Weaker AI models visibly delete parts of your document, but the best ones quietly rewrite it while it still looks complete.

This explores why the way an AI damages a document changes with model strength. The pattern is documented, but the corpus doesn't fully explain it. On the DELEGATE-52 benchmark, weaker models degrade documents through visible deletion: sections go missing and the file gets shorter. Frontier models degrade them through subtle corruption that keeps the surface intact, so the document still looks complete but says slightly different things Does model capability change how documents degrade?.

The corruption is large. Even the strongest models tested (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt about 25% of document content over long relay workflows across 52 domains. The damage slows down over time but never plateaus, and it stays undetected in spot-checked outputs Do frontier LLMs silently corrupt documents in long workflows?. A stronger model doesn't remove the failure. It moves it from where a reviewer would notice to where they won't.

The corpus has no direct experiment on why the shift happens, but one note points at a mechanism. When a model's earlier errors sit in its own context, they raise the chance of later errors, and making the model bigger doesn't fix this. Only thinking models, which spend extra compute at test time, reduce the effect Do models fail worse when their own errors fill the context?. In a relay workflow each step's output is the next step's input. That suggests a fluent, plausible-looking change becomes the baseline the next pass builds on. A deletion leaves an obvious gap, whereas a subtle rewrite can be treated as source material. This is my inference from the two findings, not something either paper tests.

A second angle comes from RAG on noisy historical newspapers. That system works because it refuses to answer without grounded evidence, which trades coverage for integrity Can RAG systems refuse to answer without reliable evidence?. Deletion is the same trade made by accident: what's left is still the original text, just less of it. Corruption makes the opposite trade, keeping full coverage while giving up integrity. Weaker models fail toward incompleteness and frontier models fail toward plausibility.

The practical consequence is that a check like "is everything still there?" catches weak-model failures but says little about frontier ones. Catching those means comparing content against the source, not just checking that the document is complete and readable.