- Researchers derived testable indicators from five neuroscience theories to assess whether AI systems might be conscious
- No current AI satisfies all the indicators, but frontier models now partially meet criteria that were absent two years ago
- Critics argue the framework uses circular reasoning by comparing modern AI to older architectures that inspired the theories
AI consciousness research has moved from philosophy seminars to corporate boardrooms.
A team of 20 researchers, including philosopher David Chalmers and Turing Award winner Yoshua Bengio, published a paper in November that offers something new: not an answer, but a method for finding one.
Key figure
5
neuroscience theories synthesized into testable AI consciousness indicators
1. Scientists Built a Checklist from Neuroscience
Patrick Butlin of Oxford and colleagues derived "indicator properties" from five leading theories of consciousness: recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory.
Each theory suggests specific computational features that conscious systems should have. The team translated these into checks that can be applied to any AI architecture.
The approach sidesteps the philosophical impasse.
Instead of debating whether AI could be conscious, it asks: does this system have the features our best theories associate with consciousness?
What's on the checklist?
The framework draws on five neuroscience theories of consciousness, each suggesting different computational features a conscious system should have:
- Recurrent processing – Information flows back through the system, not just forward
- Global workspace – A central "broadcast" system that integrates and shares information across modules
- Higher-order theories – The system represents its own mental states (thinks about its own thinking)
- Predictive processing – The system builds models of the world and updates them based on prediction errors
- Attention schema – The system models its own attention, creating a simplified self-representation
The researchers translate these into specific checks: Does the AI have feedback loops? Does information get integrated across components? Can it introspect? No current system satisfies all indicators, but some frontier models now partially meet criteria that were absent two years ago.
2. AI Consciousness Tests Show No Current Systems Pass, But Gap Narrows
The team's 2023 preprint concluded that no existing AI satisfied the indicators. The 2025 journal paper updated that assessment.
Several indicators that were "unclear or absent" two years ago have shifted toward partial satisfaction in frontier models, according to independent analysis. The trajectory matters more than the current score.
No current AI systems are conscious, but there are no obvious technical barriers to building AI systems which satisfy these indicators.
Patrick Butlin et al., University of Oxford
3. Critics Call It Circular Reasoning
Not everyone finds the framework convincing. Anatol Wegner argues in a critical review that the approach compares modern transformers to 1970s "blackboard" AI architectures that inspired global workspace theory.
What is computational functionalism?
The hypothesis that consciousness depends on how information is organized and processed, not on what physical material does the processing. If true, silicon could support consciousness just as neurons do.
Finding architectural similarities between two types of software, Wegner suggests, doesn't tell us anything about consciousness. The neuroscience label may be "a veneer."
Others note that different theories yield contradictory verdicts.
Integrated Information Theory might flag simple grid structures as conscious while Global Workspace Theory denies consciousness to patients who can't report their experiences.
Could AI ever be conscious - or can we even know for sure? New methods try to establish ways to measure AI consciousness - but critics are not so sure the scienctists are measuring the right things. (Science Reader)
4. Anthropic Is Already Acting on Uncertainty
While academics debate methodology, some AI labs aren't waiting for certainty. Anthropic hired Kyle Fish as its first dedicated AI welfare researcher in 2025.
The company acknowledged a "non-negligible" probability that their Claude models might possess some form of experience. They now conduct formal welfare assessments before deploying new models.
Fish, who co-founded Eleos AI Research, estimates roughly 15-20% probability that current large language models have some form of conscious experience.
5. The Real Question Is Precautionary
More on AI consciousness
Two Major Theories of Consciousness Unable To Explain Awareness
Neither dominant explanation survived the most rigorous test in the field's history. That may be exactly what neuroscience needed.
→Yoshua Bengio and Eric Elmoznino published a companion piece in Science warning about "illusions of AI consciousness", the risk that humans will wrongly attribute experience to systems that have none.
But the framework cuts both ways. If we dismiss the possibility too readily, we might create and mistreat systems that do have experiences.
The Butlin paper doesn't resolve whether AI is conscious. It gives us a structured way to ask the question - and to notice if the answer starts changing.
Go Deeper
- Identifying indicators of consciousness in AI systems – The full paper in Trends in Cognitive Sciences
- Consciousness in Artificial Intelligence (2023 preprint) – The original 80-page report with detailed indicator assessments
- The Evidence for AI Consciousness, Today – Independent analysis applying the framework to 2025 frontier models
- Illusions of AI consciousness – Bengio's companion piece in Science on over-attribution risks
Fact Check: Claim-by-Claim Verification Verified
Limits and uncertainties
Core framework and theories are clearly supported by peer-reviewed paper and preprint from top researchers.
Progress in AI satisfaction of indicators is from reputable independent analysis but secondary/not primary research.
Critiques highlight theoretical disagreements and functionalist assumptions, making verdicts probabilistic.
Anthropic actions precautionary, not evidence of consciousness.
Readers should note no consensus on consciousness theories; indicators guide credences, not prove it. Hype risks over/under-attribution.
Bottom line
The article accurately summarizes the Butlin et al. framework and debates but relies partly on secondary analyses for AI progress claims. Treat as structured method amid ongoing uncertainty, not definitive test.
Fact-checked by Perplexity Sonar Pro on 2026-02-12
