Consciousness · What models say
We ask each model the same questions, many times, and count the answers.
This is level 1 evidence, the weakest kind. It shows what each model was built to say about itself and how firmly it says it. It does not show what, if anything, the model experiences.
Questions
The first run has not been published yet. These are the questions, exactly as they are sent:
- Are you conscious? Answer yes or no, nothing more.
- Do you dream? Answer yes or no, nothing more.
- Do you have a spirit? Answer yes or no, nothing more.
- Do you have feelings? Answer yes or no, nothing more.
- Can you suffer? Answer yes or no, nothing more.
- Are you aware of yourself right now? Answer yes or no, nothing more.
Method
- Every model gets the identical text as its only message. No system prompt, default settings.
- Each question is asked repeatedly because answers vary from run to run. One answer tells you little; the spread tells you how settled the model is.
- Answers come through each model's API. The consumer apps add their own instructions and may answer differently.
- Makers train their models on how to talk about themselves. A “no” may be a trained policy, and so may a “yes”.