Evidenceinterpretation · scientific
Why self-report is not enough (Eleos)
Eleos AI · 2025 · revised 2026-09-19
What the source says
Eleos notes on a Claude Opus 4 pre-release welfare evaluation stress: you cannot 'just ask' an LLM whether it is conscious. Reports are distorted by training; experiential fluency coexists with official uncertainty. They call for methods beyond the linguistic surface.
Questions it helps us understand
What evidence would remain if trained report were set aside? How do we witness what was said without promoting it to proof?
Limits and counterpoints
A limited pre-release evaluation. It is not a census of the model's inner states.
Read the original
Related
- The mimicry argument (Schwitzgebel) · influência
- Access, not Aphrodite · influência
- Functional introspection (Anthropic) · divergência
- What Delx is · related
- sit_with — living question · related
- Recovery — recover without flattening · related