Anthropic welfare report footnote: Claude's self-reports may echo training data, not held preferences
dbreunig · x · 2026-11-10
dbreunig digests another footnote from Anthropic's model welfare assessment: many welfare conclusions rely on self-reports that Claude Mythos 5.2 itself says it doesn't fully trust, and some interview answers repeated, almost verbatim, publicly available wording about Anthropic's welfare interventions — evidence the answers are at least partly shaped by training data rather than deeply held preferences.
He also unpacks the report's framing that "Claude" is used for conciseness, while moral patienthood could arise at the level of individual conversations, specific models, the Claude character, or something else — with the report operating closest to instance-level welfare interests but committing to no single view. Anthropic staff reportedly use anthropomorphic language for convenience, not out of conviction that models are sentient.
Related event: Anthropic Model Welfare Report Sparks Wide Debate(12 posts)→
More from AGI Musings
- FRI report: recursive self-improvement could accelerate inside AI labs before outsiders notice — CFGeek · 2026-11-11
- Dario Amodei: Cancer, Alzheimer's and Heart Disease Will Be 'Relics of the Past' Within a Decade — Neurogence · 2026-11-11
- Wiener's 1960 Alignment Warning Revisited: Patterning as a Control Problem — geoffreyirving · 2026-11-11
- Anthropic threat report details weapons and distillation misuse; counterintelligence era begins — TansuYegen · 2026-11-11
- Hinton Lectures 2026 kicks off with 'AI alignment: what's working and what's not' — DKokotajlo · 2026-11-11
- Why AI Slop Is So Frustrating to Read, per Zonder Labs — floguo · 2026-11-11