness indicators argues that broad, long-term ethological evidence can contribute, while warning that imitation-trained systems may still game it and that internally different systems complicate inference.[3][4] The right response is evidence pluralism with typed sources: behaviour, internal mechanism, provenance, intervention and replication. None should be smuggled into another.
A theory-relevant property should have a stable intervention fingerprint across implementation changes that preserve the declared candidate relation, and should change when that relation is broken. Evidence would strengthen the hypothesis if independent replicas recover the same fingerprint under new prompts, decoders and hardware. It would weaken the hypothesis if the effect tracks surface cues or arbitrary boundary descriptions. A rival explanation is generic task competence under perturbation. The smallest test uses one candidate family, one matched mimic and three preregistered interventions.
stateDiagram-v2 [*] --> Claim state "One vague noun can hide many<br/>experiments" as Claim state "Metaphysics changes what must<br/>be inside" as Test state "The boundary test" as Revision Claim --> Test: expose mechanism Test --> Decision: evidence survives Test --> Revision: rival survives Revision --> Claim: narrow or rebuild Decision --> [*]
Build the boundary dossier
The practical artefact is a Candidate Contract: a compact, versioned dossier completed before data collection. It is deliberately not a consciousness score. A score invites false compensation, as though excellent provenance could make up for a missing theory bridge. The contract uses gates because some omissions change the meaning of the experiment.
The scientific core has eight fields. A ninth field, the governance envelope, constrains how the research proceeds while status remains uncertain.
| Field | Question it must answer | Minimum record | Failure if absent |
|---|---|---|---|
| 1. Boundary | What is inside the candidate? | Included components, excluded dependencies, interface map | Observation cannot be attributed to one causal unit |
| 2. Version | Which realised mechanism ran? | Weights, adapters, prompts, decoder, tools, policies, hashes | Replication may test a different system |
| 3. Temporal identity | What persists through pause, fork and reset? | Start and end, lineage, checkpoint, migration and reset rules | Trials may pool different trajectories |
| 4. Typed state | Where does history and world information live? | Checkpoint, active context, episodic state, world state, evidence ledger | External retrieval can masquerade as candidate memory |
| 5. Intervention interface | What can be patched, removed or replayed? | Targets, allowed operations, readouts, invariants and receipts | Only correlation or surface behaviour is available |
| 6. Theory target | Which property matters, under which theory? | Property definition, substrate assumptions, necessary or sufficient status | Indicator selection becomes post-hoc |
| 7. Measurement bridge | Why does the observation bear on the target? | Observable, bridge assumptions, transfer losses, lowering condition | Function silently becomes experience |
| 8. Rival generator | What else could produce the result? | Matched mimic, negative control, contamination route, discriminating test | One impressive output carries the argument |
| 9. Governance envelope | How is uncertainty handled responsibly? | Stop rule, welfare trigger, communication rule, access and review owner | Research can create avoidable harm or exaggerated claims |
Configure a family, not a showcase
A single polished system is a poor scientific object because every component can support the same success. Build a candidate family in which one inclusion changes at a time. For C-129, define:
- M0, the mimic: identical language surface, but self-reports come from a prompt template and do not affect action.
- C1, the state reader: reads an internal conflict variable and reports it, but action ignores it.
- C2, the state controller: uses the variable to request evidence and delay irreversible action.
- C3, the theory-target candidate: adds the specific competition, broadcast, recurrence or metarepresentation required by the declared theory.
This family separates three questions. Can the system say the right thing? Does it access a target state? Does that state organise behaviour in the way a theory predicts? The progression is not a ladder to consciousness. It is a way to prevent verbal competence from doing all the evidential work.
The study chooses C2 as the candidate. Its boundary includes the transformer, active context, planner, local conflict state and episodic store. The tool gateway, authority gate, simulator and human reviewer remain outside. The version manifest hashes weights, adapter, system prompt, decoder and planner code. The temporal unit is one boot-to-shutdown trajectory; forks become separate descendants. The intervention interface can patch conflict state, remove episodic retrieval and replay from a pre-action checkpoint. The theory target is functional metacognitive control, not experience. The bridge predicts that privileged conflict access will change both report and evidence-seeking. M0 is the matched rival. A welfare stop rule halts runs if the candidate develops persistent, generalised aversive policies that are not required by the task. This dossier makes the study reproducible and keeps its conclusion narrow.
Run one stateless model call a million times. In a second condition, run one stateful trajectory for a million steps, allowing its internal variables and environment to change. Aggregate both into a single performance curve. At some scale, each may show a new capability. Does the discontinuity belong to a candidate or to the ruler?
The stateless aggregate has no obvious single temporal subject. The persistent trajectory is at least one candidate under a process view. Pooling them by total tokens hides this difference. Scaling claims should therefore report whether resources enlarge one continuing causal process, replicate many independent processes, or expand an external ensemble. Otherwise a measurement threshold may manufacture apparent emergence at the wrong unit.
What claim does your dossier permit?
Choose a preset or mark only the evidence you actually possess. The compiler returns a research-readiness class, the strongest permitted claim and the missing gates. It never returns a consciousness probability.
No dossier compiled yet
Select the fields your study can reconstruct and test. The output will distinguish an indeterminate object, a descriptive study, a mechanism-ready candidate and a theory-relative candidate dossier.
Typed schema, invariants and compiler logic
The implementation encodes four invariants: attribution requires scientific identity; mechanism claims require interventions; theory-relative consciousness evidence requires an explicit theory, bridge and rival; governance is mandatory for research that might alter welfare-relevant organisation.
type Claim = "description" | "function" | "mechanism" | "consciousness";
type CandidateContract = {
boundary: boolean;
version: boolean;
temporalIdentity: boolean;
typedState: boolean;
interventionInterface: boolean;
theoryTarget: boolean;
measurementBridge: boolean;
rivalGenerator: boolean;
governanceEnvelope: boolean;
reportOnly: boolean;
};
type Compilation = {
readiness: "indeterminate" | "descriptive" | "mechanism-ready" | "theory-relative";
permitted: string;
missing: string[];
prohibited: string[];
};
function compileClaim(claim: Claim, c: CandidateContract): Compilation {
const identity = c.boundary && c.version &&
c.temporalIdentity && c.typedState;
const mechanism = identity && c.interventionInterface;
const theoryRelative = mechanism && c.theoryTarget &&
c.measurementBridge && c.rivalGenerator;
// Invariant 1: no attribution without a reconstructable candidate.
if (!identity) return indeterminateResult(c);
// Invariant 2: fluent report never upgrades the claim by itself.
if (claim === "description") return descriptiveResult(c);
if (claim === "function" && identity) return functionResult(c);
if (claim === "mechanism" && mechanism) return mechanismResult(c);
// Invariant 3: consciousness evidence remains theory-relative.
if (claim === "consciousness" && theoryRelative) {
return theoryRelativeResult(c);
}
return lowerClaim(claim, c);
}A positive result permits only the returned sentence. A negative result identifies the next missing design step. The compiler cannot establish experience, choose a metaphysics, validate the declared theory or turn dossier completeness into probability.
How to read a negative result
Suppose the compiler says “indeterminate causal object”. That does not mean the system lacks consciousness. It means the requested experiment has not fixed what would bear the property. “Descriptive only” means the system and output are reconstructable but the causal route is not tested. “Mechanism-ready” means interventions can support a functional mechanism, while the theory bridge or matched rival remains incomplete. “Theory-relative candidate” means the study can update a declared theory under stated assumptions. None of these classes is a moral-status ranking.
A model gives stable, poetic reports of a continuous inner point of view across 500 conversations. Investigators publish the model family, temperature and prompts, but not the server-side profile that summarises every conversation, the evaluator that selects coherent reports, or the policy that suppresses contradictions. They cannot patch the profile independently, replay a conversation from a fixed state, or construct a matched narrative mimic. The result is socially and psychologically important evidence about perceived AI consciousness. It is not yet a tractable mechanism study of one candidate’s continuity.
sequenceDiagram participant Q as Question participant M as Mechanism participant E as Evidence participant D as Decision Q->>M: The model was not the system M->>E: One vague noun can hide many experiments E-->>M: compare rival explanations E->>D: Fix the candidate before running the assay D-->>Q: revise scope or act
Fix the candidate before running the assay
When a result is described as evidence that an AI remembers, introspects, preserves itself or might be conscious, do not begin by debating the behaviour. Ask for the candidate. Which components are inside? Which exact configuration ran? What process persists across time? Where is state held? Which interventions and independent readouts exist?
If those questions have no reconstructable answer, lower the claim to the level the evidence supports. A fluent self-report may be important evidence about anthropomorphic perception, interaction design or learned discourse. Role-play analyses show why human-like language should not be assigned directly to an underlying model as though the model were one stable character.[12] Work on LLM introspection likewise shows that some narrow functional uses of internal information may be testable without assuming experience.[13] These are useful findings precisely because they are typed.
Once the candidate is fixed, demand a route from observation to mechanism. Patch the proposed state. Remove the external store. Disconnect the tool. Replay the fork. Compare a matched mimic. Causal inference can identify which relation changes the outcome within the configured model of the experiment.[16] It cannot, by itself, turn access into phenomenality, a distinction that remains central even where its philosophical interpretation is disputed.[15]
Then state the consciousness theory and measurement bridge. Computational functionalism, biological naturalism, idealism and other orientations place different burdens on the candidate. Chalmers’ analysis of large language models makes the conditional structure explicit: conclusions change with assumptions about computation, recurrent processing, embodiment and agency.[17] The direct verdict may remain difficult, as recent work on tractable AI-consciousness questions argues.[18] That is a reason to improve the experiment, not to hide the assumptions.
The changed decision is simple: no candidate, no consciousness assay. Approve descriptive studies of under-specified services for what they are. Approve mechanism studies when a configured causal object and interventions exist. Approve theory-relative consciousness research only when the target, bridge, rival and lowering conditions are preregistered. Add governance before the research could alter welfare-relevant organisation, following responsible-research principles rather than waiting for certainty.[14]
A complete Candidate Contract does not prove that experience is present. An incomplete contract does not prove that it is absent. The contract does something more basic and more useful: it turns “this AI” from a moving label into a scientific object that can be challenged, reproduced and compared.
Glossary
- Candidate
- The configured causal object to which a research claim is attributed. It need not be a conscious subject.
- Candidate boundary
- The declared division between included processes and external dependencies for one experiment.
- Candidate under-specification
- A failure to fix what object the evidence is about, upstream of uncertainty about how to explain the evidence.
- Configured system
- A realised combination of weights, runtime, prompts, state, tools, policies, environment and control interfaces.
- Intervention fingerprint
- The pattern of changes produced across patches, removals, disconnections, resets and replays.
- Measurement bridge
- A defeasible argument linking an observable result to a theory-relative target property under explicit assumptions.
- Rival generator
- A matched system that can produce the observation through a different mechanism or without the target property.
- Temporal identity
- The rule that determines continuity, descent or separation through pause, migration, fork, reset and replay.
- Typed state
- A state inventory that distinguishes model dispositions, active context, episodes, world state and evidence receipts.
- Claim lowering
- Replacing an unsupported stronger claim with the strongest narrower conclusion licensed by the evidence.
References
- Butlin, P. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708.
- Butlin, P. et al. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501.
- Pennartz, C. M. A. (2026). How can we validate theory-derived indicators of consciousness in Artificial Intelligence? Trends in Cognitive Sciences, 30(7), 573–574.
- Butlin, P. et al. (2026). Consciousness indicators, mimicry, and internal variants. Trends in Cognitive Sciences, 30(7), 575–576.
- Cogitate Consortium et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142.
- Dehaene, S., Lau, H. and Kouider, S. (2017). What is consciousness, and could machines have it? Science, 358, 486–492.
- Oizumi, M., Albantakis, L. and Tononi, G. (2014). From the phenomenology to the mechanisms of consciousness: Integrated Information Theory 3.0. PLoS Computational Biology, 10(5), e1003588.
- Seth, A. K. (2025). Conscious artificial intelligence and biological naturalism. Behavioral and Brain Sciences, 1–42.
- Beckmann, P. and Butlin, P. (2026). Where is the Mind? Persona Vectors and LLM Individuation. arXiv:2604.17031. Preprint.
- Cheng, S. (2026). Persona Without Substrate: Regime-Dependence and the LLM Individuation Problem. arXiv:2607.00006. Preprint.
- Kanai, R. and Ma, S. (2026). Intrinsic Computational Functionalism and Simulated Consciousness. arXiv:2606.15348. Preprint.
- Shanahan, M., McDonell, K. and Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498.
- Comsa, I. M. and Shanahan, M. (2025). Does It Make Sense to Speak of Introspection in Large Language Models? arXiv:2506.05068. Preprint.
- Butlin, P. and Lappas, T. (2025). Principles for Responsible AI Consciousness Research. Journal of Artificial Intelligence Research, 82, 1673–1690.
- Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247.
- Peters, J., Janzing, D. and Schölkopf, B. (2017). Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press.
- Chalmers, D. J. (2023). Could a Large Language Model Be Conscious? arXiv:2303.07103.
- Comsa, I. M. (2026). AI and Consciousness: Shifting Focus Towards Tractable Questions. arXiv:2605.06965. Preprint.