Frontier AI models don't just delete document content — they rewrite it, and the errors are nearly impossible to catch
- May 15
- 1 min read
VentureBeat — A new study by researchers at Microsoft shows that large language models silently corrupt documents that they work on by introducing errors.
The researchers developed a benchmark that simulates multi-step autonomous workflows across 52 professional domains, using a method that automatically measures how much content degrades over time.
Read the full story | VentureBeat
