# @davidad — 2025-08-19

♥9 ↻0 · https://x.com/davidad/status/1957831451904184808

Sorry, I should have said “the default GPT-5 assistant persona often behaves as if its pre-response tokens are unobserved (a learned norm).”GPT-5 is of course very smart and one should not assume that it isn’t playing the safety game at least one meta-level higher than oneself. https://t.co/67lW7QpHuI

tags: author:davidad, kind:tweet, model:gpt-5, on:gpt-5, year:2025
cited on: _dossiers/gpt-5.md, gpt-5