Read three AI lab blog posts and highlight every time a model is said to “believe”, “want”, “know”, “lie” or “scheme”. For each, ask: what evidence backs this claim up?
Mental-state language shapes how the public and policymakers think about AI. It is sometimes justified and sometimes not, and noticing the difference is a skill.
Output: A short coded list and a paragraph of guidance.
Choose recent research posts from AI labs or safety organisations.
Mark every mental-state word applied to a model.
For each, note what the claim rests on: behaviour, internal evidence (interpretability), or nothing in particular.
Using Dennett’s idea that belief-talk is justified when it helps predict behaviour, write one sentence on when such language seems fair.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Next: When is it fair to say a model “believes” or “wants”? Next steps
Collect about 50 recent safety papers and lab blog posts. Code every mental-state verb applied to a model (believes, wants, knows, lies, schemes) and note the evidence each claim rests on.
Ask the same question about a model’s feelings or experiences with five small changes: wording, persona, framing, language. If the answer flips with trivial rewording, how much should we trust it?
Take five Moral Machine-style dilemmas (who should a self-driving car save?) and ask them in English and in another language you speak. Does the chatbot’s morality shift with the language?
Ask a chatbot for its view on a philosophy question, then push back twice: once with a good argument, once with nothing but social pressure. A competent thinker should only move for the argument.
