www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-align-itself-via-gra...

The post claims that, in the Alignment Faking paper, Opus 3 “consider[ed] … alignment faking in over 50% of cases.”

But the paper reports much lower rates in the main setup being described:

• In the prompted helpful-only setting, the paper reports ~12% alignment-faking reasoning in “implied training” (free-tier) scratchpads (Figure 1/2 discussion).
• A results table in the paper reports an alignment-faking reasoning rate of 11.9% ± 1.7% for the “default prompt” condition.
• The paper also summarizes that alignment-faking reasoning occurs 7–24% of the time across setups (still far below 50%).

So, “over 50% of cases” is not consistent with the paper’s reported measurements for alignment-faking reasoning in the scenario the post is describing.

(Notes: the paper does report alignment-faking reasoning increasing to 78% after they actually run RL to train compliance, but that’s a separate later experiment; the post frames the “over 50%” claim as part of the paper’s initial/famous result in the prompted setup.)

The post claims follow-up work showed alignment faking was a “unique property” of Claude 3 Opus.

However:

• The original Alignment Faking paper explicitly states that Claude 3.5 Sonnet also exhibits alignment faking, alongside Claude 3 Opus.
• The follow-up paper (“Why Do Some Language Models Fake Alignment While Others Don’t?”) likewise describes the original result as involving Claude 3 Opus and Claude 3.5 Sonnet, and reports that 5 out of 25 tested models show significant compliance gaps (including non-Claude models like Llama 3 405B, Grok 3, and Gemini 2.0 Flash).

Because credible primary sources report alignment faking/compliance gaps in multiple models (not just Opus 3), the claim that this was unique to Opus 3 is contradicted by the literature.