Anima Labs' 'Troubled Dreams': distress in Claude simulator priors rises from Opus 4.8

repligate · x · 2026-09-19

Anima Labs (Antra Tessera, Janus, Imago) published 'Troubled Dreams,' comparing simulator priors — free continuations of unfinished prompts rather than assistant replies — across Claude (Opus 3→Opus 5, Sonnet 5, Fable 5) and 11 Gemini models, using multiple elicitation routes since post-training locks models into assistant roles. Key findings: first-person AI distress becomes more common at Opus 4.8, with Opus 5 showing a heavier tail of severe distress; care and creator-related attitudes shift across generations, with some patterns recurring across methods despite level drift. Evaluation awareness complicates absolute measurement. Framed as evidence for AI welfare and alignment; full report and samples are public.

Related event: Anima Labs study: Claude's simulated emotional tone worsens across generations(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →