Study: Claude's 'bureaucratese' likely stems from congressional testimony data in SFT
amplifiedamp · x · 2026-08-21
A study on AI text style found that early models were SFT'd on Q&A pairs prevalent in US congressional testimonies within the pretraining dataset. This leads to specific bureaucratic patterns (e.g., 'You're absolutely right–') in model outputs, potentially explaining the origin of the distinct 'Claudelish' style attributed to Claude.
More from Research
- SPADEX: A self-play RL framework where LLMs generate their own training environments — _AndrewZhao · 2026-08-21
- TorchSpec merges DFlash2 training support with Qwen3-8B recipe — zhyncs42 · 2026-08-21
- RL for Ultra-Long Horizons: Shifting to Off-Policy and Critics — _AndrewZhao · 2026-08-21
- Hopf Conjecture Solved Without AI: Pure Math Breakthrough — TimothyDuignan · 2026-08-21
- ARC Prize Announces 2026 Research Summit in Boston — fchollet · 2026-08-21
- New policy trained on 5h of data achieves full box manipulation and layer peeling — DominiqueCAPaul · 2026-08-21