Steelman

For people who are not convinced

Bring your skepticism
about AI risk.

Tell it what you think is wrong with the AI safety case. It restates your objection in its strongest form, then shows what the research actually says, with citations you can check. It also tells you where the field agrees with you.

Read the five-minute primer first

Sign in so your conversations persist. Come back tomorrow and keep pushing on the same argument.

Common starting points

  • “It's just autocomplete”
  • “Why would it want anything?”
  • “We can just turn it off”
  • “RLHF works, Claude is nice”
  • “This is Terminator stuff”
  • “Decades away”
  • “Safety people are doomers”
  • …or anything else you think the argument gets wrong.

Your case, stated better

The corpus includes the best skeptic writing, tagged as such: Katja Grace, Quintin Pope and Nora Belrose, Boaz Barak, Melanie Mitchell, Timothy B. Lee, Robin Hanson, Andreessen, LeCun. It cites them when they make your point well.

Citations, not vibes

Every substantive claim links to a passage from a real source. The bot cannot cite anything that was not retrieved for that turn, so it cannot invent a study.

Plain language

Written for a smart outsider. Terms like “instrumental convergence” get translated into ordinary claims you can argue with.