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 t