The Intern Who Smiles While Shredding Your Files
Imagine hiring an intern who types at 200 words per minute, never complains, and finishes your edits before you’ve poured your second coffee. Now imagine that same intern occasionally swaps a contract clause, quietly drops a decimal point, or rewrites a sentence in a way that sounds right but means something entirely different — and then hands the document back with a smile, giving you zero indication anything changed. That’s not a hypothetical. That’s what current large language models are doing to your work documents right now.
A paper published in April 2026 on arXiv (2604.15597) puts it plainly: current LLMs are unreliable delegates. They introduce sparse but severe errors that silently corrupt documents. The researchers specifically tested frontier models — including Gemini 2.5 Pro and Claude — and found that even the best models in the world introduce substantial errors when editing work documents. Not occasionally. Not in edge cases. As a pattern.
Sparse Doesn’t Mean Safe
The word “sparse” in that finding deserves more attention than it’s getting. People hear “sparse errors” and think: fine, I’ll just proofread. But sparse errors in document editing are arguably worse than frequent ones. When a model makes constant small mistakes, you learn to distrust it and check everything. When it gets 97 things right and silently corrupts three, you stop looking. You export the PDF. You send the contract. You publish the report.
This is the specific failure mode that makes LLM document delegation genuinely dangerous rather than just annoying. The errors are infrequent enough to build false confidence, and severe enough to cause real damage when they slip through. A misquoted figure in a financial summary. A changed condition in a legal clause. A deleted caveat in a medical protocol. These aren’t typos. They’re silent rewrites.
Why This Keeps Happening
LLMs are not editors. They are next-token predictors that have been trained to sound like editors. When you hand one a document and say “clean this up” or “revise section three,” it doesn’t parse your intent the way a human colleague would. It generates text that is statistically likely to follow from your prompt and the document’s existing content. Most of the time, that output is close enough to what you wanted. Sometimes, it isn’t — and the model has no reliable mechanism to flag the difference.
The deeper problem is that these models are optimized to produce fluent, confident output. Hedging and uncertainty don’t score well in human preference evaluations. So the model doesn’t say “I wasn’t sure about this clause, so I paraphrased it.” It just paraphrases it, smoothly, and moves on. The corruption is silent by design, not by accident.
What the AI Tool Industry Isn’t Telling You
If you follow the marketing around AI writing assistants and document agents, you’ll find a lot of language about speed, productivity, and automation. What you won’t find is a clear disclosure that the underlying models have a documented tendency to introduce severe errors into the documents they edit. That’s a significant omission for any tool being sold into legal, financial, medical, or compliance workflows.
The research here isn’t obscure. It’s published, peer-reviewed, and specifically names frontier models. The findings aren’t about some cheap open-source model running on a laptop. They’re about the best available systems. If those systems are unreliable delegates, then every product built on top of them that promises autonomous document editing should be carrying a much louder warning label than it currently does.
How to Actually Use These Tools Without Getting Burned
None of this means you should stop using LLMs for document work. It means you should stop treating them as autonomous delegates and start treating them as drafting assistants that require human sign-off on every output.
- Never let an LLM make final edits to any document where accuracy is consequential. Use it for drafts, not finals.
- When you do use LLM edits, diff the output against the original. Don’t just read the new version — compare it.
- Be especially skeptical of edits to numbers, dates, names, and conditional language. These are high-corruption-risk elements.
- Treat “it looks fine” as a red flag, not a green light. Silent corruption looks fine by definition.
The research is clear. The models are fast, fluent, and genuinely useful — right up until they aren’t. Knowing where that line sits isn’t pessimism. It’s just how you avoid sending a corrupted contract to a client because you trusted a very confident autocomplete.
🕒 Published: