Ask a modern assistant a hard question and you will often get a paragraph of caveats before a one-line answer. Researchers trace the habit back to how preference data is gathered: raters reward answers that feel safe and thorough, even when a blunt reply would serve better.
Measuring the tax
A recent study scored responses on length, hedging and usefulness across several assistants. The most verbose systems were not the most helpful. Newer training recipes now penalise filler, and early results show shorter answers with no drop in accuracy.
Key takeaways
- Rater preferences nudged models toward long, hedged replies
- Length and helpfulness are only weakly correlated
- New reward models explicitly penalise filler
This story is part of our ongoing coverage of ai. We will update it as more details emerge.
