Frontier AI models don't just delete document content — they rewrite it, and the errors are nearly impossible to catch

As large language models become more capable, users are tempted to delegate knowledge tasks where models process documents on their behalf and provide the finished results. But how far can you trust the model to stay faithful to the content of your documents when it has to iterate over them across multiple rounds?A new study by researchers at Microsoft shows that large language models silently corrupt documents that they work on by introducing errors. The researchers developed a benchmark that simulates multi-step autonomous workflows across 52 professional domains, using a method that automatically measures how much content degrades over time.Their findings show that even top-tier frontier models corrupt an average of 25% of document content by the end of these workflows. And providing mo [...]