How to Stop OpenAI Dots From Training on Your Data (2026)

Does OpenAI train on your dot's data?

It depends on your plan and one setting. OpenAI says dots follow "the same permissions and controls as ChatGPT". So on a personal account, the ChatGPT training control applies. On Business, Enterprise and Edu, workspace content isn't used to train models by default.

Quick context on why this matters more than it did for a chatbot. A dot is OpenAI's always-on personal agent, launched on September 29, 2026 and powered by GPT-6 Astra. It has its own cloud computer and browser, connects to 4,000+ apps, and works toward your goals 24/7. It also learns your preferences, standards and working habits from feedback over time.

It even does proactive research: it scans your connected apps in the background, with read-only access. The upside is real. In one OpenAI example, a freelance writer's dot noticed he'd forgotten to invoice a publication, pulled the details from the email thread and drafted the invoice, sending it only after his approval. The flip side: an agent that notices things has to read a lot of your stuff, whether or not you asked a question that day.

So there are two jobs here. Job one: make sure what you share isn't used for model training. Job two: shrink what the dot can see in the first place. The second job is where most of the real privacy wins are, because data a dot never touches can't end up anywhere.

What hasn't OpenAI said about dot data?

Quite a lot, and it's worth being plain about it. As of launch, OpenAI hasn't published dot-specific details on how long a dot's memories, conversations, background research or cloud-computer activity are kept.

Here's what we don't know yet. How long the record of your dot's browsing on its cloud computer is stored. Whether findings from proactive research are kept after you've seen them. How quickly data removed by a Reset disappears from OpenAI's systems. And whether a paused dot's memories are handled any differently from an active one's.

Here's what we do know. Dots follow ChatGPT's permissions and controls. Proactive research is read-only: during that background scan, the dot can't send messages, modify content or control your computer or browser. When the dot signs into websites for you, it uses saved passwords that are stored without the model seeing them. And sensitive tasks such as changing passwords always stay with you.

OpenAI is also candid that dots can make mistakes, and it runs a monitoring system that can pause or stop a dot over safety concerns. That's reassuring for safety. It isn't the same as a retention policy. Until OpenAI spells one out, assume anything your dot reads could be stored for some time, and decide what to connect with that in mind. For the wider safety picture, see our guide on whether OpenAI dots are safe to use.

Step 1: How do you turn off "Improve the model for everyone"?

Open ChatGPT, go to Settings → Data Controls, and switch off "Improve the model for everyone". It takes under a minute, and it's the one setting this whole article is named after.

Step 1.1. Open ChatGPT on the desktop app or on the web. Step 1.2. Open Settings (look for it under your profile). Step 1.3. Look for Data Controls. Step 1.4. Find "Improve the model for everyone" and turn it off. If the label on your screen differs slightly after an update, go by the meaning: you're looking for the option about using your content to improve OpenAI's models.

Does it cover your dot? OpenAI says dots follow the same permissions and controls as ChatGPT, so this is the control that applies. OpenAI hasn't published a separate, dot-only training switch at launch. If one appears later, turn that off too.

Two caveats. Treat the setting as forward-looking: change it on day one, not day ninety. And it's tied to your account, so it should follow you across devices, but a ten-second check on your phone costs nothing.

Also remember what this toggle doesn't do. It's about training. It doesn't stop your dot from reading your connected apps, remembering your preferences or keeping its history. That's what the next steps are for.

Step 2: Which apps should you connect to your dot?

As few as possible. Every connection is something your dot can read in the background during proactive research. Connect what it needs for the jobs you actually give it, and nothing "just in case".

Here's the trap people fall into. Say you connect your cloud drive to ask about one proposal. Naming that one document in your request doesn't narrow the connection. The connection's access is set when you approve it, not by the question you ask. If the permission covers the whole drive, the dot's background research can see the whole drive.

So work through your connections like a closet clear-out. For each app, ask: did I use this through my dot in the last month? If not, disconnect it. You can reconnect in a minute if you ever need it again. Then check the other side too: most services keep a list of connected or authorized apps in their own security settings, where you can remove access.

Two more switches deserve a slow, careful yes. With your permission, a dot can use your own computer, not just its cloud one. And it can sign into supported websites using saved passwords. Both are handy, and OpenAI says the model doesn't see those passwords. But if you don't need them, don't grant them. Less access means less that can go wrong.

Step 3: How do custom rules keep your dot in check?

Custom rules decide what your dot may do: allow an action, require your approval, or block it entirely. You edit them in the ChatGPT mobile app: look for Customize → Custom rules.

Start strict. Set "Ask before taking action" for anything consequential, like sending, publishing, deleting or changing records. Block the things you never want it to touch. You can loosen rules later, once you've watched how the dot behaves. Loosening after a mistake is the more expensive way to learn.

OpenAI adds its own layers on top. An auto-review system checks consequential actions, and sensitive tasks such as changing passwords always stay with you. Those are good defaults. They're not a substitute for your own rules, because only you know that a certain client must never get an automated email.

One honest limit: custom rules are about actions. OpenAI hasn't described them as a way to stop your dot from reading a connected app. For reading, the real control is the connection itself, which is why Step 2 comes first.

Step 4: How do you delete what your dot remembers?

Use Reset. Open your dot's profile, look for the ••• menu and choose Reset. It deletes the dot's conversations, memories and schedules. That's the cleanest way to make a dot forget you.

Before you reset, open Activity View. It shows the dot's cloud computer and its background work, split into In progress, Scheduled and Completed. Check what it's been doing and note any schedules you'd want to recreate, because Reset removes them too.

If you want the dot to stop but keep its memory, use Pause in the same menu instead. Pausing is reversible. Reset isn't.

A good habit is a monthly look at Activity View and your connections. Five minutes. Disconnect what you stopped using, and reset if the dot has picked up personal details you'd rather it didn't keep. And if you're leaving dots altogether, here's how to pause, reset or cancel your dot.

Step 5: Where should company data go?

Into a business workspace, not a personal Pro account. On Business, Enterprise and Edu, workspace content isn't used to train models by default. That's the plan-level guarantee a personal toggle can't give your whole team.

Business Premium is the business route to dots: $100 per seat a month billed annually, or $125 monthly, with your first dot included. It's also how you get a dot in the EEA, Switzerland and the UK at launch, since Pro dots exclude those regions. On Enterprise, Edu and Healthcare, dots are a beta that's off by default until an admin enables it, so your IT team gets to decide before anyone connects anything.

For larger companies, OpenAI is piloting specialist dots with their own organizational identity, credentials and company-system access. That's a conversation for your admin and your security team, not a personal settings tweak.

The simple rule for a small business: client files, contracts and finance data go in the workspace your company controls. Your personal Pro dot gets personal jobs.

Want an assistant that sees less by design?

Then pick one with fewer doors. The strongest privacy setting is an assistant that never had access to begin with. Agent Sonic is built around that: it works from what you send or forward in the chat, not from a background scan of your apps.

Sonic is a personal AI assistant that lives entirely inside WhatsApp. No app to install, no dashboard, no cloud computer. You save it as a contact and message it.

Here's the privacy shape of that. Sonic doesn't control your computer. It doesn't make purchases. It doesn't connect to 4,000 apps. It sees the messages, voice notes and files you choose to send it, plus Google Calendar, Sheets or Docs only if you connect them. When you ask it to message someone for you, it reports their reply back in your chat.

What that looks like day to day: "Remind me to call the accountant every Monday at 9." Sonic sets a recurring reminder. Forward one contract PDF and ask "When does the notice period end?" It reads that file, because you sent that file. "Message Omar and ask if the invoice was paid, then tell me what he says." It reaches out and reports back. "Every Sunday at 8, remind the building group that rent is due." It posts the reminder in the group for you.

To be fair to dots: if you need an agent that researches for hours across your tools, writes code or works inside Slack and Teams, a dot does far more than Sonic. Sonic doesn't do any of that. And we won't make claims here about Sonic's data retention; check its privacy details when you sign up, and follow the same rule as with any assistant: don't send what you wouldn't want stored.

But if your goal is to delegate reminders, follow-ups, group nudges and quick document questions while keeping your exposure small, a WhatsApp assistant that only sees what you hand it is the lighter choice. Sonic is available now. Get access at tryagentsonic.com/contact.

Frequently asked questions