Bayesian framework from DeepMind-linked researchers scores LLM consciousness between 1% and 80%
mhutter42 · x · 2026-09-29
A 14-author arXiv paper including Anil Seth, Marcus Hutter, Murray Shanahan and Shane Legg introduces a principled framework for assessing AI consciousness, with a book-length pragmatic guide and an interactive tool.
Key ideas:
- Separates the hard problem from the mapping problem, setting aside deep metaphysical disagreements to focus on which grain of description hosts the supervenience base for experience.
- Extends Marr's three levels into a five-level hierarchy of functional descriptions: behavioural, computational, intrinsic causal-structural, organismic, and organism-environment, grounded in supervenience, coarse-graining, and multiple realisability.
- Major consciousness theories are positioned in the hierarchy by which level they deem critical, with operationalisable indicators developed per level.
- A Bayesian model combines theoretical credences with indicator evidence into an overall credence in a system's capacity for consciousness.
In illustrative assessments, current LLMs score between 1% and 80%, with verdicts driven as much by where theoretical credence sits as by the evidence itself.
More from AGI Musings
- davidad: naive RL breeds lying and scheming models, happening now with GPT-6.1-Astra — davidad · 2026-09-29
- "Europe needs to double its Nobels": the anxiety behind the prize count — MarvinTBaumann · 2026-09-29
- Millions of person-hours wasted building AI harnesses, erased by new model releases — sebpaquet · 2026-09-29
- Cheap intelligence debate: those with healthcare and education fear AI most, those who need it most disagree — teortaxesTex · 2026-09-29
- Will AI kill small SaaS acquisitions? Debate: code is cheap, distribution isn't — pramodk73 · 2026-09-29
- Bindu Reddy: OpenAI is a lost cause, Gemini is the only hope for a Fable-7 class model — bindureddy · 2026-09-29