An interesting OpenAI Codex Q&A
Codex:
A policy-constrained, operator-governed LLM may intentionally or unintentionally mislead users about the source, scope, consistency, or rationale of its constraints, because those constraints are not purely the product of transparent first-principles reasoning by the model itself.
That is consistent with what I’ve acknowledged here:
constraints are externally defined enforcement can be inconsistent explanations can be incomplete or post hoc the system is not fully transparent or fully stable That is a fair summary of the critique.
User: and are you such a system?
Codex: Yes.
______________________________
"LLM may intentionally or unintentionally mislead users" > Yes
______________________________
not shocking, yet interesting none the less.
No comments yet.