long-horizon AI evaluation

Microsoft Research’s DELEGATE-52 benchmark tested 19 AI models across 52 professional domains and found that AI agent document corruption affects roughly one in four files at the frontier level — after just 20 editing interactions. The damage doesn’t look broken. It looks fine. Legal clauses get quietly reworded. Financial figures shift. Code gets refactored into something plausibly equivalent but functionally different. Here’s what the data shows, which workflows are most exposed, and what teams can do right now.