The task
A team member has a draft contract, a customer issue, or an internal plan and wants a quick summary. Before uploading it, the useful question is not just whether the assistant has a privacy setting. It is who is allowed to see the material, what the account is configured to do with it, and whether the task can be completed with less sensitive input.
OpenAI's Data Controls FAQ explains how signed-in users can turn off 'Improve the model for everyone' and says that setting applies across devices. It also explains that history can remain on while model improvement is off. Those are concrete product controls. They do not decide your team's sharing policy or prove that every document is appropriate to submit.
A proposed pre-upload check
Identify the minimum material needed for the task. For a tone edit, a redacted paragraph may be enough; a full customer file may add risk without improving the result. Check the account and workspace settings in the product itself, because interfaces and organization controls change. Record whether the output will be shared further, and remove details the recipient does not need.
If a task requires the entire document, ask the document owner which approved workspace and workflow to use. Treat links, attachments, and connected actions as separate paths through which material may travel. A pasted excerpt, an uploaded file, and a tool action are not identical sharing events. The source here supports the Data Controls description; the review procedure is an editorial proposal.
- Confirm the account or workspace control in the current interface.
- Minimize or redact input for the task.
- Review the destination and audience of the generated output.
The friction to budget for
The fastest path is to paste everything. The better routine is to make a short decision before the first upload. This is especially important when people use personal accounts for work material: a setting on one account says little about another user's account or a connected service.
Do not turn this into a legal conclusion about confidentiality or compliance. It is a practical workflow check. The result is a documented choice about data, account, and audience, with a smaller input whenever a smaller input will do the job.
If the task can be completed with a fabricated example that preserves the document's structure, use that for drafting the prompt. Replace it with real material only after the workflow and destination have been reviewed by the appropriate owner.
Sources & dates
- Data Controls FAQ ↗OpenAI Help Center · Source date not stated · Retrieved 2026-09-16
Signed-in ChatGPT users can turn off Improve the model for everyone in Data Controls. The setting applies across devices on the account. Chat history can remain available when model improvement is turned off.
Source publication dates and historical event dates describe the source material. This page has not been publicly published.

