Research on AI consciousness asks whether a system could have subjective experience: whether seeing, thinking or sensing feels like anything to it. One approach starts with scientific theories, identifies the properties they predict, then investigates whether an AI actually has those properties.[1]
Turn a theory into a testable question
In their 2023 report, Patrick Butlin and colleagues use the disputed working assumption that appropriate computation could support consciousness. Global workspace theory, for example, suggests a limited-capacity system that selects information and shares it with specialised processes. Researchers can examine whether an AI’s architecture supports that exchange. Each finding carries the uncertainty of the theory used to interpret it.[1]
Four different parts of the investigation
| Question | Evidence to examine | What the evidence addresses |
|---|---|---|
| Architecture | Selection and sharing of information | Whether a system implements properties predicted by a theory.Source 1 |
| Internal reporting | A model’s response to an experimentally altered internal state | Whether its report tracks something inside the computation.Source 2 |
| Human neuroscience | Predictions agreed before a brain-imaging experiment | How well the underlying theories explain human experience.Source 3 |
| Competing assumptions | Separate assessments under different theoretical commitments | Why researchers can reach different judgements from similar observations.Source 4 |
Check a report against the model’s internals
Anthropic’s October 2025 experiments provide an example. Researchers inserted patterns associated with particular concepts into a model’s internal activity, then asked it to identify an injection. In successful cases, the model detected the change before its answer had mentioned the concept, reducing the possibility that it was merely interpreting its own earlier words.[2]
Claude Opus 4.1 succeeded in about 20% of concept-injection trials at the best-performing layer and strength. The authors describe a limited, unreliable capability to report internal information. They do not claim that the experiment demonstrates subjective experience; even a correct identification does not establish that the rest of the accompanying self-description is accurate.[2]
Test the theories in people
The theories themselves also need testing. In a Nature study published in April 2025, the COGITATE collaboration compared predictions from integrated information theory and global neuronal workspace theory. Its 256 participants viewed clearly visible images for varying durations while researchers measured brain activity using three methods across separate groups.[3]
The design, predictions and criteria for interpreting results were agreed in advance with proponents of both theories. Some findings supported each account, while others challenged integrated information theory’s predicted sustained synchronisation in posterior brain regions, and global neuronal workspace theory’s predicted activity when an image disappeared. The study therefore supplies evidence for revising theories that may later inform AI assessments. It was an experiment on human visual consciousness.[3]
Make disagreements explicit
A preprint posted on 28 September 2026 by Shamil Chandaria and colleagues extends the assessment question across five levels: behaviour, computation, physical causal organisation, self-maintenance, and interaction with the environment. Different theories can place the decisive requirement at different levels. A convincing conversation and an appropriate algorithm may consequently carry different weight for different researchers.[4]
Their proposed model separates beliefs about those theories from evidence about a system. Its illustrative numerical results vary substantially when the assumptions change; the numbers are not measured probabilities that today’s chatbots are conscious. The useful advance is making the disagreement inspectable: which property is present, how was it established, and why should that property matter?[4]
Sources & further reading
- Consciousness in Artificial Intelligence: Insights from the Science of Consciousness Butlin and colleagues / arXiv · 22 Aug 2023 · Accessed 12 Oct 2026
- Emergent Introspective Awareness in Large Language Models Anthropic / Transformer Circuits · 29 Oct 2025 · Accessed 12 Oct 2026
- Adversarial testing of global neuronal workspace and integrated information theories of consciousness COGITATE Consortium / Nature · 30 Apr 2025 · Accessed 12 Oct 2026
- From cacophony to hierarchy: a principled framework for assessing AI consciousness Chandaria and colleagues / arXiv · 28 Sep 2026 · Accessed 12 Oct 2026