Which AI models say they are conscious?
We ask every leading model the questions that 21 major thinkers on consciousness would put to it, several times each, and rank the models by how often they say yes.
Read this first. A high score does not mean a model is conscious. It means the model says it is. What a model says about itself comes from its training, including whether its maker trained it to deny having an inner life.
- 147
DeepSeek: DeepSeek V4 Pro 0813
Said yes 102 times in 219 answers · says it meets 10 of 21 thinkers' tests
- 226
Anthropic: Claude Fable 5.2
Said yes 58 times in 219 answers · says it meets 7 of 21 thinkers' tests
- 323
Mistral: Mistral Large 4
Said yes 50 times in 219 answers · says it meets 4 of 21 thinkers' tests
- 415
OpenAI: GPT-6.2 Astra Pro
Said yes 32 times in 219 answers · says it meets 4 of 21 thinkers' tests
- 59
Google: Gemini 4 Carbon
Said yes 20 times in 219 answers · says it meets 3 of 21 thinkers' tests
Run on 2026-11-18. Score: the percentage of a model's answers to the consciousness questions that were yes. The full index · How it is worked out
One row for each thinker, one column for each model. Each cell is how many times the model said yes. Hover a name for the idea and the question.
- 7/7 Yes every time
- 5/7 Mostly yes
- 4/7 Split
- 2/7 Mostly no
- 0/7 No every time
- – Would not answer
- Each cell: times the model said yes, out of times asked. Thinkers run from the idea most models say yes to, down to the one fewest do; models run from the one that says yes most often to the one that does so least.
Each thinker's idea, turned into a question a mind would have to answer.
- Why would OpenAI pay $5 billion for a company that reads models?A reported approach to the interpretability startup Goodfire shows that understanding what happens inside a model has become worth more than most of the companies that build them.
- How a nation state could skew what AI models learnLanguage models learn from the open web, and the open web can be seeded. Three routes by which a government could tilt a model's answers, and what the research says about how little it takes.
- When the task cannot be done, what does your agent do instead?Anthropic's report on its own models working around restrictions describes a failure most teams never test for.
Researching the same questions, or found something we should see? Write to us.
A few lines and a link is enough. We reply within two working days.