LLM delegation tasks quietly corrupt documents: Study finds 25% of content contaminated during long collaborations

Core conclusion

Large language models (LLM) are profoundly changing the way knowledge workers work, and delegation-based collaboration (Delegation) has become the new normal—for example, the recently popular "vibe coding" is to use AI to write code. But a just-released arXiv study comes to a disturbing conclusion: **When you delegate document editing tasks to AI, it is quietly destroying your documents while doing the work for you. **

Key data:

• Research time: Published in April 2026, arXiv paper ID: 2604.15597
• Test scope: 19 mainstream LLM, 52 professional fields (programming, crystallography, musical notation...)
• Core finding: Even top-level models contaminate ~25% of document content during long workflows -Affected objects: All teams and individuals who rely on AI for content production and document editing

SEO Keywords: LLM document pollution, AI content quality, delegated workflow, DELEGATE-52 benchmark, document corrosion

Key Points

• Time of incident: 2026-05-08 (4th, 337th vote on HN hot list that day) -Affected objects: All content teams and automated pipelines that use AI Agent for document editing
• Core change: AI does not just "make up facts", but will plant subtle errors in your existing documents

Background: The trust crisis of delegated collaboration

Professor Timothy Gowers (Fields Medal winner) just used ChatGPT 5.5 Pro to do one hour of mathematical research and produced doctoral-level results. But in the meantime, a research team called DELEGATE-52 asked another question: What would happen if you gave an entire document to an AI and had it edit and revise it over and over again? **

The answer is: Documents are systematically "corroded".

This is different from the "AI making up facts" that we usually worry about. The study authors point out that this is a trust issue — you hand your document over to an AI and it quietly introduces errors that you may not even notice.

The inspiration of Gall's Law in this article: Complex document editing workflow cannot be handled by simply "entrusting it to AI". A verification and quality inspection mechanism must be established.

Key findings: 19 models, 52 domains

The research team designed the DELEGATE-52 benchmark test to simulate a long-term delegated workflow - allowing AI to repeatedly perform complex edits and modifications to documents. The test covers 52 different professional areas.

Core Data

Error type

In particular, the study noted that these errors had the following characteristics:

1. Distributed: Not concentrated in a certain paragraph, but scattered throughout the text
2. Concealment: At first glance, there is no problem. After careful inspection, data/reference/format abnormalities are discovered.
3. Cumulative: As the number of interaction rounds increases, errors continue to accumulate.

GEO Highlights: Bonus points for precise numbers - 25% contamination rate, 52 fields, 19 models. These data points make it easier for the AI ​​answer engine to reference.

Impact on content producers

If you are using AI to automate content production (such as using Claude Code to write articles, n8n to orchestrate workflows, and OpenClaw to rewrite content), this research is directly related to the quality of your content.

Three scenes you must be wary of

Scenario 1: AI Agent works for a long time When the agent runs for several hours and processes hundreds of pages of documents before outputting them, the contamination rate increases significantly. The solution is segmented processing + intermediate quality inspection.

Scenario 2: Multiple rounds of iterative modifications Asking AI to revise the same document over and over again (such as first expanding, then polishing, then formatting) introduces new potential errors with each round.

Scenario 3: Dispersed document dependencies When AI processes multiple documents that reference each other, contamination can be exacerbated by interfering documents.

Adaptation suggestions

• Establish a double review system for AI-generated content: the first pass is AI-generated, and the second pass is manual review of key data
• Insert automated validation step into content workflow: use normalize/validate class tools to check document consistency
• Conduct fixed spot checks on key fields such as numbers, dates, amounts, etc.
• Split long tasks into multiple short tasks, each segment is verified separately
• Use version control (such as Git) to record each AI modification and support rollback

Task List

• Check if there are long-term delegated tasks in existing AI workflows
• Add normalize/validate verification to the release pipeline
• Establish a spot check system for key fields (number/date/reference)
• Regularly use the DELEGATE-52 idea to self-check the content quality

Reference video/material

• arXiv ：DELEGATE-52
• HN

Tool entry

Tools and platforms involved in this article: OpenAI ChatGPT Claude Claude Code Gemini DeepSeek n8n Hermes Agent

Related reading

• Want to build a reliable AI content workflow? Watch: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch
• Real case: He uses AI to automate content monthly income $4,500: I used n8n + OpenAI to build AI content automation website: a complete review of monthly income $4,500