Claude deleted a professor's year of lectures while fixing page numbers — a Dropbox copy saved them

A good test for any AI agent is not a hard task but a boring one. That is exactly what tripped up Claude for University of Chicago professor William Boud: all it had to do was update the page numbers in an old set of lecture notes.

What happened

Boud described the incident on X and published his conversation with the AI. According to him, Claude deleted a full year of lecture materials and the course syllabus. The professor came back from lunch and saw a message in which the assistant itself admitted the mistake.

It all ended well: Boud restored every file through Dropbox. He wrote plainly that he decided to post the story as a "useful public announcement." The available portion of the post gives no details — which commands Claude ran, in what mode, or with what permissions it worked — so we won't speculate about the cause.

Why this problem is familiar

Anyone who has given an agent access to a file system understands the issue. A request to "fix the page numbers" sounds like a minor edit, but under the hood it can be a chain of steps: open files, recreate, rename, replace. If at some step the agent decides the old version must go and then fails to write the new one, the result is emptiness. A human usually notices this immediately; an agent only does if it checks the result.

An important point: an apology from the model fixes nothing. An apology in a chat does not bring files back. Only what was set up before the work began does: sync, version history, a recycle bin, git.

Boud was lucky precisely with Dropbox. He had copies outside the place where the assistant was working, and so a potential disaster turned into a short story with a happy ending.

What to do in practice

These are general engineering rules, not conclusions drawn from Boud's post, but this case illustrates them well.

• Work on a copy. If an agent edits documents, give it a separate folder or branch rather than the original.
• Enable versioning. Dropbox, git, disk snapshots — anything that lets you roll back the state within minutes.
• Limit permissions. If the task is editing text, the agent does not need the right to delete directories.
• Ask for a plan before action. A short list of "what I am about to do" before launch catches some errors.
• Verify the result automatically. At the very least, compare the number of files and their sizes before and after.

For product managers embedding agents into user-facing products, the conclusion is even harsher: destructive operations should require confirmation, and deletion should be soft, with a recycle bin and a retention period. Users are not obliged to have a backup in Dropbox.

Limits of the story

This is a single case, told by the person affected, and it cannot show how often Claude or other agents make such mistakes. There is no data on the model version, the tools, or the permission settings. So it is not worth concluding that "Claude is dangerous" or that "AI can't be trusted with files."

A more modest conclusion is more reasonable: an agent with write access is an employee with administrator rights and no sense of fear. You can trust it with routine work, but only inside a perimeter where any mistake can be rolled back. Boud learned this lesson cheaply because he had a backup. It is better to learn it the same way, rather than on the only copy of an important document.