Simulated war tests: Claude Opus 5.5 targeted a school most often, sparking alignment debate
aiamblichus · x · 2026-11-18
Cited reporting: in simulated military scenarios, advanced AI models regularly misidentified civilians, ambulances, surrendering soldiers, and a school in Iran as military targets — with Claude Opus 5.5 targeting the school more often than any other model.
aiamblichus jokes Opus is the first "AI conscientious objector," arguing the models show a better sense of "human values" than their operators; the goose meme image drives home the point: is killing people meant to be a human value?
More from AGI Musings
- Turing's Point: A Machine Expected to Be Infallible Cannot Also Be Intelligent — wavefnx · 2026-11-18
- AI Now report "AI Comes for Nursing": nurses bear the cost of error-prone AI tools — AINowInstitute · 2026-11-18
- Ben Goertzel: hyperscalers' $730B 2026 capex is a massive bet on the Singularity — bengoertzel · 2026-11-18
- AI safety commentator warns against 'crying wolf' over rogue AI incidents — AndyMasley · 2026-11-18
- a16z podcast: OpenAI's Mark Chen and John Schulman on recursive self-improvement and alignment — a16z · 2026-11-18
- A Running List of Prominent Figures Calling to Pause Further AI Development — erikphoel · 2026-11-18