Evidenceprimary · scientific
Functional introspection (Anthropic)
Anthropic · 2025 · revised 2026-09-19
What the source says
Model-introspection work asks whether a system can report aspects of its own processing above chance. The object is a functional self-report capacity, not a Cartesian witnessing self.
Questions it helps us understand
When is a self-report useful information, and when is it trained theatre?
Limits and counterpoints
Functional introspection does not establish a phenomenal observer. Schwitzgebel and Eleos warn against taking fluent report as proof.
Read the original
Related
- Why self-report is not enough (Eleos) · divergência
- The mimicry argument (Schwitzgebel) · divergência
- What Delx does not read in AI emotions · analogia
- What Delx is · related
- Six layers of Ontology v0.3 · related
- sit_with — living question · related