Type a question into ChatGPT or Google Gemini and your message doesn’t bounce off a machine and vanish. It travels to a company’s servers, gets saved next to your account, and — depending on two settings most people never touch — may be used to train the AI and read by a human being.

None of that is a secret. It’s written down in help pages nobody opens. So we opened them.

The essentials:

• Your chats live on the company’s servers, tied to your account, not on your phone.
• ChatGPT and Gemini both use your conversations to improve their AI by default, and a slice of chats get read by real people.
• You can switch both off in settings, but “off” isn’t the same as “not stored” — deletion runs on its own clock.

Where does your message go after you hit Enter?

It leaves your device almost immediately. The AI model is far too big to run on a phone or laptop, so your message travels over the internet to a data center, a warehouse full of computers run by OpenAI, Google, or whoever makes the tool. The model reads it there, writes a reply, and sends the reply back. The round trip takes a second or two, which is why it feels like the answer came from your screen. That warehouse has to be powered and kept cool, which is its own story: we looked at how much water one ChatGPT message actually uses.

What people miss is what stays behind. Your message and the AI’s reply are saved on those servers and attached to your account. That’s exactly why your conversation from last Tuesday is still sitting there when you reopen the app. The convenience and the storage are the same feature. You can’t have the history without the copy.

Does ChatGPT use my conversations to train its AI?

On a personal account, yes, unless you turn it off. OpenAI’s help center says it plainly: it may use your content to train the models that power ChatGPT.

Training means feeding the AI examples so its future answers get better. It’s how these systems learn in the first place, and it’s worth understanding the basics of how AI actually works if that idea is new to you. What it doesn’t mean is that your words get memorized and read back to a stranger. Your chat becomes one drop in an ocean of text.

To switch it off in ChatGPT: click your profile icon, go to Settings, then Data Controls, then turn off “Improve the model for everyone.” On mobile it’s the sidebar menu, your profile icon, then Data Controls. If you use ChatGPT without an account, the same switch is there under the ? icon in the bottom-right corner.

Three useful details. Your chats still show up in your history afterward, because the switch controls training, not saving. The setting covers your whole account, so flipping it on your laptop also flips it on your phone.

And there’s a hole in it worth knowing about. OpenAI says that if you rate an answer with thumbs up or thumbs down, the entire conversation attached to that rating can still be used for training, even with the switch off. Google’s opt-out has the same carve-out. The feedback buttons quietly reopen the door you just closed.

Does Google Gemini use my chats to train its AI?

Yes, by default, and Gemini’s switch has a different name and lives somewhere else. It’s called Keep Activity, and you’ll find it in your Google Account rather than in the Gemini app.

While it’s on, your conversations are saved to something called Gemini Apps Activity and used to improve Google’s services, including with help from human reviewers. Google is unusually direct about what that means for you. Its own privacy notice tells users not to enter data that’s confidential or that you wouldn’t want a reviewer to see.

Here’s how the two compare on the settings that matter:

	ChatGPT	Google Gemini
What the setting is called	“Improve the model for everyone”	“Keep Activity”
Where to find it	Settings → Data Controls	Your Google Account activity settings
On by default?	Yes, on personal accounts	Yes
Turning it off stops	Your chats being used to train ChatGPT	Future chats being saved and reviewed to improve Google services
What still happens	Chats stay in your history	Chats held 72 hours to answer you and guard against abuse

Do real people read AI chats?

Some of them, yes. This is the part that surprises people most.

Google says a subset of Gemini chats are read by human reviewers, including trained reviewers working for its service providers. Their job is to judge whether the AI’s answers were low-quality, inaccurate, or harmful, which is how the tool gets less wrong over time. The one case OpenAI spells out is Temporary Chats — the ones that never appear in your history — which it says may be reviewed only to monitor for abuse.

So nobody is browsing your conversations for fun, and no human sees the vast majority of chats. But the number isn’t zero, and that’s the honest way to think about it. If you’d be uncomfortable with a contractor in an office somewhere reading a particular message, don’t send that message.

How long do AI companies keep your chats?

Longer than most people assume, and the clocks are different for each company.

For Gemini, your activity auto-deletes after 18 months by default. You can change that to 3 months, 36 months, or keep it until you delete it yourself. Turning Keep Activity off doesn’t drop it to zero either — Google still holds chats for 72 hours so it can respond to you and protect against abuse.

The number that catches people out is this one: Gemini chats that were picked up for human review are kept for up to three years, disconnected from your Google Account. And Google says those reviewed chats are not deleted when you delete your activity.

For ChatGPT, saved chats stay until you remove them. Temporary Chats are deleted after 30 days on their own.

Does deleting a chat actually delete it?

Usually, on roughly a 30-day clock, with one exception that turns out to matter.

When you delete a ChatGPT conversation or your whole account, OpenAI removes it from your account immediately and schedules it for permanent deletion from its systems within 30 days. The policy adds four words that do a lot of work: unless it’s required to keep the data for legal or security reasons.

That’s not a hypothetical. In the copyright lawsuit brought against OpenAI by The New York Times, a court ordered the company to preserve chat logs that would normally have been wiped, including conversations users had deleted themselves.

OpenAI fought it and won. The company says that obligation ended on 26 September 2025, and deleted chats are back on the ordinary 30-day clock. Two leftovers are worth knowing. OpenAI is still holding a limited set of user data from April to September 2025 under legal hold, locked to a small audited team. And conversations from the UK, Switzerland, and the European Economic Area were never caught up in it at all.

The takeaway isn’t that anyone lied to you. It’s that a delete button is a promise from a company, and for several months in 2025 a court overruled that promise for millions of people. Treat deletion as very likely rather than guaranteed.

Is a work or school account different?

Yes, in both directions, and it’s worth knowing which one you’re logged into.

Business and education plans run under a contract with your organization: ChatGPT Business, Enterprise, and Edu, or Gemini through a Google Workspace account. By default, those conversations aren’t fed into training for the public models, and admins get extra controls over how long things are kept.

The trade-off runs the other way. Your organization’s administrator decides retention and may be able to see your activity. Private from the AI company doesn’t mean private from your employer. If you’re using AI at work, your company’s own policy is the document to read.

How do I stop AI tools from keeping my chats?

None of this is a reason to stop using these tools. It’s a reason to spend five minutes on the settings:

1. Flip the training switch if you’d rather not contribute. ChatGPT: Settings → Data Controls → “Improve the model for everyone” → off. Gemini: turn off Keep Activity.
2. Use a temporary chat for one-off questions you don’t want in your history. Both tools have one.
3. Shorten Gemini’s auto-delete window from 18 months to 3 if you don’t need the history.
4. Go and look at what’s already there. The same Data Controls screen in ChatGPT lets you export everything on your account, so you can read back what you’ve typed over the past year instead of guessing. Most people find it a sobering exercise.
5. Keep the genuinely sensitive things out of the box: passwords, ID numbers, medical details, other people’s private information, anything under a confidentiality agreement. Our guide to whether AI is safe to use goes deeper on where to draw that line.
6. Don’t read “off” as “not stored.” Off means not used for training. The chat still went to a server.

The mental model that works: an AI chatbot is a service you’re logged into, like email or cloud storage. You probably wouldn’t email yourself your bank PIN. That same instinct is the right one here.

Still deciding which tool to use? We’ve put ChatGPT and Gemini side by side, and there’s more in our AI safety guides.

Every setting name and retention figure here was checked against OpenAI’s and Google’s own help pages in August 2026. Both companies change these without notice, so it’s worth opening the settings yourself.

Frequently asked questions

Does turning off training delete my old chats?

No, and this trips people up. The training switch and the delete button are two different things. Turning off "Improve the model for everyone" in ChatGPT stops your conversations being used to improve the AI from that point on, but every chat already in your history stays exactly where it is. If you want the old ones gone, you have to delete them yourself. Turning the switch off also does nothing to conversations that were already used for training before you flipped it.

Is it more private to use ChatGPT without signing in?

A little, but less than people assume. Signed out, there's no account for your chats to be filed under and no history to scroll back through. Your messages still travel to the company's servers, still get stored there, and can still be used to improve the AI unless you say otherwise — the same "Improve the model for everyone" switch is available signed out, tucked under the ? icon in the bottom-right corner. Signing out changes what's tied to your name, not whether the message is kept.

Do temporary chats really disappear?

Mostly, on a 30-day timer. A Temporary Chat in ChatGPT doesn't show up in your history, doesn't create or use memories, isn't used to train the models, and is deleted from OpenAI's systems after 30 days. The one carve-out OpenAI names is that these chats may still be reviewed to monitor for abuse. So it's genuinely the most private way to ask a one-off question, but it isn't the same as the message never existing.

Can a court force an AI company to hand over my chats?

It can force the company to keep them, which is the step before handing them over. OpenAI's deletion policy has always said chats go within 30 days unless it's legally required to keep them, and that clause got tested: in the copyright case brought by The New York Times, a court ordered OpenAI to preserve logs including chats users had deleted. OpenAI says that obligation ended on 26 September 2025, and that the data it still holds from that period is locked to a small audited team and has not been handed to anyone.

Can I see everything an AI company has stored about me?

For ChatGPT, yes. The same Data Controls screen where you turn off training also lets you export your data, which gives you a file of your conversations to read back. It's the quickest way to stop guessing about what you've typed into it over the past year. Gemini's equivalent is your Gemini Apps Activity page in your Google Account, which is also where you set how long that activity is kept.

What should I never type into an AI chatbot?

Passwords and PINs, bank or card numbers, government ID numbers, medical details, other people's private information, and anything covered by a confidentiality agreement at work. The test that works: if you'd be uncomfortable with a reviewer in an office somewhere reading it, don't send it. Google says this outright in its own privacy notice, telling users not to enter data they wouldn't want a reviewer to see.