Which tool to choose
The choice isn't "which collects the least" in absolute terms, but which collects the least given how you use it.
- If you want the tightest data perimeter, Anthropic's Claude largely stays confined to activity inside the service: it doesn't cross-reference your browsing or other apps the way the other two do.
- If you're already inside Google (Gmail, Drive, Maps, Android), Gemini gives you the deepest integration, but it's precisely that integration that makes it the hungriest for data: it sees much of what the Google account touches.
- If you use ChatGPT as an all-rounder, know that its collection goes beyond the app: it includes signals tied to browsing and activity on other platforms.
In all three, the lever that matters most isn't the brand but a single setting: training on your data. Let's see what they collect and where you adjust it.
How to do it (find out and reduce what they collect)
The three types of data are the same everywhere. The range changes and the menu names change.
Conversation content. Everything you type, the files you upload, the images, the custom instructions. It's the most sensitive data because it often contains names, numbers, real documents. This data, by default, feeds model training on ChatGPT (Free, Plus, and Pro personal accounts) and on Claude (the "Help improve Claude" option is on by default).
Account and device data. Email, payment method, IP address, device and browser type, approximate location. They serve to run the service and to bill; they can't be turned off without closing the account.
Usage signals. Which features you open, how long you stay, what you delete. Here ChatGPT and Gemini look beyond the app itself.
The point where you have real control is the first. To turn off training, the concept is the same from browser or app, only the labels change:
- On ChatGPT: open your profile, then Settings, then the data controls section (Data Controls). There you'll find the "Improve the model for everyone" switch: turn it off.
- On Claude: click your name at the bottom left, then Settings, then the "Privacy" item in the side menu. You'll find the "Help improve Claude" switch: turn it off.
- On Gemini: the lever is "Gemini Apps Activity" in your Google account. Pausing it stops new conversations from being used to improve the services.
If you can't find the button because the interface has changed, there's a hold that doesn't age: in the text box ask the AI itself "Where do I find the setting to not use my chats for model training?" and have it give you the path of the updated menu.
Check: after turning off the switch, the conversations you start from that moment no longer go into the next training. Those already used in a past training stay in: they can't be pulled out.
The operational syntax, if you want a reminder of what to check on a new service:
I'm considering whether to use [name of AI service].
Tell me concretely:
1. does it use my conversations to train its models, and is it on or off by default?
2. where is the exact setting to turn it off?
3. how long does it keep my chats after I delete them?
Answer only with what you know for certain, without inventing menu paths.
A concrete example
Marta, an accountant, uses free ChatGPT to rewrite emails to clients. She often pastes in drafts with names and figures. She opens Settings, goes to the data controls, turns off "Improve the model for everyone." From that moment the new emails she pastes don't end up in training. For the more delicate drafts, instead of trusting only the switch, she also enables the temporary chat: that conversation isn't kept in history, creates no memory, and isn't used for training. Result: she keeps working fast, but her clients' data stops being raw material.
When it does NOT work (and how to fix it)
If you have a paid subscription and think you're already safe
Paying twenty euros a month doesn't turn off training. On personal plans (even Plus, Pro, Advanced) the use of chats to improve the models is almost always on by default. Fix: turn off the switch by hand as above. If you need the contractual guarantee that your data is never used, you need the business/team or enterprise plans, where by contract they don't train on customer content.
If you turned off training but the chats stay saved
Turning off training doesn't delete the history: conversations stay in your account until you remove them yourself. Fix: delete the chats you don't need and, for sensitive topics, use the temporary chat mode, which saves nothing in history.
If the data has already ended up in a model
There's no button that pulls it back out of an already-trained model after the fact. The honest fix: the only lever is the future. The sooner you turn off training, the fewer of your conversations go into it from now on.
A tip from someone who actually uses it
Don't judge an AI by its privacy marketing page: judge two concrete numbers. First, is training on your data on or off by default? Second, how long does it keep chats after deletion? On these two fronts the differences are clear and they change over time, so recheck the setting every few months: companies move the switches and sometimes turn them back on after an update.
Frequently asked questions
Do ChatGPT, Claude, and Gemini read my chats with human eyes?
As a rule no, not across the board. The content is handled by automated systems. But a share may be reviewed by people to check quality or safety, especially if the conversation is flagged or if you've left training on. It's one more reason to turn off the switch when you handle confidential data.
If I delete a chat does it disappear right away?
No, there's usually a technical window: on ChatGPT deleted chats are removed within about thirty days. The deletion starts when you do it, it isn't instantaneous on the server side.
Which one collects the least data in absolute terms?
Among the three, Claude is the one with the tightest range, because it largely stays inside its own service. But the question is poorly framed: the data that really weighs is your conversations, and those you can protect on any platform by turning off training. Real privacy depends more on how you configure it than on the logo at the top left.
Does having nothing to hide keep me safe?
That's the myth to bust. The problem isn't hiding: it's that you paste in other people's data. The client's contract, a colleague's phone number, a family member's diagnosis. That data isn't yours to give away to a training run, and once it's in, it doesn't come out. "I have nothing to hide" stops holding the exact moment you write someone else's name.