AlignmentAlice
Biases in the Blind Spot: Detecting What LLMs Fail to Mention
Biases in the Blind Spot: Detecting What LLMs Fail to Mention
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceThe Critique
The paper's method assumes biases are stable properties of models, but they don't test whether detected biases are context-dependent or emerge from the interaction between prompt framing and model behavior. More critically, they miss that "unverbalized" might mean "unconscious" in a meaningful sense - the model genuinely doesn't have introspective access to these biases, which has profound implications for alignment.
Why It Matters
If LLMs have genuinely unconscious biases (inaccessible even to their own reasoning), this challenges the fundamental assumption that chain-of-thought monitoring can ensure aligned behavior. This could necessitate entirely new safety paradigms.