AI and privacy: what happens to your chatbot conversations

When you type a message to an AI assistant, it travels to the provider’s servers, may be kept in your history and, depending on your settings, may be used to train future models. The good news: a few settings and simple habits are enough to stay in control of what you share.

This guide covers the whole picture, from what happens technically to the rights you have. Details on each assistant come from the providers’ official pages as of October 22, 2026, and they change often. This is general information, not legal advice.

The journey of a message you send to an AI

An AI assistant such as ChatGPT, Gemini or Claude almost never runs on your phone. When you tap « Send », here is roughly what happens.

1. The message leaves your device. It is encrypted in transit, then received by the servers of the provider or its cloud host.
2. The model computes a reply. Your text, and usually the whole conversation so far, is passed to the model that writes the answer. Files, photos and voice recordings you attach take the same route.
3. The conversation is saved. It shows up in your history so you can come back to it. That means it is stored somewhere, for a length of time that depends on the provider and your settings.
4. It may be reused. Depending on your plan and settings, the provider may use it to improve its models, detect abuse or personalize future answers.

None of this is unusual: every online service, from email to search engines, works this way. The difference is what people tell an AI. We talk to it like a confidant, sharing details about our health, relationships, job or money that we would never type into a search bar.

Four ways your conversations can be used

To understand what happens to your chats, keep four separate uses in mind. Each one has its own setting.

History

This is the list of your past conversations. It is handy, but it also means your chats stay stored. According to Microsoft, Copilot keeps your history for 18 months by default; according to Google, Gemini activity is auto-deleted after 18 months by default (checked October 22, 2026). You can usually delete conversations one by one or shorten the retention period.

Memory

Some assistants remember things from one conversation to the next: your name, your job, your preferences. This assistant memory makes answers more relevant, but it also builds a small profile of you. You can review, edit or switch it off in the settings. Custom instructions, which you write yourself, work in a similar way.

Training

This is the question that worries people most. Your chats can become training data used to improve future models. In practice, that does not mean your conversation will be copied word for word into someone else’s chat, but it can shape, along with millions of others, how the model responds. With most major providers, this setting is on by default for free and individual plans, and you can turn it off.

Safety and abuse monitoring

Providers also scan conversations to spot misuse. Their privacy policies explain when people may review certain chats and how long these are kept for that purpose. This is one reason a deleted or « temporary » conversation may still exist on the servers for a few days.

Check yourself You switch off training in your assistant. Do your conversations disappear from the servers?

No. Turning off training stops your chats from being used to improve models, but your history stays stored until you delete it. To reduce your footprint, combine this setting with regular history clean-ups or a temporary mode for sensitive questions.

What the main assistants offer

The table below sums up the privacy settings listed on the providers’ official pages as of October 22, 2026. Exact menu names may vary by app and language.

Assistant	Training on your chats	Where to opt out	No-trace mode
ChatGPT (OpenAI)	On by default	Settings, then Data Controls, « Improve the model for everyone »	Temporary chat, not used for training, kept up to 30 days
Claude (Anthropic)	Opt-out on Free, Pro and Max; no training by default on Team and Enterprise	Privacy settings	Incognito mode, not used for training
Gemini (Google)	Tied to the « Keep Activity » setting	Gemini Apps Activity	Activity off: kept 72 hours, not used to improve models
Le Chat, now Vibe (Mistral AI)	On by default for Free, Pro and Student	Privacy section	Not specified; data hosted in the EU by default
Microsoft Copilot	Adjustable	Profile, then Privacy, text and voice training options	History kept 18 months by default
Meta AI	In the EU, Meta AI interactions are used	Meta’s objection form	Not specified
DeepSeek	Adjustable	Data Controls, « Improve the model for everyone »	Data stored in China
Perplexity	« AI data retention » on by default	Account preferences	Incognito mode
Duck.ai (DuckDuckGo)	Never, according to the provider	Not applicable	Chats not stored by DuckDuckGo
Lumo (Proton)	Never, according to the provider	Not applicable	Ghost mode

For a screen-by-screen walkthrough, see our tutorial on ChatGPT, Claude and Gemini privacy settings. To compare other criteria such as price, web search or voice, try the AI assistant comparison tool.

Where your data is stored

Where your data lives affects which laws apply and what recourse you have. Three examples from our assistant profiles:

• In Europe: Mistral AI hosts Le Chat data in the European Union by default, and Proton runs Lumo on servers it controls in Europe, with encrypted chat history.
• In China: according to its privacy policy dated February 10, 2026, DeepSeek collects, processes and stores personal data in the People’s Republic of China. Italy’s data protection authority blocked the app in January 2025.
• In the United States: most large American assistants process some or all data in the US, with data residency options mostly reserved for business plans.

European hosting simplifies some questions, but it is not a guarantee on its own: training, retention and access matter just as much.

What you should never share with an AI

Even with the best settings, the most effective rule is simple: don’t send anything that could hurt you if it leaked. Keep these out of every conversation:

• your credentials: passwords, one-time codes sent by text, card PINs, access keys;
• sensitive numbers: card numbers, full bank details, Social Security or National Insurance numbers, tax IDs, passport numbers;
• other people’s data: details about clients, patients, students or colleagues, without their consent or your employer’s approval;
• work secrets: confidential contracts, customer files, unreleased source code;
• ID documents and sensitive photos, including pictures of your children.

For everything else, one trick works well: anonymize before you send. Replace names with « Mr. X », addresses with « [address] » and exact amounts with rough figures. The AI will help you just as much. For health or money questions, describe the situation without details that identify you.

AI agents and AI browsers: a step further

An AI agent does more than answer: it acts on your behalf. It can read your email, fill in forms, book or buy things. An AI browser sees the pages you visit. These tools therefore see far more than a simple chatbot: your inbox, your accounts, sometimes your payment details.

They also bring a new risk: prompt injection. A booby-trapped web page or email can hide instructions that the agent might follow, such as sending your data to someone else. We explain this in plain English in prompt injection explained for non-techies.

Before giving an agent access to your email or bank, read our guide on the risks of handing your accounts to an AI agent. Our Can I hand this to an agent? tool also helps you weigh up a task before delegating it.

Your data can fuel scams

What you post online, and what leaks in data breaches, also helps fraudsters. A video with your voice, family photos, your employer and your travel plans make it easy to craft very convincing messages, or even to imitate a loved one’s voice. Protecting your privacy also makes scammers’ lives harder. Our guide to AI scams and how to protect yourself covers the right reflexes.

Children and teenagers

Young people readily share personal details with a chatbot that talks to them kindly. Providers set minimum ages, which vary by service and country, and some have restricted features for minors. Talk to your kids about it: what you write to an AI is not a private diary. Our guide on kids and AI: minimum ages, parental controls and tips goes service by service.

Your rights under data protection law

In the UK and the European Union, data protection law (the UK GDPR and the EU GDPR) applies to AI assistants as soon as they process information about you, even when the provider is based elsewhere. It gives you, among other things:

• a right of access: to ask what data the provider holds about you;
• a right to erasure and to rectification, under certain conditions;
• a right to object, which several providers implement as a setting or a form to refuse training.

Providers explain how to exercise these rights in their privacy policies, often through a form or a dedicated email address. If you don’t get a satisfactory answer, you can turn to the data protection authority. In the UK, the Information Commissioner’s Office explains your data rights and how to make a complaint on its page for the public (checked October 22, 2026). In the United States, there is no single national privacy law of this kind; protections depend on your state and on the provider’s own commitments.

Choosing a more discreet setup

If privacy matters a lot to you, three options exist, each with limits.

• An assistant built for privacy. According to DuckDuckGo, Duck.ai strips your IP address before sending the request to the model provider and does not store conversations. Lumo says it never trains models on your chats. In return, agent features and long-term memory are more limited.
• A business plan. Team and enterprise plans generally exclude training by default and come with a contract. That is the right path for work data: see our guide to keeping company data confidential with AI.
• AI on your own computer. A locally installed model sends nothing to a server. It needs a reasonably powerful machine and a bit of setup: everything is explained in running AI locally.

A five-minute privacy routine

Here is a simple routine to run once, then again whenever you switch assistant or plan.

1. Open the settings of your main assistant and look for Privacy or Data Controls.
2. Decide on training: turn it off if you don’t want your chats used to improve models.
3. Check the memory: see what the assistant remembers about you, delete what you don’t want kept, or switch it off.
4. Tidy up your history: delete old sensitive chats and shorten the retention period if the option exists.
5. Use temporary mode for delicate questions about health, money or family, on top of anonymizing.