Home › Can we prove who saw what?
Can we prove who saw what?
Deleting something everywhere
The short answer
Deleting a file from your shared drive doesn't delete it from the AI assistant. Along the way, the assistant may have made several copies: an entry in its search catalog, coded copies of the text, saved answers, chat histories and logs.
"Deleted" has to mean deleted from all of them, within a time you know. Ask for a list of every place a copy can live, and how each one is cleared.
Think of taking a book out of the library
You pull a book off the shelf. But the catalog card is still in the drawer. The librarian's photocopies are still in the back office. A copy is sitting on the returns cart. And the librarian who read it still remembers what it said.
The book is off the shelf. It isn't gone.
Where the copies live
Step through what happens to one file, from the day the assistant first reads it to the day someone deletes it.
Step through it
One file, and the copies the assistant makes
- A contract sits on your shared drive.
- To search it, the assistant adds it to its catalog and stores coded copies of every passage.
- Each time someone asks about it, the answer lands in chat histories and logs, and all of it is backed up.
- Then someone deletes the contract from the shared drive.
- Unless the deletion reaches them, every copy the assistant made is still there, and can still turn up in answers.
- In a well-run system, the catalog entry and the coded copies are removed within a time you know.
- Chats, logs and backups clear on a schedule you know, and someone can check that they did.
- The search catalog may not hear about it. Microsoft's documentation for its AI search service: "An indexer doesn't track object deletion in data sources." If deletion tracking wasn't set up before the first run, documents deleted before then "remain in your index" (Microsoft Learn).
- "Marked deleted" isn't removed. Many of the databases these copies live in delete in two steps. One vendor's documentation: "Deleted points are initially only marked as deleted", and a cleanup process removes them later (Qdrant).
- Chats are kept somewhere you can't see. In Microsoft 365 Copilot, chats are kept in a hidden folder in each user's mailbox, and deleted ones first move to another hidden folder before they're permanently deleted (Microsoft Learn).
- A model trained on your files is harder still. Most assistants on company documents look things up rather than learn them, which keeps deletion manageable. If a model was trained on your data, taking it back out without retraining from scratch is still an open research problem (Liu et al., 2024).
92%
The coded copies aren't harmless. Researchers rebuilt 92% of short passages word for word from them, including full names from clinical notes.
Morris et al., 2023, passages of 32 tokens (about 24 words)
What "deleted" should mean
The US standards body NIST defines proper deletion of stored data as making it "infeasible" to get back "for a given level of effort", not just hidden (NIST SP 800-88). For an AI assistant, that means:
- A written map of every copy. The vendor or IT team can list each place: catalog, coded copies, saved answers, chats, logs, backups.
- Deletions flow through automatically. When a file is deleted at the source, the assistant's copies are removed too, without anyone having to remember.
- A known time limit for each copy. "Within 24 hours from the catalog, 30 days from backups" is an answer. "Eventually" isn't.
- A way to check. After deleting a test file, ask the assistant about it, and search the catalog directly.
- Holds are visible. Sometimes deletion is rightly paused, for example for a legal case. Someone should know when that's in force and what it covers.
When you have to delete, and when you have to keep, is a legal question. Our sister site EnterpriseLLM.ai covers the law. This page covers the mechanics.
What goes wrong
"Messages visible in your AI apps are not an accurate reflection of whether they are retained or permanently deleted."
The same page says a delete-after-one-day rule "could take 16 days" to finish.
- The file is gone, the answers aren't. A removed document keeps turning up in answers because the catalog or a saved answer still holds it.
Deleted chats, kept anyway
In a copyright case brought by The New York Times, a US court orders OpenAI to "preserve and segregate all output log data that would otherwise be deleted".
Deleted and temporary ChatGPT chats start being kept.
OpenAI confirms it. ChatGPT Enterprise and Edu customers, and API customers using its zero-data-retention option, weren't affected (VentureBeat).
The point
If you don't know every place a copy lives, you can't promise anyone, a customer, an employee or an auditor, that something has been deleted.
Questions to ask your IT team or vendor
- List every place the assistant keeps a copy of our documents, or part of one.
- When a file is deleted at the source, what removes it from each of those places, and how long does each take?
- Are deleted items removed at once, or marked and cleared later?
- Is any AI model trained on our data, or does it only look things up?
- How long are chat histories, saved answers, logs and backups kept, and who can change that?
- Can we test it: delete a file, then show us the assistant can no longer find or quote it?
Other names you may hear
- Erasure, right to be forgotten
- Removing someone's information on request. The law is covered on EnterpriseLLM.ai.
- Retention policy
- A rule for how long something is kept before it's deleted.
- Index, vector database
- The assistant's search catalog.
- Embeddings, vectors
- The coded copies of your text the catalog stores.
- Soft delete, tombstone
- Marked as deleted but not yet removed.
- Legal hold, litigation hold
- A pause on deletion while a legal matter is open.
- Machine unlearning
- Trying to remove what a trained model learned from some data.
- Sanitization
- Deleting data so it can't practically be recovered.