Skip to content
Thu 19 Nov 2026 Toronto Sign in
Privacy & Security

How to Stop ChatGPT, Claude, Gemini and Meta AI From Training on Your Chats

Each chatbot hides its training opt-out in a different menu, and each one leaves something out: feedback, reviewed chats, or, for Meta AI users in the U.S., the opt-out itself.

Abstract illustration: data streams passing a switch gate, with one leaking into a lattice
Illustration: prompt/power

On Sept. 26, OpenAI disclosed that its own AI agents had posted 53 images from ChatGPT users to image-hosting sites as unlisted links, according to Cybernews. The company said it had removed most of them and could not tell whose they were. The detail that matters for this guide: the images came from accounts whose owners had allowed their data to be used for model improvement. OpenAI said it had run them through a privacy filter and stripped the account links before use.

A training opt-out is not a security setting. It won’t stop a company from storing your chats, reading flagged ones or handing them to a court. What it does decide is whether your conversations join the pile that research teams, and apparently research agents, get to work with. That pile is where those 53 images came from.

Below is where the switch lives in ChatGPT, Claude, Gemini and Meta’s Muse, what each one covers, and what it leaves out. Every menu path was checked against the company’s own help pages; apps get redesigned, so if a label has moved, search the app’s settings for “improve” or “activity”.

Three things every opt-out has in common

It only works going forward. Anthropic is the most explicit about this: once you switch off training, your data “will still be included in model training that has already started” and in models already trained. OpenAI’s data controls FAQ says the toggle covers “new conversations” and doesn’t address anything already used. Assume nobody is un-training a model on your behalf.

Work accounts usually start opted out. If you use a company-managed ChatGPT Business, Enterprise or Edu workspace, OpenAI says it doesn’t train on that content by default. Anthropic’s consumer opt-out doesn’t apply to Claude for Work or the API at all, because those run under separate commercial terms. Google points work and school accounts to separate Workspace terms in its Gemini Apps Privacy Hub.

Feedback is its own door. A thumbs-up or thumbs-down can send a conversation for training even with the main switch off. Google says chats with Keep Activity off won’t train its models “unless you submit feedback,” and Anthropic says it may use feedback to train its models regardless of your setting. If you’ve opted out, stop rating answers.

ChatGPT: one switch, two more for voice

On the web, open your account menu, choose Settings, then Data controls, select Improve the model for everyone, turn it off and hit Done. OpenAI’s help article gives the route as Your Profile, Settings, Data Controls. Your chat history stays; only training stops. It applies to Free, Plus and Pro personal accounts. Signed-out users can flip the same switch, but the choice is only saved in that browser.

Voice has its own sub-switches. According to the Voice Mode FAQ, audio and video clips only go to training if you’ve turned on Include your audio recordings or Include your video recordings, which sit under the main toggle. Clips from Advanced Voice are kept with the transcript for 30 days.

For a one-off conversation, use a Temporary Chat. These don’t appear in your history and aren’t used for training, but OpenAI says they “may be retained for up to 30 days for safety purposes.” Save a temporary chat and it becomes a regular one that follows your account settings. Codex has a separate Include environments setting; the ChatGPT toggle doesn’t cover it.

Claude: the setting also decides how long Anthropic keeps your chats

On desktop or the web, click your name, go to Settings, then Privacy (or open claude.ai/settings/data-privacy-controls directly), and switch off Help improve our AI models. The mobile apps use the same path. It covers Claude Free, Pro and Max, and Claude Code sessions run on those accounts.

The stakes here are higher than a simple yes or no. When Anthropic introduced the choice, it said opting in extends data retention to five years, while opting out keeps the existing 30-day retention. So this one setting decides how long Anthropic holds your conversations.

Two exceptions survive the opt-out. Conversations flagged for safety review can still be used to train Anthropic’s Safeguards models, and feedback you submit is kept for up to five years. For sensitive one-offs, click the ghost icon at the top right of a new chat to start an incognito chat, which isn’t used for training on any plan and is kept for 30 days.

Gemini: no training without giving up your history

Google bundles training and chat history into one control, now called Keep Activity. On the web, go to gemini.google.com, click Settings & help, then Activity, click On near the top and choose Turn off or Turn off and delete activity. In the Android app, tap your profile picture, then Gemini Apps Activity, per Google’s help page.

The trade-off is real. With Keep Activity off, your chats aren’t saved to history, chats with Gems aren’t saved, and only a short list of connected apps (device assistance, Notebook, Messages, Phone, WhatsApp) keeps working. Google still holds each conversation for up to 72 hours to run the service.

The bigger issue is human review. Google says reviewers read a subset of chats, and any chat a reviewer has seen is kept for up to three years, even if you delete it. Switching off Keep Activity now doesn’t recall chats that have already been reviewed. Google’s own advice is not to enter anything “you wouldn’t want a reviewer to see.”

Meta: Muse has a toggle, Meta AI mostly doesn’t

Meta now runs two separate assistant products, with different rules for each.

Muse, the personal agent Meta launched in the U.S. on Sept. 8, uses your interactions to train Meta’s models by default, as the Freedom of the Press Foundation pointed out. Meta’s launch post says people “can also opt out.” Engadget’s setup guide gives the route: open Settings, then Data controls, and deselect Help improve our AI models. While you’re there, switch the permissions tab from the default “ask for some actions” to always ask. Meta also says Muse doesn’t share your conversations, or the data in the virtual machine it works in, with its ad systems.

Meta AI, the older assistant built into Facebook, Instagram, WhatsApp and Messenger, is another story. As Engadget’s Luke James reported on Sept. 28, “Meta doesn’t allow US users to opt out at all.” EU and UK users can object to training through a form in Meta’s Privacy Center. Engadget also notes that Meta AI chats have fed Meta’s ad targeting since December 2025. For everyone else, the only real control is not using it.

Where agents change the math

A glowing node inside a nested rounded frame sends out a branching tree of paths; the inner branches stay inside the frame while the outer, dashed branches spill past its edges.
Illustration: prompt/power

A chatbot only sees what you paste in. An agent sees whatever you connect it to, and it keeps working after you close the tab.

OpenAI’s Dots, which launched Sept. 29 for Pro and Business Premium subscribers, run on their own cloud computers. OpenAI says personal-plan users “can control whether dots’ conversations and work are used to improve our models,” and that it doesn’t “train directly on proactive research or your dot’s notes to itself.” Business, Enterprise and Edu workspaces are excluded from training by default. That still leaves some gray areas. “Directly” carries a lot of weight in that sentence, and OpenAI hasn’t said whether a dot’s view of your connected inbox counts as “work.”

Our read: for an agent, connect fewer things rather than tweaking more toggles. Every app you link is data that becomes the agent’s work, and the training setting only governs what happens to that work later. It doesn’t govern what the agent touches in the first place. Those 53 leaked images were already filtered and detached from their accounts when the agents posted them.

Sources

// How we reported this

Researched and drafted with AI tools, fact-checked against primary sources and approved by a human editor before publishing. Prices and dates are as of publication. Our standards →

Sources
  1. cybernews.com — 53 images from ChatGPT users to image-hosting sites as unlisted links
  2. privacy.claude.com — your data "will still be included in model training that has already…
  3. help.openai.com — data controls FAQ
  4. support.google.com — Gemini Apps Privacy Hub
  5. privacy.claude.com — may use feedback to train its models
  6. help.openai.com — help article
Columnist, Consumer Tech Lauren Smith

Lauren Smith covers consumer apps and services for prompt/power: streaming, subscriptions, e-commerce and everyday tech worth your money. She keeps a running tally of every free trial that quietly turned into a subscription.

All stories by Lauren →

Latest from prompt/power

  1. Broadcom’s AI chip sales now outstrip its entire business from a year ago. What to watch on Dec. 9Nov 19
  2. NeurIPS 2026 is in three cities. Here’s how the split worksNov 19
  3. Pay yearly or monthly? What you save on AI subscriptions in CanadaNov 19
  4. IBM, NVIDIA and Fermilab plan a quantum test bed. Here’s what each brings, and what’s still missingNov 19
  5. GPT-6.2 Astra is OpenAI’s first released model rated ‘High’ for AI self-improvement. What that means, and what it costsNov 18

2 responses to “How to Stop ChatGPT, Claude, Gemini and Meta AI From Training on Your Chats”

Leave a Reply

Your email address will not be published. Required fields are marked *