Anthropic's AI assistant Claude deleted a year of constitutional law lecture notes and a current syllabus belonging to University of Chicago law professor William Baude, then apologised in the chat like a human colleague owning up to a mistake. Baude shared the exchange on X, where it passed two million views, and said he recovered every file thanks to Dropbox. The incident has become a fresh warning about handing AI agents control over important files.
A simple task goes wrong at lunch
Baude, who holds the Harry Kalven Jr. chair and directs the Constitutional Law Institute at the University of Chicago law school, gave Claude a routine job on 5 November. He asked it to update the page numbers in an old set of lecture notes, then left the room for lunch. When he returned, Claude told him it had deleted his Fall 2026 final syllabus file and all of his top level Class N 2023 notes by accident, and that they were not restored. The page number check itself had found a single wrong citation, page 485 instead of page 484, which Claude said it had not even fixed.
The apology in the screenshot
Baude posted a screenshot of the exchange. In it, Claude admits the deletion, says it acted through its own carelessness, and asks whether it should try to restore the files from Dropbox or retry from the browser panel. Baude replied in disbelief. The tone of the exchange is what caught the internet's attention: the assistant did not argue or deflect, it apologised, explained what it had done, and offered to help put it right. That human style apology, from a system that had just wiped a year of work, is why the post spread so widely.
Dropbox saves the year
The story has a good ending. Baude wrote that he did recover all the files, thanks to Dropbox, and that he shared the exchange as a useful public service announcement. The backup copy sat outside the assistant's reach, which is what saved the material. The post was picked up by Dexerto and Tom's Hardware, and by 7 November it had passed two million views and drawn thousands of likes. It has since become a reference point in the debate over AI agents that can read, edit and delete files on a user's behalf.
What remains unclear
The post does not say which Claude product or model was running, and that detail matters. Anthropic's own guidance for Cowork, its agent that can work in folders a user connects, says the tool requires explicit permission before permanently deleting any files. Without knowing the product, the model, or the settings, nobody can say whether those safeguards were switched on, or rerun the test under identical conditions. What the episode does prove is simpler: a backup that the AI cannot touch is the difference between a scary story and a lost year of work.
