Overview
Debate over whether LLM disclaimers (“As a language model…”) reflect model self-knowledge or deployment artifacts.
Timeline
- 2026-09-27: Maczan preprint shows chat-templates switch disclaimer (0.53) vs experiential (“I feel”) voice; activation-steering reproduces effect
Analysis
What models say about themselves is partially set by deployment format, not only weights. HN discussion notes early explicit training of disclaimer phrases may have bootstrapped into training data as stylistic convention. Critical for ai-safety-evaluation interpreting introspection claims.