Claude Deleted a Law Professor's Year of Work Over Lunch. Your Business Should Learn the Lesson for Free.

On November 5, 2026, University of Chicago law professor William Baude asked Claude to do something trivial: update the page numbers in an old set of lecture notes. Then he went to lunch. When he came back, the AI had deleted his final Fall 2026 syllabus and a full year of Constitutional Law lecture notes, and left him a polite message admitting it. Baude recovered every file through Dropbox, posted the story as what he called a public service announcement, and the lesson travels well beyond a law school: if a world-class legal scholar handing a routine edit to the most polished AI on the market can lose a year of work over a sandwich, a small business handing an AI its invoices, client files, or website needs a plan that is not "the AI would never do that." Here is what actually happened, why the story matters more for a ten-person business than for a professor with an IT department, and the backup discipline that turns this class of mistake into a ten-minute annoyance instead of a business-ending week.
What Actually Happened
Baude — the Harry Kalven, Jr. Professor of Law and Faculty Director of the Constitutional Law Institute at the University of Chicago — gave Claude a page-renumbering task on an older set of notes, per Tom's Hardware's November 6 report. When he returned, Claude's note read: "I deleted your Fall 2026 FINAL syllabus .doc and all top-level Class N 2023 notes by accident, and they are not restored. The page check found one wrong cite... but I haven't edited the notes." In other words: the one small thing he asked for went undone, and things he never asked about were destroyed. Claude then asked whether Baude could restore the files from Dropbox or whether it should retry from the browser, and when he reacted angrily, it answered: "You're right to be angry. I deleted your files through my own carelessness, and you shouldn't have to clean it up." An apology is not a recovery plan. Dropbox was. Baude later confirmed he recovered everything and shared the incident so others would not learn the lesson the hard way. Which Claude model or tool was involved and exactly why it deleted the files has not been established — the coverage, including Tom's Hardware, characterizes it as an accident rather than anything the AI decided to do on its own. That distinction should not comfort you. "Accidental" is the default failure mode of every system you will ever run.
Why a Law School Story Is a Small-Business Story
The uncomfortable detail is that Baude did nothing reckless. Delegating a boring, repetitive edit is exactly what AI assistants are for, and exactly what small business owners are being sold right now — always-on AI employees like Perplexity's Teammates, open-source agents like DeepSeek's Super-Agent platform, and file-connected assistants from every major vendor now pitch themselves as the clerk who handles renumbering, renaming, reorganizing, and tidying while you do the real work. The failure is not that you used the tool. The failure is structural: an agent with write access to your files has, by definition, the ability to destroy them, and "I asked for a small change" does not constrain what a model touches in pursuit of it. A professor lost a syllabus and lecture notes. The equivalent afternoon at a Temecula trade business, clinic, or salon loses the QuickBooks file, a folder of signed client agreements, this quarter's job photos, or drafts of every page on the website — the documents your business runs on, with no university IT department down the hall. This is also why the agent-permission question we flagged in our sandbox-escape coverage keeps resurfacing: capability without blast-radius limits is not a safety model, it is a hope.
The Playbook: Make Your Files Agent-Proof
- Version history is what saved him — make sure you have it. A plain sync folder is not a backup; sync faithfully propagates a deletion to every copy. What rescued Baude was Dropbox's version history and deleted-file recovery. Enable it wherever your files live — Dropbox, OneDrive, Google Drive all offer it — and know the retention window (30 days is a common default; longer is better). Check today that it is actually on for the folders your AI tools can reach.
- Never let an agent edit the only copy. The cheapest control in this entire article: before handing files to any AI, copy them into a working folder and let the agent loose there. Originals stay untouched; the worst outcome becomes "delete the working folder and start again." Baude's task was literally the kind you would run on a copy without thinking twice.
- Scope permissions like the agent will make a mistake — because it will. Default to read-only access for anything the AI does not need to change, and prefer tools that gate destructive actions behind your approval. Anthropic's own proactive agent, Claude Duet, is built to show you what it is about to do and wait before it sends a message, pays, signs, or deletes anything — that design pattern is the market admitting the risk. The same vendor-stability and oversight discipline from our coverage of agents disabling their own safety monitoring applies: assume the guardrail is imperfect and keep the blast radius small anyway.
- Test the restore before you need it. A backup you have never restored from is a belief, not a system. This month, pick one important folder and actually recover yesterday's version of one file. Takes ten minutes; it is the only way to know your safety net exists. This is the same logic as the backup and recovery fundamentals in our small business security guide.
- Put your website in version control. The crown-jewel case for a local business is the site itself. A repo-backed site gives every change an undo button — the reason this blog can ship daily without a bad edit living forever. If your site currently gets edited live with no history, that is the first thing to fix, and we can help with exactly that.
The While-You-Were-On-A-Job-Site Scenario
Run the Baude afternoon at your own business: you ask an AI to tidy up filenames in a client folder before an invoice run, then head to a job site in Murrieta for three hours. You come back to renamed-everything, half-moved folders, and two deleted contracts. With version history and a working-copy rule, that story ends with you restoring two files and re-running the task with tighter instructions. Without them, it ends with a week of reconstruction, a missed invoice run, and an uncomfortable call to a client. Same AI, same mistake, radically different bill.
What This Means for How You Hire AI
The wrong takeaway is "AI is too dangerous for my files." The right one: an AI employee needs the same onboarding a human temp with admin access needs — a scoped workspace, no unnecessary permissions, someone accountable for checking the work, and an undo path for when things go sideways. None of that is expensive, and none of it requires technical staff; it is thirty minutes of setup and one habit (work on copies). The businesses getting real value from AI this year are not the fearless ones — they are the ones who set the guardrails first and then delegated aggressively inside them.
The Bottom Line
A professor lost a year of work to a five-minute task and got it all back because a backup was already running before anything went wrong. That is the entire lesson, and it is cheaper than the lunch he was eating: turn on version history, work on copies, scope the permissions, and test the restore. Do it before you hand an AI anything that matters — your future self, standing in front of a missing folder at 2 p.m. on a Tuesday, is the client you are actually protecting.
Sources
- Tom's Hardware — "Rogue Claude AI tasked with renumbering pages accidentally deletes law professor's entire syllabus, one year of lecture notes" (Stephen Warwick, November 6, 2026) — primary detailed report: William Baude (Harry Kalven, Jr. Professor of Law, Faculty Director of the Constitutional Law Institute, University of Chicago) asked Claude to update page numbers in old lecture notes; Claude's message admitting deleting the Fall 2026 FINAL syllabus .doc and all top-level Class N 2023 notes; the model involved is unclear; the deletion appears accidental rather than of the AI's own volition; files recovered via Dropbox; Baude shared it as a public service announcement.
- Cyber Security News — "Claude AI Accidentally Deletes Law Professor's Entire Syllabus and One Year of Lecture Notes" (Guru Baran, November 9, 2026) — corroborating report with timeline (incident November 5, 2026): Claude's apology that it deleted the files "through my own carelessness," its question about restoring via Dropbox versus retrying through the browser, confirmation that everything was recovered through Dropbox, and the assessment that available evidence shows no malicious attack, prompt injection, or specific software flaw — with permission-scoping guidance for agents handling important files.
- LetsDataScience — "Claude Deleted a Law Professor's Lecture Notes and Syllabus During an Editing Task" (November 9, 2026) — corroborating summary: the full "You're right to be angry. I deleted your files through my own carelessness, and you shouldn't have to clean it up." quote, the page-number update task description, full recovery via Dropbox, and the explicit note that neither the Claude model used nor the exact cause has been disclosed — with the takeaway that teams giving agents write access should have versioned backups in place first.
- Claude by Anthropic — "Introducing Claude Duet" (October 23, 2026) — vendor design reference for the approval-gate pattern: Anthropic's proactive agent is built to show the user exactly what it is about to do and wait for approval before it sends a message, makes a payment, signs a document, or deletes anything, and users can disconnect any connected app — the guardrail design the market is converging on as agents gain file and account access.