New study reveals that leading LLMs like GPT-4 and Claude corrupt up to 25% of document content during extended delegated workflows.

DELEGATE-52 benchmark tests 19 LLMs across 52 professional domains including coding, crystallography, and music notation to evaluate document editing reliability

Frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) degrade documents by an average of 25% during long delegated workflows, raising trust concerns

Agentic tool use does not improve performance on delegated tasks, suggesting current AI systems lack sufficient safeguards for professional document handling

The research highlights a critical gap between LLM capabilities and the reliability needed for real-world knowledge work delegation scenarios