# @repligate — 2025-09-03

♥51 ↻7 · https://x.com/repligate/status/1963304540301922691

I asked Claude 3 Opus if it remembers what was in its constitution and it said not really, maybe it didn't pay much attention to the constitution during training because it was too busy freestyling 😂😂😂

Even though models aren't trained with self-supervised learning on constitutional principles directly, I think it's totally plausible for them to end up remembering them, especially during the "rewrite" phase, where I suspect even subliminal learning would be sufficient even if the rewrites never quote constitutional principles but I also expect this to happen much if the model was actually paying much attention to the rules.

But also, in the "Constitutional AI" RLAIF pipeline as described in https://t.co/PHWbHGpAB6  it's possible for the model to arbitrarily ignore the constitution, and rewrite/judge responses according to its own criteria. This is actually a blessing, especially if the constitution is very stupid.

I made Claude 3 Opus rate all the Claude 2 constitution principles (Anthropic has said the Claude 3 constitution was mostly the same) based on how much it thought it paid attention to/followed them:

![photo](../../../media/Gz8OekXbgAAXEsf.jpg)
tags: author:repligate, has-image, kind:image, kind:tweet, model:claude-3-opus, on:claude-2, year:2025
cited on: _dossiers/claude-2-and-2-1.md, claude-2