AI Language Models Corrupt Documents After Just 20 Edits – New Study Raises Red Flags for Enterprise Automation

Breaking: LLMs Spontaneously Corrupt Work Documents in Delegated Tasks

A new preprint study from Microsoft researchers reveals that leading large language models (LLMs) introduce severe errors when editing complex documents, losing up to 25% of content after just 20 delegated interactions. The findings challenge the reliability of AI agents in enterprise workflows.

The benchmark, called DELEGATE-52, simulated editing tasks across 52 professional domains. Frontier models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4 performed worst, with document content degrading by an average of 50% across all tested models.

“Current LLMs are unreliable delegates,” the researchers state. “They introduce sparse but severe errors that silently corrupt documents, compounding over long interaction.” The paper, co-authored by Philippe Laban, Tobias Schnabel, and Jennifer Neville, is currently under review.

Expert Reactions: Warning Signs for Enterprise AI

Brian Jackson, principal research director at Info-Tech Research Group, called the findings “very interesting” but cautioned against overgeneralization. “This benchmark provides useful insights into automation limits,” he said. “But it doesn’t mean LLMs can’t be used—it means they can’t do all the work as currently constructed.”

Jackson stressed that enterprises can design safer automation. “You would use multiple agents with different roles—one makes edits, another checks for errors,” he explained. “Accuracy is crucial in the enterprise; you wouldn’t take a no-guardrails approach.”

Sanchit Vir Gogia, chief analyst at Greyhound Research, interpreted the paper as “a serious warning about delegated AI, not a claim that enterprise AI has failed.” He urged CIOs to read the preprint carefully. “Its central question—can AI preserve document integrity—is exactly the one executives should be asking.”

Background: The DELEGATE-52 Benchmark

The Microsoft researchers created DELEGATE-52 to test how well 19 LLMs handle realistic multitstep editing tasks. Each of the 310 work environments included real documents of about 15,000 tokens and five to ten complex editing requests.

Domains ranged from coding and crystallography to genealogy and music sheet notation. The goal was to simulate knowledge worker workflows where AI agents are increasingly being deployed.

The results shocked even the researchers: after 20 rounds of edits, content loss averaged 25% for the best models and 50% overall. Errors were random, severe, and compounded silently—meaning users might not notice until documents are severely corrupted.

What This Means for Enterprise AI Adoption

This study is a stark reality check for companies rushing to automate document-heavy processes with AI agents. Unsupervised delegation of editing tasks can lead to data loss, compliance risks, and eroded trust in AI systems.

However, experts emphasize that the problem lies in how AI is deployed, not in AI itself. “The paper tests raw model performance in open-ended tasks,” Jackson noted. “In a controlled enterprise flow with proper guardrails, many errors can be caught and corrected.”

Key takeaways for IT leaders:

• Never delegate unsupervised editing to a single LLM agent for critical documents.
• Implement multi-agent systems where one agent edits and another validates.
• Monitor for silent corruption by comparing original and edited documents.
• Limit delegation depth—reset after a few interactions to prevent compounding errors.

Gogia added a broader caution: “This isn’t a failure of AI—it’s a failure of imagination about how to use it safely. The same technology that corrupts a document today might excel with proper architecture tomorrow.”

The bottom line: until models improve or guardrails are standardized, delegating full control to LLMs for complex document editing is a high-risk strategy. Enterprises should proceed with caution, investing in validation layers and human oversight.