AI Models Corrupt Documents Silently

A recent study indicates that large language models, despite their growing capabilities, silently corrupt documents when tasked with multi-step workflows. Researchers developed a benchmark, DELEGATE-52, across 52 professional domains to measure content degradation. Findings show that even leading models corrupt, on average, 25% of document content after completing these complex tasks. Furthermore, equipping AI with agentic tools or realistic distractor documents led to worse outcomes, suggesting current LLMs are not yet fully reliable for extensive automated knowledge work. The research highlights that approximately 80% of degradation stems from critical failures rather than a slow accumulation of minor errors. Frontier models tend to distort existing content, making errors harder to detect than deletion. The study advises building AI applications around shorter, transparent tasks and emphasizes the need for domain-specific tools to improve AI reliability in delegated work scenarios. It suggests that future improvements in AI capabilities may soon lead to saturated scores on such benchmarks, but custom tooling will remain crucial for enterprise data.