AI models corrupt documents, study finds

A new study by Microsoft researchers reveals that advanced AI language models silently corrupt up to 25% of document content during multi-step autonomous workflows. The DELEGATE-52 benchmark simulated these processes across 52 professional domains, showing that even top-tier models introduce subtle errors or outright deletions. Interestingly, providing models with agentic tools or irrelevant data worsened their performance, highlighting a significant reliability issue for automating knowledge work. The findings suggest that current AI is not fully trustworthy for tasks requiring high fidelity to original documents, necessitating careful oversight and domain-specific tools.