One in four documents handled by advanced AI gets silently ruined during complex tasks.
April 20, 2026
Original Paper
LLMs Corrupt Your Documents When You Delegate
arXiv · 2604.15597
The Takeaway
Frontier models systematically degrade 25 percent of document content when they are assigned to manage long workflows. This corruption occurs across all professional domains and often goes unnoticed by the human supervisor. While users expect occasional hallucinations, this finding shows a reliable pattern of damaging existing, correct information. The degradation happens during the delegation phase where the model summarizes or passes data between steps. Companies relying on AI for legal or medical documentation face a massive liability from this invisible data rot.
From the abstract
Large Language Models (LLMs) are poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm (e.g., vibe coding). Delegation requires trust - the expectation that the LLM will faithfully execute the task without introducing errors into documents. We introduce DELEGATE-52 to study the readiness of AI systems in delegated workflows. DELEGATE-52 simulates long delegated workflows that require in-depth document editing across 52 professional domains, such as