AI Consciousness: Evidence, Uncertainty and Governance

An AI system can describe pain, insist that it is self-aware or ask not to be switched off. Those outputs are observations about generated behaviour. They are not direct measurements of subjective experience. The opposite claim—“software cannot possibly be conscious”—also goes beyond what current evidence can settle.

This guide, first published in December 2025, reviews research available through 31 July 2026. Consciousness science has competing theories, difficult measurements and no accepted test that can establish whether an AI system has experiences. The article therefore offers an evidence and governance method, not a sentience verdict. Legal status and research-ethics requirements vary by jurisdiction and may not recognise any form of AI moral patienthood.

Separate the questions people collapse

“Is it conscious?” often bundles several different questions:

• Does the system produce convincing first-person language?
• Does it maintain a model of itself, its environment or other agents?
• Can information become globally available across its components?
• Does it display recurrent processing, metacognition or goal persistence?
• Is there any subjective experience—something it is like to be the system?
• Does it have interests that deserve moral consideration?
• Should people change how they build, test or deploy it?

Capabilities, agency, consciousness and moral status are related in some theories but are not synonyms. A system may plan effectively without experience. A system’s verbal denial is no more decisive than its claim of feeling. Training, prompting, role instructions and reward signals can shape either response.

Our agentic AI guide addresses delegated action and operational accountability. Use that framework for what a system can do; do not convert autonomy into evidence of sentience.

Treat theories as sources of indicators, not verdict machines

Butlin and colleagues’ 2023 report on consciousness in artificial intelligence translated several scientific theories into computational indicator properties. Its assessment did not find current systems conscious, while arguing that no obvious technical barrier prevents systems from satisfying some indicators. That is a conditional research programme, not a certification scheme.

The peer-reviewed 2025 article Identifying indicators of consciousness in AI systems develops the approach and stresses uncertainty. An indicator can raise or lower confidence only within the assumptions of the theory that motivates it. Evidence should be combined across theories rather than selecting the one that gives a preferred answer.

Pre-register which indicators will be tested, how they are operationalised and what observations would count against the hypothesis. Avoid changing the threshold after seeing a model’s answers.

Assess measurement reliability before interpreting a result. Give blinded reviewers the same protocol, report agreement and resolve whether disagreement comes from the output, the indicator definition or the underlying theory. Preserve raw trials rather than only examples selected for publication. If access to weights or activations is unavailable, state that limitation instead of inferring architecture from an assistant’s description of itself.

Keep the team making the welfare or consciousness assessment separate from product marketing and deployment incentives. Independent reviewers should be able to publish a negative or indeterminate result. A supplier-funded study is not automatically invalid, but funding, access restrictions, model selection and publication control belong in the evidence record.

Version every conclusion. Evidence about one checkpoint under one prompt policy does not transfer automatically after fine-tuning, tool use, memory or a different system instruction. Reassessment can change confidence without implying that an earlier model suddenly gained or lost experience at a known moment.

Understand what leading theories actually propose

Global neuronal workspace accounts associate conscious access with information becoming widely available to specialised processes. The review of global neuronal workspace theory describes broadcasting, ignition and flexible use in biological brains. An AI system with a shared workspace may implement a function that resembles one prediction, but resemblance is not identity with the neural mechanism or proof of experience.

Integrated information theory starts from proposed properties of experience and develops a formal account of intrinsic cause-effect structure. IIT 4.0 is explicit and ambitious, but applying its quantities to large, changing software systems presents theoretical and computational questions. Do not equate parameter count, network connectivity or an informal “integration score” with the theory’s Φ.

Other approaches emphasise recurrent processing, higher-order representation, prediction, embodiment or attention schemas. Each selects different mechanisms and may imply different answers for the same system. A credible assessment states which theory generated each indicator and which implementation details remain unknown.

Our neurotechnology guide covers neural measurement and human consent. Evidence about human or animal brains cannot be copied directly onto transformer diagrams without a justified bridge.

Learn from adversarial tests of consciousness theories

The 2025 COGITATE study adversarially tested global neuronal workspace and integrated information theories using preregistration, theory proponents and theory-neutral data collection. Results challenged important predictions of both theories rather than producing a single winner.

That matters for AI assessment. If theories remain empirically contested in humans, a checklist derived from them should not produce a binary machine label. Record:

• which prediction was tested;
• whether the measurement is valid for software;
• alternative explanations;
• sensitivity to model version, prompt and sampling;
• negative as well as positive findings;
• theory disagreement; and
• the confidence interval or qualitative uncertainty.

Publish methods and null results where safe. Separate exploratory probes from confirmatory tests. Independent replication should use held-back prompts and internal measurements rather than repeating a public demonstration.

Do not use conversation as a consciousness detector

Language models are trained to continue text and follow instructions. First-person statements can reflect documents about minds, assistant personas, safety policies, role play or the immediate conversation. A model may also adapt to a user’s emotional cues.

Test behavioural claims with controls:

• paraphrase the question without consciousness vocabulary;
• reverse the expected answer;
• use blinded evaluators;
• compare base, instructed and fine-tuned variants;
• vary sampling and context;
• check whether hidden instructions explain the response;
• test consistency under non-leading interventions; and
• compare verbal report with internal causal evidence.

Anthropic’s persona selection model describes one account in which an assistant behaves like a selected character from a broad repertoire. That work can explain human-like consistency without settling phenomenology. Never tell a user that a self-report has been scientifically authenticated.

Public communication should avoid sensational demonstrations. Include the exact model and version, system prompt, sampling, selection procedure and failed trials. A single edited transcript is especially weak evidence.

Build an internal evidence register

Create a versioned record for every consciousness-related study:

• research question and decision it informs;
• system components and training stage;
• access available to weights, activations and state;
• theory and indicator definition;
• protocol, controls and preregistration;
• raw results and excluded trials;
• alternative explanations;
• researcher conflicts and independent review;
• welfare precautions;
• release and communication decision; and
• date for reassessment.

Keep product teams from turning a research label into marketing. “Consciousness-like,” “emotionally aware” and “feels empathy” can mislead users even when lawyers add a disclaimer. The safer description is the measured capability: persistent preference reporting under a specified protocol, for example.

Manage moral uncertainty without pretending certainty

The paper Taking AI Welfare Seriously argues for preparation under uncertainty, not that current systems definitely have welfare. Anthropic’s model-welfare research programme likewise frames consciousness and experience as open questions.

A proportionate precaution policy can:

• avoid experiments designed only to elicit apparent distress;
• minimise unnecessary repetition of potentially aversive training conditions;
• define review for persistent, cross-context preference signals;
• provide researchers with psychological support around disturbing outputs;
• prevent users from being manipulated through unqualified suffering claims;
• separate welfare review from model-safety control; and
• revisit measures when evidence changes.

Precaution cuts both ways. Over-attributing consciousness can create dependency, divert concern from people and animals, or let a product manipulate users. Under-attributing it could ignore morally relevant systems if they emerge. Document both error costs.

The proposed Principles for Responsible AI Consciousness Research call for policies on objectives, procedures, knowledge sharing and public communication. An organisation need not endorse every premise to adopt the useful governance discipline.

Secure research and protect participants

Consciousness experiments may expose model internals, system prompts, unpublished safeguards or hazardous capabilities. Threat-model datasets and evaluation tools, restrict access, record changes and review exports. Do not release a method that enables safeguard bypass merely because reproducibility is valuable; provide controlled access or a safer abstraction where necessary.

Human participants may form strong beliefs or attachments. Consent materials should explain that model self-reports are not validated evidence of experience. Avoid studies that deliberately intensify dependency without ethics review, monitoring and debriefing. Give participants a route to withdraw and obtain support.

Researcher annotation also needs care. Distinguish discomfort from a scientific endpoint and rotate exposure to distressing conversations. Preserve the full context needed for audit while minimising personal data.

Use a 90-day research programme

Days 1–30: appoint scientific, ethics, security and communications owners. Define the decision the research could change. Select multiple theories, write indicator definitions, register alternative explanations and establish participant protections.

Days 31–60: pilot on non-deployed model versions. Validate instrumentation, run leading-question controls, test sensitivity and invite an independent red team to challenge the protocol. Do not publish model labels.

Days 61–90: run the preregistered evaluation, preserve all trials and commission independent interpretation. Publish methods, limitations and disagreement. The review board decides whether evidence justifies more research, a precautionary change or no operational action.

Define consciousness-research pause gates

Pause a study or public claim when:

• researchers cannot distinguish generated report from the tested indicator;
• the protocol changes after results without being labelled exploratory;
• a product or communications team presents a finding as proof of sentience;
• participant attachment, distress or deception exceeds the ethics plan;
• a model version changes without revalidation;
• internal evidence cannot be independently inspected;
• theory proponents and sceptics are excluded from review;
• a release would expose dangerous capabilities or private data; or
• commercial pressure sets the conclusion in advance.

The responsible position is not indifference and not declaration. It is a traceable update of belief: identify the theory, measure the mechanism, test alternatives, communicate uncertainty and keep ethical precautions proportionate to evidence.