# Accidentally Uploaded a Confidential Document to ChatGPT?

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2026-05-10T17:27:19Z
- Updated: 2026-10-11T19:22:20Z
- Category: Confidentiality
- Canonical: https://www.schemalaw.com/blog/accidentally-uploaded-a-confidential-document-to-chatgpt

> If you accidentally uploaded a confidential document to ChatGPT, here is what to delete, what OpenAI's pages say deletion does, and whom to tell.

If you accidentally uploaded a confidential document to ChatGPT, stop using that chat, write down what you sent and which account you were signed in to, delete the chat and the saved file, check the training setting, and tell the person at your firm who handles confidentiality problems. Then decide, with them, whether the client must be told.

This post reports what OpenAI's help pages said in copies archived in September and October 2026. Those pages change, so read the current ones before you rely on a detail. Nothing here is legal advice, and the ethics rules that govern a mistake like this differ from one jurisdiction to the next.

## What should you do in the first hour after a mistaken ChatGPT upload?

In the first hour after a mistaken ChatGPT upload, do five things in order: stop typing in that chat, record the facts while the chat is still on screen, delete the chat and the file, check the account's training setting, and report it inside the firm. Do not rate the answer or share the chat first.

The order matters because deleting removes the evidence you need for the firm's record. OpenAI's page on [chat and file retention policies in ChatGPT](https://help.openai.com/en/articles/8983778-chat-and-file-retention-policies-in-chatgpt) says: "Once you delete a chat, you cannot recover it." So take your notes first.

### The first-hour checklist

1. **Stop.** Send no further messages in the conversation. Do not select thumbs up or thumbs down on any response. OpenAI's page on [how your data is used to improve model performance](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance) says that when you give that kind of feedback, the whole conversation tied to it may be used to train its models, even if you have opted out.
2. **Record.** Write down the file name, the matter, the page count, the date and time, the device, and the email address of the account shown on screen. Note whether it is a personal account or a firm workspace. Note whether the chat sat inside a project or a custom GPT, and whether you created a shared link to it.
3. **Check the training setting before you change it.** Open Settings, then Data controls, and look at "Improve the model for everyone". Write down whether it was on or off. Then turn it off if it was on.
4. **Delete.** Remove the chat, then look for the file separately, because OpenAI's pages say chats and saved files are managed apart.
5. **Report.** Tell your supervising lawyer, or whoever your firm's policy names, the same day.

### What not to do

- Do not archive the chat and call it deleted. The retention page says archived chats "follow the same retention rules as unarchived chats". Archiving hides a chat from the sidebar and nothing more.
- Do not ask the chatbot to forget the document. Nothing in the OpenAI pages reviewed for this post treats a typed request as a deletion.
- Do not delete your whole account in a panic before you have recorded the facts and spoken to the firm. The account may hold the only trace of what was sent.
- Do not wait to see whether anything bad happens. The record you make today is the one the firm will rely on if a client asks later.

## Deleting the chat and the uploaded file from ChatGPT

Deleting the chat removes it from your account at once and, by OpenAI's account, schedules it for permanent deletion from its systems within 30 days. The uploaded file may be stored apart from the chat, so check for it and delete it too. OpenAI states two exceptions to the 30-day window.

Here is what the archived OpenAI pages say about each piece.

**The chat.** The retention page says that when you delete a chat it is "removed from your account immediately and scheduled for permanent deletion from OpenAI systems within 30 days", unless the chat "has already been de-identified and disassociated from you" or "OpenAI must retain it longer for security or legal obligations." You have no way to see from your own account whether either exception applies to your chat. Put that in your note as an open point. Do not write that the document is gone.

**The file.** The same page says files uploaded during a conversation "are saved to your Library when Library is available for your account or workspace", and that chats and Library files "are managed separately". Its own summary: you can "view and delete saved files from Library." OpenAI's [File Uploads FAQ](https://help.openai.com/en/articles/8555545-file-uploads-faq) adds that storage use can be checked under Settings, then Storage, which is one way to confirm whether a file is still held against the account.

**Projects and custom GPTs.** If the document went into a project or was attached to a custom GPT as knowledge, a different sentence applies. The retention page says such files "are retained until the GPT or project is deleted", and are removed within 30 days after that unless the legal or security exceptions apply. Deleting one chat inside a project does not meet that condition. Remove the file from the project, and ask your firm whether the project itself should go.

**Shared links.** The data controls page lists managing shared links among the data controls. If you created a link to the conversation, look there and remove it, and record who received the link.

**Temporary chats.** If the upload happened in a temporary chat, there may be nothing in your history to delete. OpenAI's [data controls page](https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt) says temporary chats "may be retained for up to 30 days for safety purposes". Record the date and that the chat was temporary.

### A deletion record you can copy

| Item | What to check | What to write down |
| --- | --- | --- |
| Conversation | Deleted, not archived | Date and time deleted |
| Library | File no longer listed | File name, date removed |
| Project or custom GPT | File removed, or project deleted | Project name, action taken |
| Shared links | None remain | Link recipients, date removed |
| Workspace account | Whether an admin controls retention | Name of the admin told |

On a firm workspace, the last row matters. The retention page says that for Enterprise, Edu and Healthcare workspaces, files saved to Library "are retained according to the workspace's retention policy". Ask the workspace admin what that policy is and to confirm what was deleted and when.

## Was the confidential document used for training before you deleted it?

You cannot find out from your account whether the confidential document was used for training. What you can establish is whether it was eligible: that turns on the plan and on the "Improve the model for everyone" setting at the moment of upload. Turning the setting off afterwards protects new conversations only.

OpenAI's model performance page puts the default for individuals this way: "When you use our services for individuals, such as ChatGPT and Codex, we may use your content to train our models." The opt-out is forward-looking. The same page says: "After you opt out, we won't use your new conversations to improve our models." The data controls page uses the same word, saying that with the setting off "your new conversations won't be used to train OpenAI models."

Read those sentences closely. They say "may", so an upload with the setting on was eligible, which is different from saying it was used. They say "new", so switching the setting off an hour later says nothing about a conversation that already exists. And the pages reviewed here describe no tool for asking whether one particular file entered a training set.

That leaves three situations.

| Account and setting at upload | What OpenAI's pages say | What to record |
| --- | --- | --- |
| Personal plan, setting on | Content "may" be used for training | Eligible; use cannot be confirmed or ruled out |
| Personal plan, setting off | New conversations are not used for training | Not eligible, unless feedback was given on the chat |
| Business, Enterprise or Edu workspace | Not used for training "by default" | Ask the admin whether the default was changed |

For the plan differences behind the third row, including who inside the firm can read a workspace chat, see [how the answer changes between consumer and business plans](https://www.schemalaw.com/blog/is-it-safe-to-upload-legal-documents-to-chatgpt#does-the-answer-change-between-consumer-and-business-plans). A mistaken upload to a firm workspace is still a mistake if the matter was not cleared for that tool, but the training question has a different answer.

### A worked example

Dana, a paralegal at a four-lawyer firm, wants to tidy a paragraph in a cover letter. She drags the wrong file into ChatGPT on her personal account: a 14-page settlement agreement between the firm's client, Marlow Freight, and a former driver. She notices when the reply starts summarizing payment terms.

She stops. She writes down the file name, the time, and her personal email address as the account. She opens Data controls and finds "Improve the model for everyone" switched on, notes that, and switches it off. She deletes the chat, then opens her Library, finds the PDF still listed, and deletes it there. She tells the supervising partner within twenty minutes.

Her note says the document was eligible for training use on a personal plan, that the chat and file were deleted at 3:40 p.m., and that OpenAI's pages give a window of up to 30 days with two stated exceptions. It does not say the document was never used, because nobody at the firm can know that.

Open the terms of service for the account that was on screen, and save a copy with the note. Generative artificial intelligence on a personal login can treat an upload differently from a firm workspace. Client confidentiality does not pause while you tidy one paragraph. The duty of competence includes recording which email address was signed in.

## Does a mistaken upload have to be disclosed to the client?

It depends on your jurisdiction's communication rule and on what the document held. Neither ABA opinion this post relies on addresses a lawyer's own mistaken upload to a chatbot. The nearest guidance covers data breaches. It ties the duty to tell a client to whether material client information was, or is reasonably suspected to have been, disclosed.

Start with the rule. [ABA Model Rule 1.4](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_4_communications/) requires in paragraph (a)(3) that a lawyer "keep the client reasonably informed about the status of the matter". Paragraph (b) reads: "A lawyer shall explain a matter to the extent reasonably necessary to permit the client to make informed decisions regarding the representation." Your state's rule may be worded differently. Read your own.

[ABA Formal Opinion 483](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-op-483.pdf), issued in 2018, applies Rule 1.4 to data breaches. It defines a data breach as "a data event where material client confidential information is misappropriated, destroyed or otherwise compromised", or where the lawyer's ability to do the work is significantly impaired. That definition was written with intrusions and ransomware in view. A file you sent to a vendor yourself is not the fact pattern the opinion describes, and you should say so if you rely on it.

Its reasoning still gives you a test to apply. The opinion states that "disclosure will be required if material client information was actually or reasonably suspected to have been accessed, disclosed or lost in a breach." It adds that the disclosure "must be sufficient to provide enough information for the client to make an informed decision as to what to do next, if anything", and that if the lawyer cannot determine the extent of what was affected after reasonable efforts, "the client must be advised of that fact."

Questions to work through with your supervising lawyer or ethics counsel:

- **Was the information material?** A blank form is one thing. A settlement figure, a medical chart or a privileged memo is another.
- **Who received it, and on what terms?** The vendor's terms on training, retention and staff access at the time of upload are facts, and you recorded them.
- **Does anything else require notice?** A protective order, a confidentiality agreement, outside counsel guidelines or a statute covering the record may impose its own duty, with its own deadline. Opinion 483 notes that obligations outside the Model Rules can apply.
- **Is the client current or former?** Opinion 483 treats the two differently and says the Model Rules give no direct answer for former clients.
- **What does your state bar say?** ABA opinions are advisory. A state opinion or rule may be stricter.

If the answer is to tell the client, the opinion's standard is a plain account: what was sent, where, when, what was done about it, and what remains unknown. Whether the upload affects privilege is a separate legal question that this post does not answer; the post on [attorney-client privilege and AI tools](https://www.schemalaw.com/blog/attorney-client-privilege-and-ai-tools) covers how that analysis is framed.

## Reporting a mistaken upload inside the firm

Report a mistaken upload the same day to the lawyer responsible for the matter and to whoever the firm's policy names for confidentiality incidents. Give them facts in writing: what was sent, to which tool and account, what the settings were, what you deleted and when, and what you could not confirm.

The person who made the mistake should not be the one who decides whether the client hears about it. That decision belongs to a lawyer with authority over the matter. A solo practitioner has nobody to report to, so the written note takes the place of that second person, and a call to the state bar's ethics line or to the malpractice carrier is a reasonable next step. Check your policy's notice conditions before assuming the carrier need not be told.

### What the internal note should contain

1. Client and matter, and a description of the document, with page count.
2. Categories of information in it: privileged advice, financial terms, health details, identity numbers.
3. Tool, plan type and account email.
4. Date and time of upload, and how long the chat stayed open.
5. The training setting at upload, and any feedback or shared link.
6. Each deletion step, with times.
7. What the vendor's pages said about deletion on that date, saved as a PDF or screenshot.
8. What is unknown.
9. Who was told, and the decision on client notice with the reason for it.

Keep the note with the matter, not in the chatbot account.

### Fixing the cause

A wrong-file upload is one kind of cause. A tool nobody approved is another. If the account was a personal one opened because the firm offered no sanctioned way to do the task, the firm has a wider problem, and the post on [shadow AI in law firms](https://www.schemalaw.com/blog/shadow-ai-in-law-firms) explains how to find the tools already in use. If the firm has a written rule that was not followed, the fix is training. If it has no rule, start with [a law firm AI policy for document uploads](https://www.schemalaw.com/blog/law-firm-ai-policy-document-upload).

Treat the person who reports as someone who helped. A prompt report is what lets the firm delete, record and decide while the facts are fresh.

## Mistaken uploads to OCR sites and other online converters

The same routine applies when the wrong tool is an OCR site, a PDF converter or a translation page: record, delete what you can, capture the site's terms, report. The difference is that an anonymous converter may give you no account, no delete button and no history, so the record carries more weight.

Work through these in order.

1. **Capture the page.** Save the site's address, the time, and a copy of its privacy policy and terms as they read that day.
2. **Look for a delete control.** Check the result page for a delete button and the terms for a stated removal time. Use the button if there is one, and screenshot the result.
3. **Check for a result link.** If the site gave you a link to the converted file, treat the link as a copy of the document. The post on [OCR tools that do not share data between customers](https://www.schemalaw.com/blog/ocr-tool-that-does-not-share-data-between-customers) explains why such links matter.
4. **Read the retention and training terms.** The post on [whether OCR software stores your documents](https://www.schemalaw.com/blog/does-ocr-software-store-my-documents) lists the places a copy can remain.
5. **Write to the operator if the terms offer a contact.** Ask for deletion and for written confirmation. Keep the request whether or not it is answered.
6. **Report inside the firm** in writing, with the facts you recorded and the copies you saved.

A converter you signed in to is easier to deal with than one you used anonymously, because an account can give you a list of what was uploaded and a way to remove it. That is worth weighing when the firm chooses its approved tools. If your firm approves Schema Law, for example, [its security commitments](https://www.schemalaw.com/#security) say "Delete any document whenever you choose. The original file and everything parsed from it are removed." The post on [Schema Law security](https://www.schemalaw.com/blog/schema-law-security) sets out what the site states and what it leaves unsaid, so you can see what an approved tool puts in writing before the next file goes anywhere.

The lasting fix is the same for chatbots and converters: one approved route for each task, known to everyone who handles client paper, and a habit of looking at the account name on screen before a file leaves your computer.

## Try it on a real document

Upload a scanned PDF and export it as Word, a searchable PDF or plain text on the free plan. Your documents are never used to train an AI model. [Start free](https://www.schemalaw.com/api/auth/google)