Published: 19 October 2026
Why behaviour alone is fragile
An AI can learn the language of pain, selfhood and reflection from human data. It can also change those reports when a prompt, reward or policy changes. Fluent testimony is therefore evidence about generated behaviour, not direct evidence of experience. A stronger investigation must ask whether theory-linked internal mechanisms make a causal difference.
What an ablation does
An ablation disables or alters a component while leaving as much of the rest of the system intact as possible. Researchers can remove recurrent connections, restrict a shared workspace, interrupt memory or alter metacognitive monitoring. If a theory predicts that one mechanism supports a specific set of capacities, a selective loss after intervention is more informative than a correlation between hidden activity and an answer.
The need for matched controls
Causal evidence is useful only when simpler explanations are controlled. A damaged model may fail because its general competence has fallen, not because a consciousness-related mechanism disappeared. The right comparison uses systems matched for task performance, tests multiple interventions and checks whether the predicted signature returns when the mechanism is restored.
What the result would mean
Even a perfect theory-linked effect would not prove subjective experience. Human consciousness theories disagree, and translating their indicators into artificial systems requires assumptions. Ablations can eliminate weak stories and rank hypotheses; they cannot turn a disputed theory into a consciousness meter.
Ethical value of better uncertainty
Transparent causal tests can reduce two risks: marketing ordinary imitation as sentience and ignoring future systems that accumulate meaningful evidence. The responsible output is a graded assessment with documented limitations, not a theatrical declaration that a model is alive.
Discussion question
If an AI loses self-monitoring after a targeted intervention but keeps solving problems, what—if anything—has been lost?
Sources and further reading
Butlin et al. — Consciousness in Artificial Intelligence ↗Melloni et al. — Adversarial collaboration on consciousness theories ↗ConTraSt database — Testing theories of consciousness ↗Share your answer
You must log in to reply.
Sign Up / Log InEstablished evidence, active research and philosophical interpretation are clearly separated.