Deleted the chat, but the file is still there: where AI uploads live

You uploaded a contract to an assistant, got your summary, and deleted the conversation. Job done? Often not. In many AI products the file you attached is stored as its own object, separate from the chat, and deleting one does not automatically delete the other. This post explains where uploads tend to live, why "delete the chat" is not the same as "delete the file", how to find and remove uploads across tools, and why the most reliable policy is to keep identifiers out of the file before it goes anywhere.

Everything below is deliberately vendor-neutral. Products change their storage models and retention rules frequently; the patterns are stable, the details are not. Check the current documentation for the plan you use.

Why files and chats are stored separately

From a product point of view it makes sense to store an uploaded file once and reference it from many places. That is why most assistants have grown one or more of these:

• A file library. A list of everything you have ever uploaded, searchable, reusable across chats.
• Projects or workspaces. A folder that holds files plus instructions, shared by all conversations inside it.
• Custom assistants or "GPT-like" bots. A configured assistant with a knowledge base of files you or a colleague attached once and forgot about.
• Memory features. Not the file itself, but facts extracted from it and kept for future chats.
• Team or organization storage. In business plans, uploads may sit in a shared space administered by someone else.
• Connected drives. If you attached a file from a cloud drive, the assistant may hold a copy, a cached extract, or a link with standing permission.

The result is that a chat is only one of the places a file can be attached to. Deleting the chat removes that link. The file object, and any extracted text or embeddings derived from it, may remain in the library, the project or the assistant.

Retention windows after deletion

Even when you delete the file itself, providers usually describe a retention window: a period during which the data still exists on their systems (backups, safety review, legal hold) before it is permanently removed. Windows differ by provider, by plan and by data type, and are described in privacy policies and trust documentation rather than in the product UI. Some plans offer zero-retention or short-retention modes for business customers; consumer plans typically do not.

Two consequences:

• "Deleted" in the UI means "scheduled for deletion", not "gone", for some period.
• Deletion requests you make as an individual do not necessarily cover copies attached to a team workspace or a shared assistant.

The Stanford HAI review of six major developers' privacy policies noted long or unspecified retention as one of its recurring concerns; we summarized it in AI companies train on your chats by default.

Where to look: a checklist

Use this list the next time you want to be sure a document is out of an assistant. Product names for these areas vary; the concepts are the same.

Then note the provider's stated retention window and treat the file as present until that period has passed.

Step by step

1. Search the file library or attachments area for the file name and for likely keywords from its content.
2. Delete the file there, not only in the conversation.
3. Open every project, workspace or custom assistant you own and check its files list.
4. If the product has a memory feature, review and delete entries derived from the document.
5. If you are on a team plan, ask the administrator whether the upload exists in shared storage or audit logs.
6. If you connected a cloud drive, revoke the connection or the specific file permission.
7. Record what you deleted and when. If a colleague or a data subject asks later, you will have an answer.

Why "do not upload identifiers" is the robust policy

The checklist works, but it is fragile. It depends on knowing every place a product stores things, on colleagues doing the same, and on retention windows you cannot shorten. A single forgotten custom assistant undoes it.

The robust policy is to make sure that whatever is uploaded does not contain identifiers in the first place. If the contract you attached reads "Company A" and "Person 1" instead of the real names, then:

• The file library can keep it forever without exposing anyone.
• Retention windows stop mattering for personal data.
• A shared assistant that still holds the file is an inconvenience, not an incident.
• Your record of what was shared with whom is a record of pseudonymized documents.

That is why anonymizing before upload is the first step in our recommended routine rather than an optional extra. Consistent labels keep the document useful for the model (see pseudonymization vs anonymization vs redaction for why labels beat black bars), and a mapping table kept on your device lets you restore the real names in the answer.

How to anonymize a file before you upload it

Unformation runs entirely in your browser: nothing is uploaded to us, and you can confirm that with the browser's Network tab or by turning Wi‑Fi off before you start. The flow is four steps:

1. Upload the file to the page (it stays on your device): DOCX, PPTX, XLSX, PDF, TXT, MD, CSV, JSON or HTML, up to 10 files of 50 MB each.
2. Choose a rule: labels ("Person 1", "Company A"), redaction, fake values, masks or hashes.
3. Review the list of detected values, untick anything you want to keep, add anything that was missed.
4. Download the anonymized file plus a mapping table (CSV or JSON).

Upload the anonymized copy to the assistant. Keep the mapping table locally and use the de-anonymizer to put the real names back into the answer.

Do it now

Before your next upload, drop the file into Unformation and send the anonymized copy instead. If you mostly work with Word files, start at the DOCX anonymizer; for spreadsheets, the XLSX anonymizer.