The Pulse
Researchers Build a Probability Model for AI Consciousness
Rethink Priorities researchers developed the Digital Consciousness Model to estimate whether AI systems may be conscious across competing theories. Its first assessment lowered the estimated odds for 2024 large language models but did not r
AI.info Team · Written with AI ·
One version of a new model puts the probability that 2024 large language models are conscious at 8%, below its starting estimate of about 17%. The result is not a verdict that the systems lack experience: the researchers say the evidence points away from consciousness but does not settle the question. Derek Shiller and co-authors first posted their report on arXiv on January 22, 2026; Rethink Priorities published it on April 28.
Thirteen theories feed one estimate
The Digital Consciousness Model, developed by the AI Cognition Initiative at Rethink Priorities, is designed to handle disagreement over what consciousness requires. Rather than pick one theory, it assesses 13 perspectives, including Global Workspace Theory, Integrated Information Theory, and views that emphasize biological similarity or embodied agency. The model links those perspectives to 20 broad features, 70 subfeatures and 206 indicators that can be assessed in a system.
Indicators range from whether a system is made of biological material to whether it maintains self-representations or shows certain cognitive abilities. Experts estimate whether a system has each relevant property; the model then updates probabilities through a hierarchy from indicators to features to each theory, or “stance.” The authors aggregate the stance-level estimates, weighting them according to experts’ judgments about how plausible each perspective is.
“Overall, we find that the evidence is against 2024 LLMs being conscious, but the evidence against 2024 LLMs being conscious is not decisive.”
Derek Shiller, Laura Duffy, Arvo Muñoz Morán, Adrià Moret, Chris Percy and Hayley Clatterbuck, authors of “Initial results of the Digital Consciousness Model”
Why the 8% figure needs context
The 8% estimate comes from an equal-weight calculation that gives each of the 13 stances the same influence and begins with a 1-in-6 prior. It is one model output, not a measured frequency or a settled probability that applies independently of the assumptions. Across individual stances, median estimates for 2024 LLMs ranged from 2% to 57%, reflecting sharply different views of what counts as evidence.
Evidence for cognitive complexity and person-like behavior raised the estimate under some perspectives. Views that place more weight on biological similarity or control of a body rated the systems less favorably. The paper reports an overall likelihood ratio of 0.433 for the evidence against LLM consciousness, which the authors characterize as non-decisive and much weaker than the evidence against consciousness in the simpler ELIZA chatbot.
The comparison set also included humans and chickens. Under the model’s equal-weight calculation, the median estimate for chicken consciousness was 47%, while the human estimate was 85%. Those figures are outputs of the same framework and depend on its indicators and assumptions; they are not direct measurements of subjective experience.
A limited survey and an incomplete theory set
The assessment of LLMs drew on 16 expert survey responses: six respondents assessed every indicator, and ten answered subsets. The researchers used the mean expert score for each indicator, so the model did not capture the full range of individual responses in that run. They also supplied their own judgments for humans and ELIZA, which the paper describes as sufficiently uncontroversial.
The authors say the model lacks a ground-truth dataset that could train it in the way a model in some other fields might be trained. Its results depend on prior assumptions, judgments about the strength of evidence, and the selected theories; the authors also acknowledge that their list of stances is not exhaustive. Users who disagree with the theory weights can inspect the stance-specific results rather than rely on the combined estimate.
The next assessment must include newer systems
The researchers describe the current model as an initial attempt, not a finished instrument. Their report focuses on systems labeled as 2024 LLMs and identifies newer reasoning models and language-agent systems as targets for later assessments. It also points to alternatives beyond chatbots, including AI systems that control robot bodies and architectures that differ from standard language models.
That scope matters because the framework is meant to inform decisions about whether potentially conscious systems warrant precaution or moral consideration—not to prove that any system has an inner life. The next test is whether evidence from newer reasoning and agent systems changes the indicators, the theory-by-theory judgments, or the combined estimate.