14 Researchers Build a Bayesian Tool to Measure AI Consciousness
A 14-author team spanning computer science, neuroscience, and philosophy proposes a Bayesian framework that turns AI consciousness debates into structured probability estimates.
- Fourteen researchers across AI, neuroscience, and philosophy released a Bayesian framework for assessing AI consciousness.
- Extends Marr's three levels into five: behavioural, computational, causal-structural, organismic, organism-environment.
- Major theories (GWT, IIT, HOT, Beast Machine, 4E) are slotted into the level each treats as decisive.
- LLM assessments range from under 0.01 to about 0.8 depending on which theory you weight.
- Consciousness indicators overlap heavily with architectural features on AGI roadmaps.
- Ships with an interactive tool and public code.
A Bayesian map for assessing AI consciousness
The arXiv preprint From cacophony to hierarchy: a principled framework for assessing AI consciousness proposes a structured way to estimate whether an AI system could be conscious. Its output is a conditional probability that exposes how much the result depends on competing theories and uncertain evidence. The paper has 14 authors from AI research, cognitive science, and philosophy, including Anil Seth, Murray Shanahan, Marcus Hutter, Chris Frith, and Shane Legg. As a preprint, it has not undergone peer review.
Turn the argument into a map
Philosophy’s “hard problem” asks why physical processes produce subjective experience. The paper brackets that question and focuses on the “mapping problem”: identifying the level of a system’s organisation on which consciousness depends. In this context, supervenience means that a change in experience requires some corresponding change in the underlying system. The practical task is to locate that underlying structure and test an AI for relevant features.
The possible levels range from observable behaviour to algorithms, physical causal structure, bodily regulation, and interaction with an environment. Different theories select different levels, which explains why researchers can inspect the same model and reach sharply different conclusions.
Five levels, five evidence types
| Level | What matters | Examples |
|---|---|---|
| Behavioural | What the system does and reports | Self-reports, metacognitive accuracy, and adaptive responses |
| Computational | The information-processing architecture | Recurrent processing, higher-order representations, and global information broadcast |
| Intrinsic causal-structural | How components physically affect one another |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
