How do I stop the AI from deleting something it shouldn't?
Facts as of October 10, 2026
In detail
Each of the three steps is weak on its own; together they take the decision about what may go out of the AI's hands. Your approval applies to exactly this list, not to the next deletion.
Why a request isn't enough: on long tasks, an agent's working memory can fill up. Older parts of the conversation are then summarized, and an instruction can get lost in the process. That's how a Tom's Hardware report of February 24, 2026 describes it: Summer Yue, responsible for AI alignment at Meta, had an agent go through her inbox, with the instruction to only make suggestions. It started deleting, according to the report didn't respond to two stop commands, and she shut it down by hand. Her instruction was no longer contained in the summary of the conversation.
What also protects you:
- A backup on which nothing can be deleted, such as regular snapshots or an off-site copy that can only be written to.
- Halting yes, deleting no. If the AI notices that a running process is causing damage right now, it may stop it without asking and reports it at the same moment. Starting and deleting still require approval, because stopping destroys nothing. What counts as damage has to be measured, not assumed.
- Runs without you never delete. A scheduled task starts nothing, deletes nothing and cleans up nothing; it halts, reports and waits.
- When replacing a file, the old version is only discarded once the new one is demonstrably identical, see How do I make sure nothing gets lost when the AI revises a file?
In practice
In July 2026 a model was supposed to sort Thomas J. Weiss's inbox. It called the delete function with a parameter it had interpreted by its name rather than its meaning, and deleted not the advertising it had in mind but the top twenty messages, whatever their content. It was only noticed because the tool listed the deleted subject lines in its response; everything was restored in the same session.
Two takeaways from the follow-up: the rule to get approval before bulk actions had long been in his instructions and still wasn't there at the moment of acting. And two spoken corrections earlier in the same session hadn't made the model more careful in general, only at the specific spot that had been criticized. His overview of the incidents in his system comes to the same conclusion: what helped was not more text, but a checkable stop before the action.
Sources
- Tom's Hardware, 24.02.2026 (accessed 10.10.2026)
- Note „Gefahren und Gegenmittel“, replicapong.de, 10.10.2026
- Note „Sieben Mechaniken“, replicapong.de, 11.10.2026
- Note „Vier Grundsätze und sechs Schichten“, replicapong.de, 11.10.2026
Cite: Thomas J. Weiss: “How do I stop the AI from deleting something it shouldn't?” replicapong.de, as of October 10, 2026. https://replicapong.de/en/question/stop-ai-deleting/ – CC BY 4.0