← Retour au blog
tech 9 May 2026

LLMs Corrupt Your Documents When You Delegate

Large Language Models (LLMs) promise to revolutionize knowledge work. However, a recent study shows they often corrupt documents during delegation.

Article inspired by the original source
LLMs Corrupt Your Documents When You Delegate ↗ arxiv.org

Introduction

Large Language Models (LLMs) have stormed the tech world, promising to revolutionize how we handle knowledge work. With the emergence of interaction paradigms like 'vibe coding', delegating tasks to these models has become commonplace. But can we really trust them? A recent study, 'LLMs Corrupt Your Documents When You Delegate', highlights a major issue: these models can corrupt documents when used for delegation tasks.

The DELEGATE-52 Experiment

To assess the effectiveness of LLMs as delegates, researchers developed DELEGATE-52, a large-scale experiment simulating delegation workflows across 52 professional domains, from coding to music notation. The results are concerning: even frontier models like Gemini 3.1 Pro and GPT 5.4 corrupted an average of 25% of document content during long workflows.

Details of the Experiment

The DELEGATE-52 experiment involved 19 different models. Researchers found that, despite technological advancements, no current model succeeded in maintaining document integrity over time. The errors introduced are often subtle but can accumulate and become severe.

Degradation Factors

Several factors were identified as increasing the likelihood of document corruption:

  • Document Size: Larger documents are more prone to corruption.
  • Interaction Length: Longer interactions increase the risk of error introduction.
  • Presence of Distractor Files: Irrelevant files can also exacerbate degradation.

Limitations of Current Tools

Research shows that agentic tool use does not compensate for the degradation observed. This raises questions about the reliability of LLMs for critical delegation tasks.

Implications for Businesses

For businesses considering using LLMs to automate document processes, these findings are an important reminder of the need to carefully monitor document integrity. Automation should not come at the expense of accuracy and reliability.

Conclusion

The promise of LLMs is great, but their application in delegated workflows must be carefully evaluated and monitored. Decision-makers need to be aware of potential risks and implement robust verification systems to minimize errors.

Let's discuss your project in 15 minutes.

LLM document integrity automation delegation AI reliability
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call