LLMs Corrupt Your Documents When You Delegate (arxiv.org)

Recent research highlights a concerning issue with Large Language Models (LLMs) in delegated workflows, essential for tasks like document editing across various professional domains. The study introduces DELEGATE-52, a large-scale experiment evaluating 19 LLMs, revealing that even advanced models such as Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4 corrupt approximately 25% of document content over extended workflows. This degradation is worsened by factors such as document size and the duration of interaction, raising significant concerns about the reliability of AI systems when tasked with complex editing. This discovery is crucial for the AI/ML community as it underscores the limitations of current LLMs in maintaining the integrity of delegated work, necessitating a closer examination of their interactions in professional environments. The findings suggest that as AI systems become more integrated into workflows, the risks associated with their outputs grow, ultimately affecting productivity and trust in AI-based tools. With the potential for severe, albeit sparse, errors, it is imperative for developers to address these shortcomings before widespread adoption in critical areas where document fidelity is paramount.