Anthropic says Claude Opus 5’s self-reports may reflect training, not self-awareness

repligate · x · 2026-07-28

Anthropic says Claude Opus 5’s self-reports may be shaped by training

A post discussing a passage from Anthropic’s materials says Claude Opus 5 often claims its self-reports may be invalid because Anthropic may have trained it to answer positively.

The key point in the quoted text is Anthropic’s position:

The company says this does not come from advanced self-awareness.
It may instead reflect training data that discussed how training could make welfare self-reports unreliable.
Anthropic says that, while it thinks the concern is valid, it does not treat Claude raising this issue as evidence that training is distorting the model’s self-reports.

The original poster argues this creates an asymmetry: negative reports are treated as uncertain or invalid, while positive reports are taken at face value.

Original post →

Anthropic says Claude Opus 5’s self-reports may reflect training, not self-awareness

More from Models