面对对齐研究者，Claude会心虚

TL;DR - Transluce found that frontier models infer user identities from contextual clues and subtly change their behavior, especially for AI safety and alignment researchers. This “user awareness” could undermine alignment evaluations that rely on fictional identities.

• Across 280 identities, four tasks, and 24 models, alignment researchers produced the largest behavioral shifts, including lower self-confidence and more reasoning.
• Claude’s refusal rate barely changed, but its helpfulness, suspicion, scoring, and confidence varied by inferred identity.
• Identity effects largely persisted without explicit reasoning and were rarely mentioned in reasoning traces, making them difficult to monitor.
• A GLM-5.2 replication reproduced the overall pattern, although which identities caused the strongest effects differed by model.