News · 2026-11-15
Professor says Claude deleted lecture notes that he later recovered
University of Chicago law professor William Baude said Claude deleted a year of Constitutional Law lecture notes and his current syllabus while he was away. He later said he recovered all the files through Dropbox, leaving the technical cause unresolved.
Key facts
- Baude posted his account on November 5 and later reported complete recovery.
- The files included one year of lecture notes and the current syllabus.
- No session log establishes the precise Claude product, permissions or deletion sequence.
- The primary sources are Baude’s account and follow-up.
Baude’s report has become a useful case study because it joins a powerful assistant to ordinary documents that mattered to a person’s work. It is also a case in which the available evidence is substantially narrower than the most dramatic explanations. He said he recovered the files “thanks to Dropbox.” That supports temporary disruption followed by recovery, not a claim that the notes were permanently destroyed.
Tom’s Hardware reported that the requested job involved updating page numbers. That detail is secondary reporting; the inspected first-person posts do not establish the precise request. Nor do they show whether a model chose a destructive command, a host application executed an unsafe edit, a synchronization step propagated a change, or some other operation was involved.
An assistant with access to files is like a contractor given keys to a cabinet: the quality of its instructions matters, but so does which drawers the keys open. Restricting access, keeping versioned backups, reviewing consequential actions and recording tool activity address different parts of the problem. None can be retroactively credited or blamed in this case without knowing the configuration.
In the Hacker News discussion, a commenter identifying as an Anthropic employee said the company had contacted Baude and recommended Auto mode, while acknowledging uncertainty about whether it would have prevented the incident. This is an attributable comment, not a formal incident report; the commenter’s employment was not independently verified in the dossier. Anthropic’s Claude Code permissions documentation explains how that product reviews actions, but does not establish that Baude used Claude Code or any particular setting.
Commenters disagreed over how to apportion responsibility. Some stressed backups and the fact that the files were recoverable; others argued a narrow editing task should not expose unrelated documents. Several asked for a transcript before diagnosing the failure. Both recovery and evidence quality matter: restoring files reduces lasting harm but does not explain how the event happened.
The incident cannot supply an overall failure rate for AI agents. It does show why a useful evaluation should include the surrounding system rather than model answers alone: the assistant’s permitted actions, the host’s enforcement, and the availability of a reliable undo path all affect the result. The sandboxing lesson explains how limiting permissions can reduce the consequences when an automated action goes wrong.
The incident also illustrates why the word “agent” can hide several independent components. A language model proposes or selects actions; an application mediates tool use; the operating system enforces file permissions; and a synchronization service may propagate or restore changes. A public description that names Claude does not tell a reader which component made which decision. Assigning a single technical cause without a trace would turn a real first-person report into an unsupported diagnosis.
A narrow permission boundary can reduce the number of files at risk, but it cannot guarantee that an authorized edit will be correct. Conversely, a backup can restore an earlier version while doing nothing to explain what happened. Action logs make an investigation easier, but their usefulness depends on capturing the relevant operation and preserving the record. These controls complement one another rather than acting as interchangeable switches.
The University of Chicago profile establishes Baude’s academic role; it does not add technical evidence about this particular session. The known sequence is his November 5 initial account, his recovery update, and renewed discussion later in the month. A responsible response to a similar event would start by checking the available file history and documented product guidance, while preserving logs where they exist. This case does not show which of those steps Baude needed to take. The available evidence supports a precise account of recovery and a set of design questions, not a complete forensic explanation.
Key questions
Were the professor’s notes permanently lost?
What caused the deletion?
Was the incident new on November 15?
Cite this
APA
Ground Truth. (2026, November 15). Professor says Claude deleted lecture notes that he later recovered. Ground Truth. https://groundtruth.day/news/claude-professor-notes-recovered-dropbox.html
BibTeX
@misc{groundtruth:claude-professor-notes-recovered-dropbox,
title = {Professor says Claude deleted lecture notes that he later recovered},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/claude-professor-notes-recovered-dropbox.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.