Published: 13 October 2026
Why words are not enough
A language model can say “I feel afraid” because those words fit a conversational context. That performance may be impressive, but it does not reveal whether anything is felt. The same output could be produced by systems with very different internal organization. A scientific test must therefore move beyond interviews. It should identify mechanisms predicted by theories of consciousness, measure them inside the system and ask whether controlled changes alter behaviour in the expected way.
From checklists to competing predictions
Researchers have proposed indicators derived from global-workspace, recurrent-processing, higher-order and predictive-processing theories. Useful indicators might include recurrent feedback, information available across subsystems, metacognitive access and a stable model of self and environment. No single indicator proves consciousness. The stronger approach is to specify rival explanations before testing and design an intervention that makes them predict different outcomes.
Intervention matters
Suppose an AI appears to integrate vision, memory and planning. Researchers could selectively interrupt the connections that broadcast information while preserving local task performance. If a theory predicts that global availability is required for consciousness, the intervention should remove theory-linked capacities in a reproducible pattern. Similar tests could probe recurrent loops or metacognitive monitoring. The point is causal evidence: changing the proposed mechanism should change the predicted signatures.
The calibration problem
Human and animal studies supply the reference cases, yet consciousness science itself remains divided. A major 2025 adversarial collaboration tested integrated information theory and global neuronal workspace theory with fMRI, MEG and intracranial recordings. The results supported some predictions while challenging central claims of both. If our biological theories are unsettled, confidence about machine experience must remain limited. AI evaluation should report uncertainty rather than turn a provisional score into a verdict.
A responsible protocol
A credible programme would publish model architecture details, preregister tests, use independent evaluators and separate behavioural ability from evidence about experience. It would also test for imitation: can the same “conscious” behaviour be elicited by prompting or training alone? Until converging evidence exists, current systems should be described by demonstrated capabilities. Ethical monitoring is sensible, but marketing language is not a consciousness detector.
Discussion question
Which result would be more persuasive to you: a machine describing an unexpected private experience, or a causal experiment showing the internal mechanism predicted by a theory?
Sources and further reading
Butlin et al. — Consciousness in Artificial Intelligence ↗Cogitate Consortium — Adversarial testing of consciousness theories ↗Share your answer
You must log in to reply.
Sign Up / Log InEstablished evidence, active research and philosophical interpretation are clearly separated.