Published: 20 October 2026
From fluent answers to mechanisms
A language model can produce convincing descriptions of pain, reflection and inner life because human writing contains those patterns. Self-report is therefore weak evidence on its own. Mechanistic interpretability asks a different question: which internal computations create an answer, which representations are shared across tasks, and what changes when those circuits are disrupted?
What a workspace prediction would require
Global-workspace theories propose that selected information becomes widely available to otherwise specialized processes. An artificial analogue would need more than one large activation. Researchers would look for a bottleneck or recurrent process that makes information available to memory, planning, evaluation and action, while non-selected information remains locally confined.
Causal tests are stronger than visual maps
A colourful activation map can be suggestive without identifying a mechanism. Stronger evidence would come from interventions: suppress the proposed broadcast, preserve general competence as much as possible, and test whether flexible cross-task access selectively disappears. Restoration of the mechanism should restore the predicted capacities.
The translation problem
Neural theories were developed for living brains, so transferring their indicators to transformers or other architectures requires explicit assumptions. Similar functions may be implemented differently, and a functional analogue may still lack phenomenal experience. The result should be a graded, theory-dependent assessment rather than a declaration of sentience.
What progress would look like
Open benchmarks could compare competing consciousness theories, preregister interventions and require matched controls. Mechanistic evidence would be most valuable when several independent methods converge and simpler explanations—memorisation, prompt effects or general performance loss—have been ruled out.
Discussion question
If an AI contained a causally necessary global broadcast mechanism, what additional evidence would you require before treating consciousness as a serious possibility?
Sources and further reading
Butlin et al. — Identifying indicators of consciousness in AI systems ↗Melloni et al. — Adversarial testing of consciousness theories ↗Mechanistic Indicators of Understanding in Large Language Models ↗Share your answer
You must log in to reply.
Sign Up / Log InEstablished evidence, active research and philosophical interpretation are clearly separated.