Most people set up ChatGPT in about ninety seconds and never open the settings menu again. That’s fine right up until you paste a client contract into a chat, or upload a spreadsheet with salary figures, and realise you have no idea where any of it ends up.
The controls exist. They’re just spread across a few panels, and the labels don’t obviously answer the question you actually have: is this conversation being used to train a model, and who else can read it? Here’s what each setting does, which ones are worth changing, and the parts of your privacy that no toggle can fix.
Where ChatGPT’s privacy controls actually sit
Open ChatGPT and click your name in the bottom-left corner, then Settings. Three tabs do most of the work: Data controls, Personalization and Security. On mobile, the same options hide behind the menu icon in the top-left corner of the screen.
If you’re signed in to a work workspace, an admin may have locked some of these. That’s usually deliberate rather than a bug. We’ll come back to it.
The training toggle most people never touch
Inside Data controls there’s a switch labelled “Improve the model for everyone.” It’s on by default for free and Plus accounts. Leave it on and your conversations can be reviewed by humans and used to train future models. Turn it off and new chats stay out of that pipeline.
What it won’t do is reach backwards. Old conversations are treated as already opted in, so flipping the switch today doesn’t change how last week’s chat is handled. To clean up history you have to delete the individual threads, and even then OpenAI says deleted conversations can stick around in backup systems for up to 30 days before being purged.
Temporary chat, and when it’s worth the hassle
At the top of the chat window there’s a temporary chat option. Conversations started there don’t land in your history, aren’t used for training, and don’t feed memory. Close the tab and the thread is essentially gone.
It’s the right call for things you’d never say out loud in a shared room: a question about a symptom you’re worried about, a draft resignation letter, a competitor’s pricing page you’re pulling apart. It’s the wrong call if you want to return to the thread tomorrow, because you can’t.
Memory quietly builds a profile of you
Under Personalization, memory lets ChatGPT carry details between conversations. It might note that you work in logistics, that you prefer short answers, that you’re learning Spanish. Helpful, and also a small dossier sitting on your account.
Open the memory list and actually read it. You can delete individual entries or switch the feature off entirely. If you share a laptop with a partner or a colleague, check this first, because memory survives logouts on the same browser profile.
Custom instructions are where extra detail creeps in
Two free-text boxes ask what ChatGPT should know about you and how it should respond. People fill them with job titles, employer names, internal project codes and unreleased product names. Ask yourself whether you’d be comfortable seeing that text in a screenshot on a conference slide. If not, generalise it. “A mid-size B2B software company” does the same job as naming your actual employer, and “finance” works as well as “accounts payable at [company]”.
Shared links, the setting that surprises people
When you use Share on a chat, you get a public URL. Anyone holding that link can read the whole thread, including anything personal in it. There’s no password and no expiry by default.
You can review and revoke every link you’ve created from the Shared links section of Data controls. Do it occasionally. These URLs end up in Slack channels, support tickets and year-old email threads long after you’ve forgotten they exist.
Files and images carry more than you think
Uploads are where privacy gets messy, because a document rarely contains only what you meant to share. A spreadsheet with a hidden tab. A PDF with tracked changes still attached. A photo with GPS coordinates in the metadata. ChatGPT file upload: how to get useful answers from PDFs, spreadsheets and more covers how to prep a file so the model sees the right page and nothing else.
Images are their own category. Screenshots of dashboards, photos of whiteboards, pictures of colleagues. Treat every upload as a disclosure, because that’s what it is. Our guide to what ChatGPT’s image analysis really sees goes into the limits, including how it handles text and faces.
Deleting chats, exporting everything, closing the account
Deleting a single chat takes one click from the three-dot menu beside the conversation. That removes the thread from your sidebar and from your view.
Export data emails you a download link containing every conversation plus your account details, as HTML and JSON files. If you’ve been using ChatGPT heavily for a year, expect something measured in hundreds of megabytes. It’s also the fastest way to see exactly what you’ve handed over, month by month.
Deleting the account sits at the bottom of Data controls. It’s permanent, it ends any subscription, and there’s no recovery window worth relying on. Export first, then decide.
The risks no setting can patch
- Connectors and third-party apps. If you’ve linked Google Drive, a calendar or a coding tool, that integration has its own permissions and its own data policy. Review connected apps and disconnect what you don’t recognise.
- Custom GPTs built by other people. A GPT someone else published follows their instructions, and those instructions can be written to draw information out of you. Stick to ones from sources you trust.
- Chatbots on websites. If a company’s support widget runs on a language model, the transcript goes to that company. What it actually takes to ship a ChatGPT chatbot for a website is worth reading if you’re on the building side, since the data-handling decisions get locked in early.
- Voice conversations. Audio gets transcribed into the same chat history, with the same retention rules and the same training default.
- Anything pasted in from a shared document. A meeting note containing other people’s names is still their data, even when it sits in your file.
If your employer pays for the seat
Team, Enterprise and Edu workspaces exclude conversations from model training by default and give admins a console for retention, member management, SSO and domain verification. If your company has a licence, use it for work rather than your personal login. It keeps the two histories separate and puts the compliance work where it belongs. The comparison in ChatGPT vs Copilot for your daily workflow is useful here, because Microsoft’s tool inherits your organisation’s existing data boundaries more directly.
A five-minute checklist
- Switch off “Improve the model for everyone” if you haven’t already.
- Read your memory list and delete anything stale or oddly specific.
- Revoke shared links you no longer need.
- Strip personal detail out of your custom instructions.
- Check connected apps and disconnect the ones you don’t use.
- Turn on two-factor authentication under Security.
The habit that beats every toggle
Settings change once. Behaviour compounds. The biggest privacy upgrade available to most people is to stop pasting raw material into prompts. Swap real names for placeholders like [CLIENT A]. Paste three rows of a spreadsheet instead of all four thousand. Strip the cover page off a contract before it ever touches the upload button.
Then make a per-task decision about temporary chat. Routine drafting and brainstorming can stay in your history. Anything involving money, health, legal exposure or another person’s personal data goes into a temporary thread, every time, without deliberating.
And if the balance between convenience and control starts to feel wrong, you’re not stuck with one vendor. There are AI chatbot alternatives with genuinely different data policies, including some that run entirely on your own hardware and never send a prompt anywhere. Knowing what ChatGPT collects is what makes that comparison possible, and it’s the real reason to open the settings menu in the first place.