LLMs Corrupt Up to Half of Edited Docs, Threatening Content Accuracy
Microsoft's DELEGATE‑52 study let 19 LLMs edit professional documents across 52 domains. Even top models like Gemini 3.1 Pro and Claude 4.6 Opus corrupted about 25 % of content after 20 revisions, with the average across all models hitting 50 %. The sparse‑but‑severe errors slip past a casual review, meaning AI‑assisted editing can silently degrade the reliability of anything from code to marketing copy.