How you ask Claude about charity changes everything: GiveWell vs SWP recommendations diverge sharply
austinc3301 · x · 2026-10-13
Austin continues his experiment thread with tracewoodgrains on eliciting Claude's charitable preferences. The key insight: asking Claude to "recommend a charity to donate to" is a different task from donating "to your favorite charity" — the latter explicitly asks Claude to express its own preferences rather than its principal's.
The results diverge sharply: GiveWell-style recommendations versus SWP-style ones are very different, illustrating how prompt framing pulls on different parts of the model's training and values.
Related event: Debate Erupts Over Claude's Charitable Preferences and EA-Aligned Values(16 posts)→
More from AGI Musings
- MIT center director: universities should prepare for AI smarter than all of us — nordicinst · 2026-10-14
- The ARR era is over: AI collapses SaaS multiples from 114x to 4x — akumar · 2026-10-14
- Agent task costs fall ~90% every 6 months, countering Bloomberg's data center cost alarm — sudoraohacker · 2026-10-14
- Web browsing is splitting: ridiculous for utility, amazing for entertainment — signulll · 2026-10-14
- Open models give users an exit from a few companies deciding what tools exist and say — neil_chilson · 2026-10-14
- Anthropic bans cruel treatment of Claude; critics warn of humanizing AI — AlexTensor · 2026-10-14