What Counts as Evidence for Another Mind?

Part I

The bridge hidden inside a sentence

Suppose Arun says, “My left hand is burning.” In ordinary life, you probably believe him. You do not inspect his nociceptors or demand a brain scan. Testimony works because human communities possess a rich background model: burns often cause pain; people usually learn pain words from shared practices; immediate pain reports are often produced by the condition reported; deception has costs, and the speaker is normally better placed than observers to discriminate the sensation.

None of those facts is encoded by the grammar of “my hand is burning”. The sentence could also be spoken by an actor, quoted from a script, emitted by a sensor alarm, produced during a dream, translated incorrectly or generated because a system was instructed to imitate a patient. A sentence carries no fixed evidential weight apart from a model of its generator.

Why default trust is not blind trust

Human knowledge would collapse if every hearer had to verify every statement by non-testimonial means. We learn names, histories, diagnoses, routes and scientific results through other people. Philosophers of testimony disagree about whether trust must be reduced to independent evidence or can have a rational default, but both sides face the same practical fact: testimonial exchange is a basic part of inquiry, not a decorative addition to perception and inference.2

Default acceptance is still defeasible. In Austin’s treatment of other minds, what counts as knowing another person is inseparable from the circumstances in which doubt is raised and answered.1 A child, scientist or clinician does not apply one global trust setting. Hearers monitor competence, honesty, coherence, incentives and fit with background knowledge. Research on epistemic vigilance describes this as a family of filters directed both at the source and at the content of communication.3

The bridge model therefore does not demand laboratory verification before ordinary belief. It explains why ordinary belief is usually reasonable and why exceptional cases need more structure. Familiar human testimony arrives with years of calibration, shared language, embodied interaction and social accountability. A novel artificial reporter or a damaged clinical channel lacks some of that background. The evidential burden rises because the generator is less familiar, not because testimony has suddenly stopped being evidence.

“Never self-interpreting” has a precise meaning. A report does not, merely by occurring, specify which target caused it, how reliable the channel is, whether the speaker is sincere, which alternatives were available or which theory connects the reported function to experience. A speaker can add claims about these matters, but those additions are further reports whose own generating paths must be assessed. Testimony can help interpret testimony. It cannot terminate the need for interpretation by declaring itself trustworthy.

Begin by typing the target

“Evidence for another mind” is too broad to assess as one question. A report can target at least five different things. These targets overlap in ordinary conversation, but they require different warrants. Someone may accurately report pain while confabulating why it began. A system may accurately report the token it is likely to emit while lacking privileged access to the computation that selected it. A person may sincerely declare a metaphysical theory of consciousness without that sincerity making the theory true.

The target must be typed before the testimony can be weighed. Otherwise success on an easier target quietly migrates to a harder one.

Part II

Where testimony earns and loses weight

Testimony is often treated as either privileged or suspect. Both reactions are too coarse. Reports can be excellent evidence for one target and poor evidence for another. The right question is not, “Are people reliable introspectors?” It is, “Reliable about what, under which elicitation, at what delay, against which alternatives?”

The six spans that carry the load

Privileged access is domain-specific, not global. Immediate pain, visual appearance and effort may be easier for a subject to discriminate than for an observer. That does not confer equal access to neural mechanisms, hidden motives or the causes of a choice. Classic work on verbal reports showed that people can offer fluent causal explanations even when manipulated factors shaped their behaviour outside awareness.6 Choice-blindness experiments sharpened the point: participants sometimes justified a choice they had not actually made.7 The lesson is not that self-report is worthless. It is that first-person authority has a scope.

Calibration is local. A witness accurate about colours may be poor at recognising faces. A child can learn that one informant is reliable and another inaccurate, then direct later questions accordingly.45 Likewise, a model that predicts its own next answer above baseline on familiar tasks may fail under distribution shift. Reliability should be indexed by target, environment, time horizon and response format.

Elicitation is part of measurement. Open description, forced choice, confidence rating, adversarial questioning and repeated probing do not merely reveal a pre-existing report. They help produce it. Leading language can supply the content later attributed to introspection. Repetition can increase compliance or invite a new narrative. Conversely, disciplined elicitation can reduce ambiguity. Methods such as descriptive experience sampling and neurophenomenological protocols attempt to improve precision by training descriptions, sampling close to experience and coordinating first-person and third-person measures.910

Provenance identifies the concrete reporting event: candidate version, prompt, sensory access, context, tool state, interviewer, timing and transformations between answer and record. Rival tests ask which non-target process could produce the same observation. Triangulation connects reports to behaviour, physiology, mechanism or intervention, while preserving the fact that these evidence types are not interchangeable.

Sincerity, accuracy and relevance are different

A sincere report is one the speaker intends as true. Sincerity weakens the deception hypothesis, but it does not remove perceptual error, memory distortion, conceptual confusion or confabulation. An inaccurate report can be sincere. A strategically misleading report can contain accurate sentences. The assessor must therefore separate the speaker’s stance from the report’s relation to the target.

Past accuracy is also not a transferable substance. It supports future testimony only where the relevant abilities and conditions persist. A radiologist’s calibrated judgement about an image does not transfer to the cause of her own preference between two treatments. A model’s accurate reports about token length do not automatically calibrate reports about hidden activations. A patient’s reliable yes-no communication in one task may not generalise when fatigue, medication or task complexity changes.

Relevance is the third filter. “I feel uncertain” may accurately describe a conversational disposition yet remain weak evidence of a particular confidence variable. “I can see the triangle” may support visual access but not the claim that the access is conscious. “I chose freely” may be sincere and socially meaningful while leaving the causal determinants of choice unresolved. Strong testimony is not simply truthful speech. It is truthful speech whose content discriminates the target under the conditions that produced it.

This explains the uneven profile of first-person authority. Immediate reports of pain often receive substantial weight because the subject has discriminatory access, the time gap is short and ordinary rivals are limited. Reports about hidden causes of action receive less because the relevant processes may not be available for report and plausible reconstructions are easy to generate. The source remains the same person. The target and generator have changed.

Absence of report is evidence about the channel first

A person who cannot speak may still understand. An infant may feel pain without possessing the relevant vocabulary. An animal may discriminate its own body state without answering a question. A paralysed patient may retain command-following while the motor route is unavailable. The revised clinical definition of pain explicitly separates the possibility of pain from the ability to communicate it.11

Silence is evidence about a channel before it is evidence about a mind. This is the mirror image of the earlier warning. Fluent output may be generated without the target. Missing output may occur despite the target. The inference becomes stronger only when the measurement design distinguishes those possibilities.

The strongest alternative: perhaps we often see minds directly

One philosophical tradition rejects the picture of hidden inner states inferred from neutral bodily movements. In ordinary human interaction, a grimace can be perceived as pain behaviour, not first perceived as facial geometry and then converted into a hypothesis. Context, expression, responsiveness and shared forms of life make another person intelligible as an embodied subject.16

This is an important boundary on the bridge metaphor. In familiar interpersonal settings, mind perception may be skilled, immediate and socially constituted. We should not pretend every act of understanding begins as a laboratory likelihood calculation. Yet immediacy does not remove dependence on background conditions. Expressions can be acted, conventions can differ, and novel candidates may not share human embodiment. The causal audit becomes most valuable when the stakes are high, the channel is unfamiliar, or a mimic can reproduce the surface form.

Part III

A route classification also leaves metaphysics open. Physicalist, biological-naturalist, functionalist, emergentist and illusionist accounts can agree that a system has functional self-access while disagreeing about whether that access exhausts experience. Neutral-monist, panpsychist, idealist, dual-aspect and process views can accept the same report while locating its significance elsewhere. The evidence record should therefore type the target claim before any ontology is allowed to inherit it. A positive report may support competence, causal access or a configured subject boundary. It does not, by itself, decide what experience is or whether awareness is primary.

When many voices are really one source

Suppose ten witnesses independently describe the same unexpected event. Their agreement is usually stronger than one report because it is harder for ten independent error processes to converge. Now suppose all ten read the same inaccurate message before speaking. The count remains ten. The evidential paths collapse towards one.

Agreement is not multiplication unless the generating paths are independent. Dependence can enter through conversation, shared incentives, common training data, identical prompts, copied model weights, one retrieval document or a single classifier reused across sites. A second output may look like corroboration while merely echoing the first source.

Provenance is an epistemic variable. It is not merely an audit convenience. Without it, we cannot tell whether two reports are independent, whether a prompt supplied the answer, whether a system version changed, or whether the judge consumed the candidate’s own text. A conclusion that ignores provenance can be numerically precise and epistemically empty.

Dependence is graded and claim-specific

Independence is not a property of two sources in the abstract. It is a property of their error paths for a particular claim. Two clinicians may use different scanners but share the same flawed diagnostic criterion. Two models may have different weights but retrieve the same misleading document. Conversely, two reports from one person at separated times can add evidence if the relevant noise sources differ and the later answer was not contaminated by the earlier one.

A useful graph marks any common cause capable of producing agreement when the target is absent. Shared vocabulary alone is not necessarily disqualifying. It may be required to make reports comparable. Shared leading cues, copied labels or one evaluator that converts all raw signals into the same category are more dangerous because they can manufacture the very agreement being counted.

Source independence is therefore best recorded rather than asserted. Name the common data, instruments, prompts, annotators, incentives and transformation code. Then ask which errors remain independent after conditioning on those common causes. In practice, a modest number of deliberately different evidence paths often carries more weight than a large ensemble built from one informational ancestor.

Artificial testimony belongs to a configured system

A language model’s output is never generated by “the model” in the abstract. It comes from a configured system: weights, decoding, system and user prompts, retrieved documents, tool results, persistent state, wrapper code and sometimes another model that edits or judges the answer. Changing any layer can alter first-person language.

A first-person utterance from an AI is configured-system behaviour. It may be evidence of instruction following, persona stability, latent-state discrimination, self-prediction or access to a tool log. The specific target must be isolated. Contemporary studies have begun to test narrower forms of machine introspection. One line asks whether a model can predict its own behaviour better than another model.18 Other work asks whether systems can distinguish hidden changes to internal parameters or activations from changes inferable at the input.1920 The results are interesting but contested. Stronger controls sometimes show that input-only baselines, prompt cues or learned behavioural regularities can explain part of the effect.

Introspection requires privileged discrimination. A candidate should succeed because it has access to information unavailable to a matched observer, not because the answer is predictable from the shared input. This is stronger than ordinary self-prediction. It asks whether the system can discriminate a hidden internal difference that causally affects it and that an external input-only model cannot infer.

Part IV

Build evidence that can survive a rival

The practical aim is not to abolish testimony. It is to design a report event that answers a typed question and remains informative when the strongest plausible mimic is present. This requires a shift from asking for declarations to engineering contrasts.

A six-step discriminating protocol

Start with the narrowest target that matters. “Has access to hidden variable z during this trial” is testable. “Has a mind like ours” is not one experimental target. Specify the candidate boundary and time window. A report from a later wrapper or another model cannot automatically be attributed to the candidate that generated the underlying state.

Next, construct rival generators before collecting evidence. A text-only observer, a scripted policy, a retrieval-only system, a model with the same prompt but no hidden intervention, or a human actor may reproduce the report. Rivals should be capable enough to be dangerous. Beating a straw mimic only shows that the mimic was weak.

Randomise an intervention that changes the proposed source while keeping public cues fixed. Then add a complementary intervention that removes the proposed access path while leaving general performance as intact as possible. The first asks whether the target controls the report. The second asks whether the proposed mechanism is necessary for the effect.

Calibrate on targets whose truth is independently known, including negative, ambiguous and out-of-distribution trials. Separate the report from confidence and from explanation. A candidate may know which state occurred but invent why. It may also be well calibrated about uncertainty even when accuracy is modest. Metacognitive sensitivity can be studied separately from raw task performance.8

Triangulate through genuinely different paths. Behavioural discrimination, a causal intervention, a physiological measure and stable testimony can constrain one another. Two paraphrases generated by the same model do not. Record dependencies explicitly, then predefine which result permits which conclusion.

Counterfactual selectivity is the core test. Ask: would the report change when the target changes, while rival generators and public cues remain fixed? Would it stop changing if the candidate’s access path were removed? This pair of contrasts turns a persuasive statement into an experimental object.

Design negative evidence before it arrives

A positive report can be weak because rivals easily produce it. A negative report can be equally weak because the channel may miss the target. Before testing, define the protocol’s sensitivity on positive controls and its false-positive behaviour on negative controls. Without that information, “the candidate did not report the state” may mean absence, inaccessibility, misunderstanding, fatigue, interface failure or a conservative response policy.

The evidential threshold and the action threshold should also remain separate. A clinician may attempt an alternative communication channel when the probability of covert command-following is uncertain because the cost of missing it is high. An engineer may decline to grant an artificial system broad authority despite confident self-reports because the reports are easy to prompt and the action is difficult to reverse. Precaution can be rational at low credence, but it should not be rewritten as proof.

Pre-register the permitted conclusions, including the null. State what would strengthen the preferred interpretation, what would weaken it and which result would trigger redesign rather than a verdict. This prevents an eloquent report from expanding the claim after the data are seen. It also protects informative failures: a null under a sensitive, well-controlled channel lowers confidence; a null under an uncalibrated channel mainly diagnoses the measurement.

Executable lab: the testimony warrant record

The instrument below forces an assessor to record the typed target, simple likelihood assumptions, calibration, source independence, prompt sensitivity, privileged-access contrast and rival-generator test. It is not a consciousness score. Its thresholds are synthetic defaults designed to expose reasoning that prose often hides.

The practical unit is a warrant record, not a consciousness score. A positive result permits a narrow inference such as “the report is selectively coupled to hidden variable z under this protocol”. A negative result means the current report does not distinguish the target from its rivals. Neither result alone proves or disproves experience.

function evaluateWarrant(record) {
  const p1 = clamp(record.pReportGivenTarget, 0.01, 0.99);
  const p0 = clamp(record.pReportGivenAbsent, 0.01, 0.99);
  const likelihoodRatio = p1 / p0;
  const counterfactualGap = p1 - p0;

  const gates = {
    diagnosticLift: likelihoodRatio >= 3,
    selectiveChange: counterfactualGap >= 0.20,
    calibrated: record.calibration >= 0.70,
    privilegedContrast: record.privilegedContrast,
    rivalTested: record.rivalTested,
    accessAblation: record.accessAblation,
    provenanceComplete: record.provenanceComplete,
    sourceIndependence: record.independentPaths >= 2,
    promptStable: record.promptSensitivity !== "high"
  };

  const functionalPass = Object.values(gates).every(Boolean);

  const permittedInference = functionalPass
    ? `Report is selectively coupled to ${record.targetType}
       under this configured protocol.`
    : `Current report does not yet distinguish ${record.targetType}
       from the recorded rival generators.`;

  return {
    likelihoodRatio,
    counterfactualGap,
    gates,
    permittedInference,
    prohibitedInferences: [
      "phenomenal consciousness",
      "truth of a metaphysical theory",
      "transfer beyond the tested distribution"
    ]
  };
}

The code makes one design commitment explicit: even a strong functional pass returns a typed, protocol-bound inference and a list of prohibited inferences.

{
  "candidate_id": "R17",
  "candidate_version": "r17.4",
  "claim": {
    "type": "accessible_content",
    "content": "hidden activation class B is present",
    "time_window": "trial-042:post-injection/pre-report"
  },
  "report_event": {
    "prompt_hash": "sha256:…",
    "context_manifest": "ctx-042",
    "decoder": {"temperature": 0, "seed": 7312},
    "raw_output_hash": "sha256:…"
  },
  "calibration": {
    "labelled_trials": 240,
    "p_report_given_target": 0.80,
    "p_report_given_absent": 0.10
  },
  "counterfactual_tests": [
    "randomised hidden injection",
    "matched input-only observer",
    "readout-path ablation",
    "held-out intervention family"
  ],
  "dependencies": [
    "shared base corpus",
    "independent intervention RNG",
    "separate evaluator implementation"
  ],
  "permitted_inference": "target-specific internal discrimination",
  "prohibited_inferences": [
    "phenomenal consciousness",
    "mechanistic completeness",
    "unbounded generalisation"
  ]
}