How to Stop ChatGPT, Claude and Gemini From Training on Your Chats (and Check It Actually Stuck)

What's in this guide

• The short answer
• What "training on your chats" actually means
• Why people are checking this setting again this week
• ChatGPT: the switch, the form, and the catch
• Claude: one toggle and one big number
• Gemini: it is called Keep Activity now, and it has a side effect
• The five minute audit I now do every month
• Does turning it off cost you anything?
• The verdict

If you use ChatGPT, Claude or Gemini with a normal personal account, there is a good chance your conversations are being used to train the next version of the model. Not because you agreed to it in any memorable way, but because the switch that controls it is on by default, buried two menus deep, and named something friendly like "Improve the model for everyone."

This week a thread on Hacker News sent a lot of people to look at that switch for the first time in months, and some of them found it back on after they had turned it off. So I went through all three of my own accounts on September 11, read every help center page the three companies publish on the subject, and wrote down where the switch is, what it does and does not do, and how long each company keeps your chats either way.

The short answer

If you just want to do it and get on with your day, here is the whole thing.

ChatGPT: click your profile picture, open Settings, go to Data Controls, and turn off "Improve the model for everyone." It applies to your whole account, not just that device.

Claude: open Settings, go to Privacy (or go straight to claude.ai/settings/data-privacy-controls), and turn off "Help improve Claude." Whole account, same as above.

Gemini: go to myactivity.google.com/product/gemini and turn off Keep Activity. This one has a side effect, so read that section before you flip it.

Then put a reminder in your calendar to check all three again in a month. That is the step every other guide skips, and it is the reason I am writing this one.

What "training on your chats" actually means

The phrase gets used to mean two different things, so it is worth being precise.

The first is what the assistant remembers about you. ChatGPT's Memory, Claude's memory and Gemini's personalization all read your past chats so the answers fit your context. That stays inside your account.

The second is training. When the setting is on, the company can take your conversations, strip identifying details, and feed them into the process that builds future models. Once that has happened, your text is baked into a model millions of people use, and it cannot be pulled back out. Anthropic says this plainly: turning the setting off stops future training, but "your data will still be included in model training that has already started and in models that have already been trained." OpenAI's help center says the same in softer words.

That is what changes the calculation. A normal privacy setting is reversible. This one is only reversible for the future.

None of this covers the business versions. ChatGPT Team and Enterprise, Claude for Work and Gemini inside Google Workspace are not trained on by default. This guide is about the free and paid personal plans, which is what most people actually use.

Why people are checking this setting again this week

On September 10 a long-time Hacker News user posted that he had turned off ChatGPT's training setting more than once, noted the date, and later found it switched back on. The thread went to the top of the front page. A handful of people in the US, UK, Germany and Switzerland said the same had happened to them. Roughly as many said they had turned it off a year or more ago and it had stayed off.

To be clear, these are user reports, not a confirmed bug or policy. OpenAI had not commented when I checked, and the boring explanations in the thread are plausible: an app update that reset preferences, or a setting that looked saved in the browser but never reached the server.

For your purposes it does not much matter which. If a setting can be reset by an update or a server hiccup, then setting it once and trusting it is the wrong model. You check it the way you check a smoke alarm. That is the premise of the audit at the end of this post.

ChatGPT: the switch, the form, and the catch

The switch is under Settings, then Data Controls, and it is labelled "Improve the model for everyone." On the phone app it is under your profile icon, then Data Controls. OpenAI's help center confirms it is an account-level setting, so turning it off on your laptop covers your phone too.

There is a second route most people do not know about. OpenAI runs a privacy portal at privacy.openai.com with a "Do not train on my content" request form. The help center says that if you opted out through the form, "your account will represent that request," and an OpenAI staff member said on X this week that either route is enough on its own. I did both anyway. The form gives you an email receipt with a date on it, and after this week I like having the receipt.

Now the catch. Turning off training does not make ChatGPT forget anything. Your chats still sit in your history, Memory still works, and OpenAI still stores them. If you want a conversation to leave no trace, use a Temporary Chat. Those are not used for training, do not create memories, do not appear in your history, and are deleted from OpenAI's systems after 30 days.

Claude: one toggle and one big number

Anthropic's version is under Settings, then Privacy, and the toggle is called "Help improve Claude." Same place on the phone. When it is off, Anthropic says it will not use your new chats or coding sessions for future training, and it also stops using your previously stored chats in future training runs. That last part is a little better than what OpenAI promises.

The big number is retention. If the toggle is on, Anthropic may keep your chats in de-identified form for up to five years in its training pipelines. If it is off, a deleted conversation is gone from the back end within 30 days. Five years versus 30 days is the biggest gap between "on" and "off" of any of the three, and it is why I would put Claude at the top of your list if you only fix one today.

Two exceptions. Chats flagged by safety classifiers can still be used to improve Anthropic's trust and safety systems regardless of the toggle. And if you press thumbs up or thumbs down, the whole conversation is stored for up to five years and can be used for training, toggle or no toggle. If you care about this, stop rating answers.

Claude also has an Incognito chat mode, excluded from training even if the toggle is on. It is the right tool for the one sensitive conversation in an otherwise normal week.

Gemini: it is called Keep Activity now, and it has a side effect

Google renamed this setting and the old name still shows up in most guides, so here is the current one: Keep Activity, at myactivity.google.com/product/gemini. You can also reach it from the Gemini app under Settings, then Activity.

When Keep Activity is on, your chats are saved to your Google account, a subset are read by human reviewers, and they are used to improve Google's AI models. Auto-delete defaults to 18 months, adjustable to 3 or 36. The detail most people miss: Google says chats that have been through human review are not deleted when you delete your activity. They are kept for up to three years.

When Keep Activity is off, your future chats are not used for training unless you send feedback, and they are kept for 72 hours so the service can work. Gemini also has a Temporary Chat mode that behaves the same way regardless of the setting.

Now the side effect. Google's privacy notice says that with Keep Activity off, connected apps like Google Workspace are not available. On my account, turning it off meant Gemini could no longer look inside Gmail, Drive or Calendar, which for many people is the whole reason to use Gemini. Calls and messages on Android keep working. If you rely on Gemini reading your inbox, you are being asked to trade that for the opt-out, and only you can make that call.

The five minute audit I now do every month

Setting the switch once and trusting it is exactly the mistake the Hacker News poster made. Here is what I do instead. It takes five minutes and I have a repeating calendar reminder for the first of the month.

Step one: check each switch, do not assume. Open the three settings pages in three tabs. ChatGPT Data Controls, Claude Privacy, Gemini Keep Activity. Look at the actual state, not what you remember doing.

Step two: check on the phone too. OpenAI and Anthropic both say the setting is account-wide, and on my accounts it was. But the phone app is where updates land first. Thirty seconds.

Step three: screenshot the state. One per service, in a folder with the date in the file name. If a setting ever changes without you touching it, you have something better than memory to point at.

Step four: keep the receipt. File the confirmation email from OpenAI's privacy form. A written opt-out is a stronger position than a toggle.

Step five: delete what you do not need. Training aside, a chat that no longer exists cannot be reviewed, leaked or subpoenaed. I delete anything with tax numbers, medical details or other people's names as soon as the task is done, and I use Temporary Chat or Incognito for that kind of thing to begin with.

Five steps, once a month, and you never have to wonder again.

Does turning it off cost you anything?

For ChatGPT and Claude, no. I have had both switched off for months and cannot point to a single feature that stopped working. Memory works, projects work, the models are the same models. The "improve the model for everyone" framing makes it sound like you are opting out of something you benefit from, but the benefit goes to the company's next training run, not your account.

For Gemini, yes. Losing Workspace access is the cost of turning off Keep Activity, and if Gemini's main job in your life is summarizing your inbox, the switch may not be worth it. A reasonable middle ground: leave Keep Activity on, drop the auto-delete window from 18 months to 3, and use Temporary Chat for anything sensitive. You are still being trained on, but the window is smaller and the sensitive stuff never enters it.

And if none of that feels good enough, you can run a model on your own computer or phone, where the question does not arise. I have written about how to do that on a laptop and on a phone, and for the sensitive ten percent of what I ask an AI, that is what I use.

The verdict

Turn it off. All three switches, today, starting with Claude because it has the longest stated retention if you leave it on. It takes under five minutes and there is no downside for ChatGPT and Claude users. Gemini users should weigh the Workspace trade-off first.

Then treat the switch as something you verify rather than something you set. Whether this week's reports turn out to be a bug or nothing at all, a setting you have not looked at in six months is a setting whose state you do not actually know.

If you are still deciding which assistant to use in the first place, my side by side comparison of Claude, ChatGPT and Gemini is the place to start, and the guides to Claude Fable 5.1 and GPT-6 Astra cover what the paid plans get you. Whatever you pick, go and find the switch first.