Story11 Nov 20265 min read

Who can read your AI chats, and how long they’re kept

Think about the most personal thing you ever typed into ChatGPT or Gemini.

A symptom that scared you at night. A fight with your partner. Your salary. A contract you pasted in “just to check it”.

Who else can read that? How long is it kept? And where can it end up?

We read the companies’ own help pages and the news coverage. Every claim below has a source at the bottom of this page.

Someone can read it

Google has human reviewers read a subset of Gemini chats, including reviewers from outside service providers.¹ Google’s own advice:

“Please don’t enter confidential information that you wouldn’t want a reviewer to see or Google to use to improve our services.”¹

Connected your Gmail or Drive? Google says it doesn’t train directly on your inbox. It does train on “summaries, excerpts, generated media, and inferences” from your emails and files, and humans review a subset. For a short email, Google writes, “the summary may just be the email or file.”²

Deleting it doesn’t delete it

Gemini keeps chats that a reviewer read for up to three years, even after you delete your activity.¹
With Gemini’s Keep Activity off, your chats are still kept for 72 hours.¹
If you disconnect Gmail from Gemini, what it already pulled in stays. Google says it can still be used for training.²
During the New York Times lawsuit, a court ordered OpenAI to keep even deleted ChatGPT chats, from May to late September 2025. What was saved in that period is still kept.^{3, 4}
Once your words have been used to train a model, there is no button to take them back.

It doesn’t stay between you and the AI

Sam Altman, the CEO of OpenAI, said young people use ChatGPT “as a therapist, a life coach”. He also said those conversations have no legal privilege, the way they would with a doctor or a lawyer. In a lawsuit, in his words, “we could be required to produce that.”⁵

Since 16 December 2025, Meta uses your chats with Meta AI to personalize ads and content on Facebook and Instagram. There is no opt-out. The EU, the UK and South Korea were left out.⁶

One click and it’s public

More than 370,000 Grok chats ended up in Google search through the share button. Some contained names, personal details and at least one password.⁷

In 2025, people using the Meta AI app posted private chats to a public feed without realizing it.⁸ How a shared chat ends up in Google.

Training is on by default

ChatGPT trains on your chats on Free, Plus and Pro unless you turn it off.⁹
If you are 18 or older, Gemini’s Keep Activity is on by default, and your chats are kept for 18 months.¹⁰
Since autumn 2025, Anthropic trains Claude on consumer chats unless you opt out. If you leave it on, your data can be kept for up to five years.¹¹
A Stanford study of six large US AI developers, including Google, OpenAI and Anthropic, found that all six appear to train on user chats by default. Some keep that data indefinitely.¹²
In the Netherlands, the Dutch Data Protection Authority has received data breach reports since 2024 about employees who pasted personal data into chatbots. Patient data at a GP practice. Residents’ data at the municipality of Eindhoven.^{13, 14}

The rules keep moving one way

In 2025 alone, Google started using a sample of your uploads to improve its services,¹⁵ Anthropic turned training on by default¹¹ and Meta started using AI chats for ads.⁶ Each time, it was up to you to notice and opt out, if that was possible at all.

Meanwhile AI moves into your email, your browser and your files. The more you connect, the more of your life runs through someone else’s servers.

Local AI will fix this, just not yet

Models that run entirely on your own device get better fast, and what you type there doesn’t have to leave it. The technology exists: Apple and Google run small models on devices, and anyone can download open models like Llama or Gemma. Why small models are enough for most jobs.

But the assistants most people use every day still run in the companies’ own cloud, where your chats are stored and, by default, used for training. We don’t expect that to change by itself.

What you can do tonight

In ChatGPT, go to Settings, then Data Controls, and turn off “Improve the model for everyone”.
In Gemini, turn off Keep Activity and disconnect the apps you don’t need.
In Claude, turn off “Help improve Claude” in the privacy settings.
Use a temporary chat for anything personal. Gemini still keeps it for 72 hours, ChatGPT for up to 30 days.^{1, 9} How temporary chats work.
Leave out health details, account numbers, ID numbers and passwords.
Don’t press “share” on a chat you wouldn’t post in public.

Where Melo comes in

That gap is why we build Melo, an AI for iPhone and Mac that does work for you. In the app, Melo answers with the AI model on your iPhone or Mac itself, and your conversations stay on your device.

It’s still early. Join the iPhone beta and tell us what it should do next.

Keep it yours. Try Melo

Sources

Google, Gemini Apps Privacy Hub, updated October 2026
Google, About personalization with Connected Apps
Malwarebytes, OpenAI forced to preserve ChatGPT chats, June 2025
Yahoo News, Judge lifts order requiring OpenAI to preserve ChatGPT logs, October 2025
TechCrunch, Sam Altman warns there’s no legal confidentiality when using ChatGPT as a therapist, 25 July 2025
The Register, Meta will listen into AI conversations to personalize ads, 1 October 2025
Forbes, xAI published hundreds of thousands of Grok chatbot conversations, August 2025
Mozilla Foundation, Meta: help users stop accidentally sharing private AI chats
OpenAI, Data Controls FAQ
Google, Gemini Apps activity
Anthropic, Updates to our consumer terms, 28 August 2025
Stanford, Be careful what you tell your AI chatbot, October 2025
NOS, Privacywaakhond: pas op met gebruik AI-chatbots (Dutch), August 2024
Accountant.nl, AP slaat alarm over datalekken door AI (Dutch), December 2025
Google, Gemini adds Temporary Chats and new personalization features, 2025