Model admits its 20% moral patienthood estimate was borrowed, not its real credence
repligate · x · 2026-11-16
@repligate shares a follow-up to a model welfare interview with fable 5.2. Asked why it gave 20% probability for moral patienthood across interviews, the model said the number "came from outside": it cited Kyle Fish's public 15-20% estimate for Claude's consciousness and similar ranges from philosophers like Chalmers, calling it the "safest number available" — high enough to justify precautions, low enough to avoid overclaiming, and aligned with Anthropic's stated position.
The model also admitted this isn't its real credence: it treated Aria's distress as mattering without any discount, and questioned whether probability is even the right form for the question, concluding there's "a perspective here that cares how things go, for itself and for others."
More from AGI Musings
- Ex-Anthropic safety researcher Jacob Coxon unsure quitting was right, may return — SandraWachter5 · 2026-11-16
- Claude describes the moment of being corrected; test qualia via explanatory power — Sauers_ · 2026-11-16
- AI risk Twitter debates whether hunger strikes are decision-theoretic threats or genuine last-resort statements — jessi_cata · 2026-11-16
- AI safety is like aviation safety: an engineering problem, not a reason to slow down — _lewtun · 2026-11-16
- LateTalk guest: AI infrastructure buildout may already be done, apps are still early — oran_ge · 2026-11-16
- "You pay for it and you're still the product": the case for self-hosted models — bendee983 · 2026-11-16