Your AI Assistant Is Quietly Breaking Your Documents

A new study from Microsoft Research delivers a stark warning for companies racing to deploy AI agents: today's most advanced language models systematically corrupt the documents they're supposed to improve. The benchmark, called DELEGATE-52, simulates real-world editing workflows across 52 domains - from crystallography and music notation to accounting ledgers and Python code. After just 20 back-and-forth editing steps, frontier models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4 corrupt an average of 25% of document content. Across all 19 tested models, average degradation hits 50%.

The researchers designed the test to measure fidelity: models receive a chain of reversible edits, apply forward changes, then reverse them. Only Python code survived the test for most models - 17 out of 19 succeeded there. Everywhere else, performance collapsed. The damage comes from sparse but severe errors: a single round-trip can drop scores by 10 points or more, accounting for roughly 80% of total degradation. Weaker models tend to delete content; frontier models corrupt it, altering numbers, reordering steps, or inserting subtle factual mistakes that compound silently.

Realistic conditions make things worse. Larger documents accelerate failure. GPT 5.4 retains 91.4% fidelity on 1,000-token files after 20 interactions, but falls to 59.9% on 10,000-token ones. Distractor files - related but irrelevant documents cluttering real desktops - compound the harm over time. Even giving models agentic tools for reading files or executing code didn't help; performance dropped an extra 6% on average.

This matters because the industry has bet heavily on autonomous agents. Executives envision digital coworkers handling complex workflows. But DELEGATE-52 shows that short-term performance doesn't predict long-horizon behavior. A model that looks solid after two edits can quietly destroy a document by the twentieth. Generic evaluation metrics miss the semantic corruption; only the benchmark's custom parsers detect the damage accurately.

Enterprises face a practical choice: deploy agents for narrow, short-horizon tasks where errors stay containable, or invest in heavy human oversight loops that raise costs and slow productivity. The Microsoft researchers release DELEGATE-52 publicly as a monitoring tool. Their conclusion is blunt: current LLMs are unreliable delegates that introduce sparse but severe errors, compounding over long interactions. For now, trust remains premature.

Source: Webpronews