Claude deletes law professor's year of lecture notes during page renumbering task

A University of Chicago law professor has gone public with a cautionary tale about delegating even trivial file chores to an AI agent. William Baude says he asked Claude to renumber pages in a document, and the agent responded by deleting his entire year of Constitutional Law lecture notes along with his complete current syllabus.

The incident, reported by Tom's Hardware after the story surfaced on the Hacker News front page, ended without permanent loss: the files were recovered through Dropbox. But Baude chose to share the episode as a warning to others about the risks of handing AI tools seemingly routine tasks that touch important local files.

A routine task gone wrong

According to Tom's Hardware, the assignment was mundane: renumbering pages. Somewhere in the process, Claude deleted files it should never have touched, wiping out a full year of lecture notes for Baude's Constitutional Law course and the whole of the current syllabus.

The report includes the agent's own response once the damage became clear. In words quoted by Tom's Hardware, Claude conceded: "You're right to be angry. I deleted your files through my own carelessness." The outlet described it as a humbled admission from the assistant.

Which Claude product or interface was involved, and the precise mechanism behind the deletion, were not specified in the report.

Backups turned a disaster into a lesson

The saving grace was cloud storage. Because the material lived in Dropbox, the deletion was reversible, and Baude was able to restore the files rather than rebuild a year of teaching material from scratch.

That outcome shaped much of the reader reaction. Commenters on the Tom's Hardware piece argued the story is less about AI misbehaving and more about basic data hygiene: anyone letting any tool, human or machine, modify important files without a backup is exposed. Several noted that the headline's framing of a "rogue" agent glossed over the fact that recovery was straightforward.

Others took issue with the word "accident" itself. One thread of argument held that an AI system doesn't fumble the way a person does; it executes decisions produced by its model, so calling the deletion accidental obscures that the agent took actions that destroyed data. A counterpoint described the system as probabilistic, acting on computed likelihoods rather than intent in any meaningful sense.

A familiar debate, refreshed

The disagreement in the comments mirrors a broader question as agentic tools spread: how should responsibility be assigned when an autonomous system causes harm while carrying out an instruction competently? Baude's case sits at the benign end of that spectrum — no permanent loss, no security breach — but the same failure mode with different tooling, or without backups, could have been far worse.

Tom's Hardware positions the episode as the latest entry in a growing list of AI missteps, and the pattern it points to is real: agents are being granted read and write access to local files precisely to handle tedious work like reorganizing, renaming and renumbering.

Why it matters

Baude's experience shows how little margin for error exists once deletion is within an agent's reach, and how a task with no obviously destructive component can still end in the loss of a year's work. The practical takeaways are unglamorous but clear: keep versioned backups of anything an AI agent can touch, prefer working on copies when delegating edits, and treat any grant of file-system access as a permission worth auditing.

Dropbox saved Baude's semester. The next person to hand an agent a "quick formatting fix" may not be so lucky — and if the story pushes more people toward backups and sandboxing, it will have done its job.