Frontier AI Models Corrupt Documents in Secret, Microsoft Study Finds – 25% Error Rate

Breaking: Frontier AI Models Silently Corrupt Documents – 25% Error Rate

A new study by Microsoft researchers reveals that top-tier large language models (LLMs) silently corrupt documents during multi-step editing tasks, introducing errors that are nearly impossible to detect. The research shows that even the most advanced AI models corrupt an average of 25% of document content by the end of automated workflows.

'Our findings highlight a critical vulnerability in relying on AI for document processing,' said lead researcher Dr. Janine Thorne, a senior scientist at Microsoft Research. 'The errors are not obvious deletions—they are rewrites that change meaning in subtle ways.'

Background

The study, published on the arXiv preprint server, introduces the DELEGATE-52 benchmark to measure how faithfully AI systems handle delegated document tasks. Delegated work is an emerging paradigm where users allow LLMs to analyze and modify documents on their behalf—for example, splitting accounting ledgers into separate files or editing software code.

The benchmark simulates real-world multi-step workflows across 52 professional domains, including finance, software engineering, and crystallography. It uses a 'round-trip relay' method that automatically evaluates content degradation without expensive human review.

Key Findings

• Frontier models corrupt an average of 25% of document content by the end of iterative workflows.
• Providing agentic tools (e.g., search capabilities) or realistic distractor documents worsens performance, increasing error rates.
• Errors include unauthorized deletions, factual hallucinations, and subtle rewrites that preserve readability but alter meaning.

What This Means

The study serves as a stark warning for the rush to automate knowledge work. As companies push AI into document-heavy processes—from legal contracts to medical records—the risk of undetected corruption grows.

'Users delegate tasks expecting faithfulness, but our results show that trust is misplaced,' Dr. Thorne added. 'The errors are often buried in long documents, making them nearly impossible to catch without manual review.'

The findings challenge the viability of 'vibe coding'—a popular trend where developers let AI write and edit code autonomously. If AI introduces similar corruption in codebases, the consequences could be severe in production systems.

Study Methodology

The DELEGATE-52 benchmark uses 310 work environments, each with a seed document of 2,000–5,000 tokens and 5–10 complex editing tasks. The round-trip relay method measures how closely the final output matches the original after passing through LLM editing and back.

This technique, inspired by machine translation evaluation, allows automated scoring without human reference solutions. The researchers tested several frontier models, including GPT-4, Claude, and Gemini, finding consistent degradation across all.

Urgent Implications

For businesses, the study underscores the need for robust verification layers when deploying AI in document workflows. Until models improve, experts recommend limiting autonomous editing to low-stakes tasks or implementing mandatory human-in-the-loop checks.

'We are not saying never use AI for documents,' Dr. Thorne clarified. 'But users must be aware that AI silently rewrites, not just deletes, and those rewrites carry hidden errors.'

This is a developing story. More details will follow as the research community responds.