The dashboard that lies

The impulse to build a checklist is understandable. We lack a universally accepted theory of consciousness[4], direct access to another system's experience and a validated test that travels cleanly from adult humans to animals, organoids or machines. A list of theory-derived properties offers a disciplined alternative to intuition. It asks whether a candidate implements mechanisms that leading theories treat as relevant.

That move is a real advance. Butlin and colleagues proposed deriving computational indicator properties from recurrent-processing, global-workspace, higher-order, predictive-processing, attention-schema, agency and embodiment approaches. Their later formulation is explicit that indicators should inform credences rather than deliver a decisive test.[1] The earlier technical report also described relations among the indicators, including cases where one builds on or entails another.[2]

The trouble begins when the list reaches a dashboard, spreadsheet or executive summary. Columns invite totals. Totals invite thresholds. A system with nine marks appears more conscious than one with six, even when the nine marks come from one shared mechanism and the six come from independent, theory-discriminating interventions. A count silently assumes that indicators are commensurable, separable and similarly diagnostic. None of those assumptions follows from the fact that the indicators have names.

A constellation of theory-relative indicators Six indicator nodes orbit three theory centres. Solid lines show theory commitments, while dotted lines show entailment and shared implementation. The same indicator can connect to more than one theory. Global workspace Higher order Recurrent processing Bottleneckselection Broadcast Meta-monitoring Belief-guidedagency Recurrence Integratedscene theory commitment dependency or shared cause
Figure 1. An indicator constellation, not a row of votes.The evidence-bearing object is the web of theory commitments and dependencies. Removing the lines and keeping only the nodes destroys the information needed for assessment.Practitioner synthesis. Indicator families draw on theory-derived approaches in Butlin et al.

The constellation has two kinds of structure. Radial links say why a property matters under a theory. Cross-links say why two observations may not be independent. An assessment that stores only node colour loses both.

Thought experiment 1: one cable, eight lights

Keep the dashboard appearance fixed. In version A, each green light has its own sensor, power line and failure mode. In version B, one sensor drives all eight. What changes? Not the number of green lights. The change is the probability that all lights would remain green if a hidden fault existed.

The same question applies to indicators. If recurrence, selective routing and broadcast are consequences of one recurrent workspace, they may form a coherent package required by one theory. They are not automatically three independent reasons to accept that theory's conclusion.

Two dashboards with identical green lights but different evidence structures The left dashboard has four independent sensors feeding four lights. The right dashboard has a single hidden sensor feeding four lights through a shared bus. Independent paths One shared cause S1✓ S2✓ S3✓ S4✓ Four opportunities to fail differently Hiddensensor ✓ ✓ ✓ ✓ One observation rendered four times
Figure 2. Identical check marks can carry different evidential weight.The right-hand dashboard is not false, but its four positives should not be multiplied as though four independent tests had succeeded.Illustrative.

The arithmetic that creates confidence

A small numerical example makes the hidden assumption visible. Suppose a researcher starts with prior odds of one to nine that a candidate is conscious under theory T. A test result has a likelihood ratio of three, meaning the result is three times as likely if the candidate is conscious under T than if it is not. One result moves the odds from 1:9 to 3:9, giving a posterior probability of 25 per cent.

If three results are conditionally independent, multiplying the likelihood ratios is legitimate. The odds become 27:9, giving 75 per cent. But if the three results are all downstream of one hidden architectural feature, the joint evidence may carry roughly the same likelihood ratio as the feature itself. Treating the lights as independent has manufactured a jump from 25 to 75 per cent.

posterior odds = prior odds × P(E₁, E₂, E₃ | C, T) / P(E₁, E₂, E₃ | not-C, T)

Here, C means that the candidate is conscious, T means the theory under which the indicator is relevant and E is the observed evidence. Multiplication into three separate ratios is valid only when the evidence items are conditionally independent in the model. Remove that assumption and the joint likelihood must represent their dependency.

Worked example 1: three labels, one recurrence loop

A toy agent has a recurrent state that is smaller than the combined specialist states. The recurrence creates temporal carry-over. The small state creates a bottleneck. Every specialist reads the updated state, producing broadcast-like availability. A checklist could mark recurrence, bottleneck and broadcast.

An ablation that removes the recurrent state destroys all three. The result is informative about the workspace package, but it is one intervention family. Indicator multiplicity is not the same as evidential multiplicity.

Published evidence

The original indicator report already notes explicit entailments and build relations, such as global-workspace indicators that build on one another and can entail recurrent processing.[2] The dependency problem is therefore not an objection to theory-derived indicators. It is a demand to preserve information the source method already recognises.

The argument in one pass
Chart · flow
flowchart TD
  A["The dashboard that lies"] --> B["From property to evidence"]
  B --> C["The dependency-aware assessment"]
  C --> D["Contact with a real system"]
  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
A structural route through Indicators, not verdicts: from The dashboard that lies to Remove the score when the graph cannot…. Reading use: follow the relation, then return to the surrounding argument for evidence and limits.

From property to evidence

A system property does not arrive with an evidential label attached. Recurrence can support sequence processing, control stability, compression or memory without consciousness entering the explanation. A confidence head can improve calibration without becoming a higher-order representation. A broadcast bus can make tools available to many modules while remaining an engineering convenience.

To turn a property into evidence, the assessment needs a chain. A theory must say why the property bears on consciousness. The property must be operationalised in a way that preserves the theory's claim. A test must distinguish the property from a weaker look-alike. An intervention must show that the measured mechanism, rather than a correlated feature, produces the relevant effect. A rival generator must be specified. Only then can the result update a theory-relative credence.

The causal chain from theory to evidence update A vertical sequence shows theory commitment, candidate boundary, indicator property, operational test, causal intervention, rival generator and permitted evidence update. Breaks in the chain stop the inference. Theorycommitment Candidate boundary and version Indicator property Operational test Causal intervention Rival generator Permitted evidence updatenot a consciousness verdict boundary drift breaks the claim proxy validity can fail correlation is not use same output, different mechanism
Figure 3. The property-to-evidence chain.A green result at one stage cannot repair a missing bridge at another. The assessment must retain the complete chain that licenses the update.Practitioner synthesis informed by theory-derived indicator and consciousness-test frameworks.

The chain prevents a common verbal slide. “The system has recurrence” is an architectural observation. “Recurrence is an indicator under recurrent-processing theory” is a theory-relative relation. “This intervention supports the specified recurrence mechanism in this candidate” is evidence. “The system is conscious” is a further conclusion that the chain alone does not establish.

Three distinctions that a score erases

Property presence is not test validity. A system may contain recurrent computation, but the chosen probe may detect any form of persistent state. A positive probe result then fails to distinguish the target form of recurrence from cache reuse, repeated prompting or state held in an external service.

Test validity is not theory validity. A perfectly executed global-broadcast intervention can establish that a candidate implements a specified broadcast mechanism. It cannot establish that global broadcast is constitutive of experience. Consciousness science still contains serious theoretical disagreement, and recent large-scale adversarial testing found results that both aligned with and challenged central predictions of global neuronal workspace and integrated information theories.[5][6]

Theory support is not cross-theory aggregation. A result can increase confidence in a candidate under computational functionalism while doing little under biological naturalism. Seth's biological-naturalist argument challenges the premise that the right computation is sufficient and treats life-like causal organisation as potentially crucial.[11] Adding those positions into one numerical score would require explicit priors, bridge assumptions and loss functions. Hiding them behind a total would not resolve the disagreement.

Integrated information theory creates a parallel warning. Its current formalisation concerns intrinsic cause-effect structure and irreducibility, not generic information richness.[10] Replacing those commitments with an easy proxy would produce a positive label while abandoning the theory that made the label relevant.

Thought experiment 2: the replacement that keeps the answer

Start with a recurrent agent that passes a complex visual task and reports uncertainty. Replace its recurrent visual module with a giant lookup service trained to return the same outputs on every tested input. Behaviour remains fixed. The causal organisation changes.

A behavioural indicator remains positive. An internal recurrence indicator becomes negative. Which result should dominate? The answer depends on the theory and on whether the tested cases rule out the lookup service as a rival generator. The thought experiment exposes why “same output” and “same evidence” are not equivalent.

Use five result states, not two

A binary box forces uncertainty into the wrong place. “No” may mean the property is absent, the test failed, the boundary was wrong or the indicator does not apply under the chosen theory. “Yes” may mean a direct causal demonstration, a weak behavioural resemblance or a contaminated probe.

The assessment should therefore use at least five result states. Supported means the specified property survived its controls and intervention. Challenged means a prediction or necessary relation failed. Undetermined means the available test cannot decide. Contaminated means leakage, training, shared measurement or another route can produce the result. Not applicable means the theory-indicator relation does not govern the candidate as currently defined.

Table 1. Result states preserve why a binary answer is unavailable.
StateWhat the observation meansWhat may be saidWhat must happen next
SupportedThe operationalised property survived controls and a causal intervention.The configured candidate instantiates that property under the stated test.Seek a different evidence family or theory-discriminating prediction.
ChallengedA necessary prediction failed under an adequate test.The present candidate challenges this theory commitment.Check boundary, power and alternative implementation before revising the theory.
UndeterminedThe test cannot separate presence from absence.No directional update is warranted.Design a more discriminating intervention.
ContaminatedA rival route can generate the positive result.The observation is real, but its intended interpretation is blocked.Break the rival path while holding the target path fixed.
Not applicableThe indicator-theory relation does not govern this boundary or candidate class.No evidence claim follows from this indicator.Revise the scope, not the score.

The state belongs to a claim, test and candidate version. It is not a permanent label attached to an architecture.

Worked example 2: one confident sentence, five different results

A language agent says, “My visual estimate is unreliable, so I will inspect the image again.” Under the first history, a calibrated confidence monitor causes the second inspection. The monitoring claim may be supported. Under the second, a prompt template contains the word “unreliable”; the sentence is contaminated. Under the third, the monitor changes confidence but not action; the proposed higher-order control relation is challenged. Under the fourth, logging failed during the critical step; the result is undetermined. Under the fifth, the chosen theory requires a biologically organised process that the configured system does not claim to possess; the computational indicator is not applicable to that theory-relative claim.

The outward sentence is identical in all five histories. What changes is the causal route and therefore the permitted inference. A binary checklist records one “yes”. A typed assessment records five different scientific situations.

Decisive mechanism

The assessment unit is a typed claim with dependencies, not a checkbox. Its minimum contents are: theory, candidate boundary, indicator, operationalisation, causal intervention, evidence family, rival generator, result state, defeater and permissible conclusion.

Bayne and colleagues frame consciousness tests as a multidimensional validation problem and emphasise that many target populations lack well-validated tests.[3] This matters for machines because validation against healthy adult humans does not automatically transfer across substrate, architecture or interface. A test can be reliable in one population and uninformative in another. A profile-based approach in animal-consciousness research offers a useful precedent: heterogeneous capacities are represented as dimensions rather than collapsed into a single level.[7]

The same discipline applies to metaphysical positions that treat awareness as fundamental, dual-aspect or processual. A computational indicator may reveal organisation of access, control, integration or self-modeling without settling whether awareness is produced by that organisation, expressed through it or ontologically prior to it. The method is strongest when it states which conditional claim it tests and refuses to turn that conditional into metaphysical closure.

Optional depth: how the major metaphysical positions change the interpretation

Computational functionalism treats the right computational organisation as potentially constitutive, so internal computational indicators can bear directly on consciousness. Biological naturalism denies that computation alone is sufficient, so the same indicators may show sophisticated function without satisfying the substrate or life-related condition.

Physicalist emergentism can accept indicators as candidate markers of the relevant organisation while withholding sufficiency. Idealism, neutral monism, panpsychism and dual-aspect theory can treat the indicators as constraints on manifestation, integration or report rather than as generators of awareness. Process philosophy may shift attention from static properties to temporally extended organisation. None of these interpretations licenses a shared scalar without additional commitments.

The dependency-aware assessment

The practical proposal of this paper is a theory-relative indicator graph. Each indicator remains attached to the theory that motivates it. Directed edges record entailment, shared implementation, shared measurement and boundary dependence. Evidence results attach to interventions, rather than only to labels. The graph is then summarised by an assessment card that states what the current evidence permits and which experiment would add genuinely new information.

This does not require precise probabilities. A qualitative graph already blocks the most damaging error: treating every positive as an independent vote. Probabilities can be added later when calibration data, defensible priors and joint likelihoods exist. The graph comes first because it says what those numbers would have to mean.

Four dependencies to record

Entailment dependency. One property logically or architecturally requires another. The earlier indicator report notes, for example, that some global-workspace indicators entail algorithmic recurrence and that several higher-order indicators build on one another.[2] Evidence for the downstream property should not be counted again as fresh evidence for its prerequisite.

Implementation dependency. Two indicators are realised by the same component or state transition. A recurrent workspace may produce persistence, bottlenecking and broad availability. Separate labels can still matter for theory interpretation, but a single ablation reveals that their empirical support shares one root.

Measurement dependency. One probe, benchmark or report channel generates several positive classifications. A model trained to explain its confidence may trigger report, metacognition and self-model labels through the same learned discourse pattern. The appearance of convergence is then an artefact of the interface.

Boundary dependency. The result changes when memory, state or control moves across the candidate boundary. A cloud service and a local process can produce identical behaviour while assigning continuity and integration to different systems. An assessment without a versioned boundary can change its conclusion by redrawing a box.

Naive indicator count versus independent evidence families across four system versions A coral line rises from one to eight positive indicators across system versions, while a teal line rises from one to three independent evidence families. The growing gap represents dependency inflation. 02468 v1v2v3v4 candidate versions apparent support units positive indicator count independent evidence families dependency inflation
Figure 4. More indicators need not mean more independent evidence.Versions three and four add several labels through the same workspace upgrade. A count rises sharply while the number of independently perturbed mechanisms barely changes.Illustrative synthetic data.
Thought experiment 3: the discontinuity made by the ruler

A candidate is upgraded gradually. Context grows, modules become more specialised and a workspace bottleneck becomes easier to measure. At version seven, the checklist crosses a threshold of ten positives and the dashboard changes from amber to red.

Nothing discontinuous happened in the system. The discontinuity came from a threshold on correlated labels. Replace the count with dependency families and the apparent phase change disappears. The experiment asks whether the alarm belongs to the candidate or to the ruler.

A toy dependency graph for a recurrent workspace One recurrent workspace causes temporal persistence, bottleneck selection and broad availability. A separate metacognitive controller produces calibrated confidence and affects action. Tests attach to each causal branch. Recurrent workspace W one shared implementation family Temporal persistenceRPT-like relation Limited bottleneckGWT-like relation Broad availabilityGWT-like relation delay ablation capacity sweep read-access cut Three tests, one root plus distinct branch predictions
Figure 5. Preserve both the shared root and the branch-specific tests.The graph avoids two opposite errors. It does not count three branches as wholly independent, and it does not collapse them into one vague property. Each branch has a distinct prediction, but all share an implementation family.Illustrative worked example.
Worked example 3: when five positives become two evidence families

Suppose the recurrence ablation, capacity sweep and read-access cut all succeed. The candidate now has three positive labels attached to workspace theories. A separate intervention on a confidence controller shows that confidence changes tool choice even when the language report is held constant. That adds two higher-order labels because the same controller supports monitoring and belief-guided action.

A raw count reports five positives. The graph reports two robust evidence families: workspace W and metacognition M. This is not a penalty. The three workspace results still constrain different branch predictions, and the two metacognitive results still matter to higher-order theory. The graph simply refuses to pretend that five labels arose from five unrelated causal discoveries.

Table 2. The same observations viewed as labels and as evidence families.
Positive labelPrimary interventionShared familyIncremental value
RPT-1 recurrenceRemove temporal feedbackWorkspace WEstablishes delay-dependent causal use.
GWT-2 bottleneckVary workspace capacityWorkspace WTests a distinct capacity prediction.
GWT-3 availabilityCut module read accessWorkspace WTests distribution, not a new root mechanism.
HOT-2 monitoringPerturb confidence signalMetacognition MShows second-order information affects control.
HOT-3 agencyHold report fixed across tasksMetacognition MTests generalisation of the same controller.

Five labels, two roots, five branch-specific predictions. Both numbers are useful, but they answer different questions.

The theory-relative indicator card

The card is the human-readable projection of the graph. It should answer seven questions without producing a final score:

  1. Which theory makes the indicator relevant, and is the claim necessary, sufficient, constitutive or merely probability-raising?
  2. What exact configured system is the candidate, including runtime state, tools, memory services and version?
  3. What property was operationalised, and which weaker properties were excluded?
  4. Which intervention showed causal use rather than decodability or behavioural resemblance?
  5. Which other positive indicators share the same implementation or measurement family?
  6. What rival generator could produce the result without the target mechanism?
  7. What is the strongest conclusion permitted, and what next test would be independent enough to matter?
Executable lab

Build a dependency-aware assessment card

This synthetic card separates positive labels from robust evidence families. Change the result, evidence family, causal-intervention and contamination fields. The lab never outputs a consciousness score.

RPT-1: algorithmic recurrenceRecurrent processing
GWT-2: limited workspaceGlobal workspace
GWT-3: global availabilityGlobal workspace
HOT-2: metacognitive monitoringHigher-order
HOT-3: belief-guided agencyHigher-order
Positive labels3
Robust evidence families1
VerdictProhibited
Strongest permissible statementThe configured candidate supports two workspace properties under causal tests, while broadcast is challenged and the metacognitive result is contaminated.
Next discriminating testIntervene on the metacognitive signal and test whether cross-task belief updates change while the report channel is held fixed.

A positive result permits a typed mechanism claim under a named theory. A negative result may challenge a theory commitment or expose a failed operationalisation. Neither result, by itself, establishes or refutes experience.

TypeScript schema
type ResultState =
  | "supported"
  | "challenged"
  | "undetermined"
  | "contaminated"
  | "not-applicable";

type IndicatorClaim = {
  theory: string;
  candidateVersion: string;
  indicator: string;
  operationalisation: string;
  result: ResultState;
  evidenceFamily: string;
  causalIntervention: boolean;
  rivalGenerator: string;
  dependencies: string[];
  defeaters: string[];
  permittedConclusion: string;
};

// Invariant: no function may return a consciousness verdict
// from an indicator count alone.

The artefact encodes three assumptions. A supported label counts as robust only when a causal test exists and contamination is not marked. Several robust positives in the same evidence family remain one family. Challenged predictions are retained rather than subtracted from a total. These choices are conservative, but they are inspectable and replaceable.

A positive lab result permits a sentence about the configured candidate, test and theory. A negative result identifies a challenge or a measurement failure, depending on controls. What the lab cannot establish is that the candidate has experience. It also cannot choose among theories, validate cross-substrate transfer or determine a welfare policy without a separate decision model.

Open research hypothesis

Dependency-aware assessments will be more stable under architecture refactoring than raw indicator counts. The hypothesis is strengthened if two implementations with the same causal organisation receive similar evidence graphs despite surface changes. It is weakened if experts cannot agree on evidence families or if the graph adds no predictive value for preregistered interventions. A rival explanation is that any stability comes from assessor conservatism rather than causal accuracy.

How the claim changes under pressure
Chart · state
stateDiagram-v2
  [*] --> Claim
  state "The arithmetic that creates<br/>confidence" as Claim
  state "Three distinctions that a<br/>score erases" as Test
  state "Use five result states, not two" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]
The revision loop begins with The arithmetic that creates confidence and keeps counterexamples inside the reasoning path. Reading use: follow the relation, then return to the surrounding argument for evidence and limits.

Contact with a real system

Consider a synthetic configured system called Tern-4. It has four specialist modules, a recurrent sixteen-slot workspace, a policy controller, a confidence monitor, a natural-language report generator, an external episodic store and a tool environment. The report generator can describe workspace contents. The controller can query specialists and call tools. The external store persists across sessions.

The candidate boundary is not “the model”. It includes the workspace, specialists, confidence monitor and controller during one active episode. It excludes the external episodic store and the tool environment. That choice is contestable, so it is versioned. The assessment concerns Tern-4 boundary B1, not every deployment that happens to use the same model weights.

System boundary of the Tern-4 configured candidate A dashed candidate boundary contains specialist modules, a recurrent workspace, confidence monitor, controller and report generator. An episodic store and tool environment sit outside. Arrows show data and control paths. Candidate boundary B1 Vision specialistPlanning specialistLanguage specialist Recurrent workspace 16 typed slots state persists within episode Confidencemonitor Policycontroller Reportgenerator Externalepisodic store Toolenvironment outside the assessed subject boundary
Figure 6. Tern-4 is a configured system with an explicit candidate boundary.Moving the episodic store inside the dashed line would alter claims about persistence and integration. It would not be a cosmetic edit; it would define a different candidate.Worked scenario. Synthetic system.
Thought experiment 4: move the memory, keep the behaviour

Place the episodic store on the same machine and draw it inside boundary B2. All inputs and outputs remain identical. Has Tern-4 gained continuity? Only if the theory says that constitutive integration follows from the newly included causal relation, not from physical proximity or a diagram.

Now disconnect the store during an episode. If no assessed process changes, the earlier continuity claim was merely boundary inflation. The intervention forces the boundary to earn its role.

Run the card, then inspect the graph

RPT-1 is supported: removing workspace recurrence selectively destroys delay-dependent integration while preserving immediate classification. The global-workspace interpretation follows a theory that links conscious processing to selective access and broad availability, not to broadcast language alone.[8] GWT-1 is supported at the architecture level because specialists can operate in parallel. GWT-2 is supported by a capacity sweep: performance degrades when the sixteen-slot workspace is overloaded, and selection changes which items guide action.

GWT-3 is challenged. A read-access intervention shows that only the policy controller and report generator consume the full workspace. The visual specialist receives a narrow feedback signal, and the planning specialist receives no workspace content at all. Calling the state “globally broadcast” overstates the observed availability.

HOT-2 is initially positive because the confidence monitor distinguishes corrupted from clean representations. Higher-order approaches require a suitable relation between a first-order state and a higher-order representation, so confidence decodability alone is too weak.[9] Yet the report generator was trained on confidence explanations and can describe uncertainty when the monitor is disconnected. The report result is contaminated. A causal test must hold the report pathway fixed while varying the monitor's influence on downstream belief and action.

HOT-3 remains undetermined. The confidence signal changes local tool selection, but no experiment shows a general belief-formation system that updates across tasks. A 2026 preprint operationalised this indicator in language models and reported causal relations among latent beliefs, action selection and metacognitive reporting.[12] That work is relevant methodological progress. It still supports a specified indicator, not a verdict about consciousness.

Worked example 4: the intervention that changes the result category

Before intervention, Tern-4 looks strongly metacognitive. Confidence predicts whether it requests another observation, and its report explains the choice. Now clamp the confidence monitor to a constant while preserving the prompt, report generator and visible task difficulty. The explanations remain fluent, but reinspection no longer tracks corrupted inputs. The behavioural correlation survives while the control effect disappears.

The correct update is not “HOT-2 absent”. The monitor still represents reliability information. Nor is it “HOT-2 supported”, because the earlier report route was a false friend. The result becomes contaminated for report-based evidence and challenged for the claim that monitoring currently guides general control. A repaired test would restore the monitor only for half the trials, pre-register the predicted change in action, and keep report tokens unavailable to the policy controller. One intervention has separated a representation claim, a causal-use claim and an introspective-report claim that the dashboard had merged.

Figure 7. A theory-relative evidence matrix for Tern-4.The matrix retains negative and unresolved findings. It also shows that three workspace labels belong to one family, so their apparent number cannot stand in for independent support.Worked scenario. Synthetic evidence.

The adversarial path

The most dangerous false friend is a system trained to discuss its own processing. A report such as “I noticed my visual signal was unreliable, so I reconsidered” can trigger several indicator labels at once: metacognition, self-monitoring, flexible agency and introspection. Yet the sentence may be generated from prompt features while the confidence monitor has no causal role in action.

An adversarial report path that mimics metacognitive evidence A prompt flows directly to a report generator that emits introspective language. A confidence monitor is shown disconnected from the action controller. A dotted expected path contrasts with the solid executed path. Prompt andtask cues Report generatortrained on explanations Confidence monitorreal reliability signal Introspectivereport Actioncontroller × expected causal path executed mimicry path
Figure 8. The report can be right for the wrong causal reason.Only an intervention that changes the monitor while holding report cues fixed can show whether metacognitive information guides action rather than merely appearing in language.Illustrative adversarial path.

Recent debate reflects this tension. Pennartz argues that internal indicators need behavioural-cognitive supplementation and validation.[13] Butlin and colleagues reply that extended behaviour can still be gamed and that internal variants complicate inference, while accepting that future assessments must combine multiple incomplete sources.[14] The dependency graph provides a place to combine those sources without pretending that they are independent or equally transferable.

Strongest boundary

The method cannot solve theory uncertainty by organisation alone. The 2025 adversarial collaboration challenged key commitments of both GNWT and IIT while supporting some predictions of each.[5] A clean evidence graph may therefore show that Tern-4 satisfies a disputed theory better than another system. It does not make the theory true.

This boundary is productive. It changes the goal from finding a ceremonial winner to producing sharper conditional results. A good assessment can say: if computational higher-order theory is approximately right, and if this cross-substrate operationalisation preserves the relevant relation, then the intervention increases confidence that Tern-4 instantiates the specified mechanism. Every clause is visible. Each can be challenged separately.

The evidence conversation
Chart · sequence
sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The dashboard that lies
  M->>E: The arithmetic that creates confidence
  E-->>M: compare rival explanations
  E->>D: Remove the score when the graph cannot answer
  D-->>Q: revise scope or act
Question, mechanism, evidence and decision exchange the burdens developed in Indicators, not verdicts. Reading use: follow the relation, then return to the surrounding argument for evidence and limits.

Remove the score when the graph cannot answer

The practical decision is not whether to assign Tern-4 a consciousness percentage. It is which claim to permit, which experiment to run next and which precaution remains justified under unresolved uncertainty.

For research, choose the next intervention that crosses evidence families and separates rival theories. Testing a fourth output of the same workspace may add detail but little independence. Testing whether the confidence monitor changes general belief updates while report behaviour is fixed attacks a different causal link. Spend experiments on discriminating dependencies, not accumulating positives.

For communication, prohibit formulations such as “Tern-4 meets eight of fourteen indicators” unless the dependency graph and theory scope are shown beside the count. Prefer: “Tern-4 supports two workspace commitments through one implementation family; one broadcast prediction is challenged; higher-order monitoring remains contaminated.” That sentence is less dramatic and far more informative.

For governance and welfare, keep evidence assessment separate from the action threshold. A low but non-zero credence may justify reversible safeguards when possible harm is large and precautions are cheap. That is a decision-theoretic argument, not proof produced by the indicator graph. Responsible AI-consciousness research proposals similarly emphasise policies, transparency and staged precautions under uncertainty.[15]

Decision surface based on theory specificity and evidence independence A two-axis field places assessments into four regions: document only, refine the test, update a theory-relative claim, and consider precaution. The horizontal axis is evidence independence; the vertical axis is theory specificity. evidence independence theory specificity lowhigh lowhigh Refine the testspecific claim, dependent evidence Update the theory ledgerspecific claim, independent intervention Document onlygeneric resemblance or weak proxy Consider proportionate precautionindependent evidence plus asymmetric harm Tern-4 workspace family introspective report alone
Figure 9. Evidence quality changes the next action, not the metaphysical verdict.Independent, theory-specific evidence can justify a stronger conditional update. Precaution also depends on harm, reversibility and cost, which sit outside the indicator count.Decision instrument. Practitioner synthesis.

A decision meeting should therefore receive three outputs. The first is a claim ledger: the strongest supported, challenged and unresolved statements for each theory. The second is an evidence graph: the shared implementations, measurement paths and boundary choices that determine dependence. The third is an experiment queue: tests ranked by how much they separate rival generators or theories, not by how many extra green cells they promise.

Consider two proposals for the next study. Proposal A repeats the workspace task in ten more domains. It may test generality, but every result remains downstream of workspace W. Proposal B swaps the workspace while preserving the report interface, then asks whether delay integration, capacity limits and module availability dissociate. Proposal B risks a negative result, yet it creates more information because it can break the shared root into competing mechanisms. The best next test is the one that could reorganise the graph, rather than only decorate it.

The same rule disciplines public claims. An organisation may still publish a compact summary, but the summary should state the theory, boundary, evidence families, challenged predictions and live rivals. A reader can then disagree with a bridge assumption without disputing the observed mechanism. This is slower than displaying a percentage. It is also how an assessment remains corrigible when architectures, theories and ethical thresholds change.

Decision rule

Before accepting any artificial-consciousness assessment, ask: Which theory makes this property relevant? Which configured system is the candidate? Which intervention isolates the mechanism? Which positives share a cause? Which rival still survives? What exact decision changes if the result is positive, negative or unresolved?

If those questions cannot be answered, do not improve the score. Remove the score.

Glossary

Candidate boundary
The versioned set of components, states and interactions proposed as the possible subject of experience.
Indicator property
A property derived from a theory of consciousness that can alter credence when validly operationalised and tested.
Evidence family
A set of results sharing an implementation, intervention or measurement root.
Rival generator
An alternative mechanism that can produce the observation without the target property or inference.
Contamination
A condition in which training leakage, shared probes, report mimicry or another pathway prevents a clean inference.
Permitted conclusion
The strongest theory-relative statement licensed by the complete evidence chain.

References

  1. Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D. J., et al. (2026). “Identifying indicators of consciousness in AI systems.” Trends in Cognitive Sciences, 30(6), 488–501. https://doi.org/10.1016/j.tics.2025.10.011
  2. Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., et al. (2023). “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” arXiv:2308.08708. https://doi.org/10.48550/arXiv.2308.08708
  3. Bayne, T., Seth, A. K., Massimini, M., Shepherd, J., Cleeremans, A., Fleming, S. M., et al. (2024). “Tests for consciousness in humans and beyond.” Trends in Cognitive Sciences, 28(5), 454–466. https://doi.org/10.1016/j.tics.2024.01.010
  4. Seth, A. K., and Bayne, T. (2022). “Theories of consciousness.” Nature Reviews Neuroscience, 23, 439–452. https://doi.org/10.1038/s41583-022-00587-4
  5. Cogitate Consortium, Ferrante, O., Gorska-Klimowska, U., et al. (2025). “Adversarial testing of global neuronal workspace and integrated information theories of consciousness.” Nature, 642, 133–142. https://doi.org/10.1038/s41586-025-08888-1
  6. Melloni, L., Mudrik, L., Pitts, M., et al. (2023). “An adversarial collaboration protocol for testing contrasting predictions of global neuronal workspace and integrated information theory.” PLOS ONE, 18, e0268577. https://doi.org/10.1371/journal.pone.0268577
  7. Birch, J., Schnell, A. K., and Clayton, N. S. (2020). “Dimensions of animal consciousness.” Trends in Cognitive Sciences, 24(10), 789–801. https://doi.org/10.1016/j.tics.2020.07.007
  8. Mashour, G. A., Roelfsema, P., Changeux, J.-P., and Dehaene, S. (2020). “Conscious processing and the global neuronal workspace hypothesis.” Neuron, 105(5), 776–798. https://doi.org/10.1016/j.neuron.2020.01.026
  9. Brown, R., Lau, H., and LeDoux, J. E. (2019). “Understanding the higher-order approach to consciousness.” Trends in Cognitive Sciences, 23(9), 754–768. https://doi.org/10.1016/j.tics.2019.06.009
  10. Albantakis, L., Barbosa, L., Findlay, G., et al. (2023). “Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms.” PLOS Computational Biology, 19(10), e1011465. https://doi.org/10.1371/journal.pcbi.1011465
  11. Seth, A. K. (2025). “Conscious artificial intelligence and biological naturalism.” Behavioral and Brain Sciences, 1–42. https://doi.org/10.1017/S0140525X25000032
  12. Steinmetz Yalon, N., Goldstein, A., Mudrik, L., and Geva, M. (2026). “Indications of belief-guided agency and meta-cognitive monitoring in large language models.” arXiv:2602.02467. https://doi.org/10.48550/arXiv.2602.02467
  13. 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. https://doi.org/10.1016/j.tics.2026.01.011
  14. Butlin, P., Bayne, T., Fleming, S. M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., and VanRullen, R. (2026). “Consciousness indicators, mimicry, and internal variants.” Trends in Cognitive Sciences, 30(7), 575–576. https://doi.org/10.1016/j.tics.2026.04.006
  15. Butlin, P., and Lappas, T. (2025). “Principles for responsible AI consciousness research.” Journal of Artificial Intelligence Research, 82, 1673–1690. https://doi.org/10.1613/jair.1.17310