This post continues my 2020 article on privacy. That one was about our clicks; this one is about our thoughts.
In 2020 I wrote that if you spend 40% of your day at a computer, you experience 40% of your thoughts being watched. I’m afraid that number is rising steeply. We have such intimate conversations with chatbots that they give a full view into our minds; big companies store them, and sooner or later they will use them to manipulate us.
Why is this different from search?
We discuss things with AI that we would never type into a search engine, and often wouldn’t tell another person. Health, relationships, money, conflicts at work, raising children. The assistant is patient, doesn’t judge, has time for you at two in the morning, and on top of that calls our insights brilliant. That’s exactly why we open up to it.
These systems increasingly remember, too. Since 2025, ChatGPT can draw on all your previous conversations to give more personalised answers. Sam Altman, the head of OpenAI, put it this way: it points towards AI systems that get to know you over your life. It really is convenient. But think about it: this is exactly the profile the advertising industry used to piece together from clicks. Only now you write it yourself, in your own words.
In part one I wrote that Cambridge Analytica worked by making sure every target group got a solution to its own fears in the ads. To do that, they had to work out from clicks and likes what you were afraid of. A conversation partner who knows your train of thought doesn’t have to guess. And it doesn’t try with an ad – it answers, argues, and speaks to you in exactly the tone that works on you. This is a new level of mapping how we think. And it’s also an opportunity to influence, on a new scale, how we think and what we believe. For someone who uses these tools a lot, I don’t think anyone or anything else can have such an impact. I’m not saying they abuse it, but the power is there.
What happens to what you type?
This varies by provider and plan, and it changes often, so it’s worth checking from time to time. A few things that happened in 2025–2026:
- Training by default. On most free and individual plans, your conversations may be used to improve the models unless you turn this off. Anthropic, for example, introduced in 2025 that conversations with Claude may be used for training unless the user opts out, in which case they can be kept for up to five years.
- People may read them. Google shows some Gemini conversations to human reviewers. These aren’t deleted even if you delete your history: they’re kept for up to three years, though disconnected from your account.
- Advertising. Since December 2025, Meta has been using conversations with Meta AI to personalise ads on Facebook and Instagram, with no way to opt out. The EU is excluded for now, thanks to the GDPR I already praised in part one. And in early 2026 OpenAI started testing ads in ChatGPT’s free and low-cost plans. They promise not to share conversations with advertisers. I believe them. But in part one I also wrote that the next management or shareholder pressure can rewrite everything.
- Accidental publicity. In the summer of 2025 it turned out that some conversations shared in ChatGPT had appeared in Google search results, including job applications, relationship advice and conversations about mental health. OpenAI quickly removed the feature, but it shows how easily things can slip out of our hands.
One more important thing: an AI assistant has no doctor–patient or attorney–client confidentiality. What you type sits on a company’s servers and can even be requested in legal proceedings.
So should we stop using it?
I use it too, including for work, and it’s very useful. That’s not what I’m saying. What I am saying is that it’s worth using as consciously as a search engine or a browser. A few things I do:
- Turn off training. You’ll find it in the settings of all three big providers: in ChatGPT under Settings → Data controls → “Improve the model for everyone”, in Claude under Settings → Privacy → “Help improve Claude”, and in Gemini with the “Gemini Apps Activity” / “Keep Activity” switch.
- Review the memory. If the assistant remembers you, check what it has noted. It can be surprising. If you don’t need it, turn it off.
- Use temporary chats. Most providers have a mode that doesn’t go into your history or the memory. Use it for sensitive topics.
- Don’t type anything you wouldn’t write down for a stranger. Full name, home address, health data, other people’s personal matters, company secrets. If you need a text reviewed, take the names out first.
- Use a business plan for work. Business plans usually have stricter data handling and don’t train on conversations by default. If everyone at your company pastes client data into their own free account, now is the time to deal with it.
- Run AI on your own computer. These days a smaller language model runs on an ordinary laptop, for example with Ollama or LM Studio. It’s not as smart as the big ones, but it’s plenty for rewording an email or summarising a note, and nothing leaves your computer. I’m experimenting with this myself. And these are improving fast too.
- For truly private matters, use Duck.ai, even from Tor Browser. You can chat with several popular models without signing up – just be careful not to dox yourself.
- Choose a provider built on privacy. Lumo is the assistant from Proton (the makers of Proton Mail and Proton VPN): it doesn’t keep logs, doesn’t train on your conversations, stores them encrypted, and runs in Europe. You can try it without signing up. France’s Mistral is also a European alternative, but on the free plan turn off training in the settings first.
Finally
The question is the same as in 2020: it’s not whether you have something to hide, but who ends up holding this much power. Only now they’re no longer collecting what you click on, but how you arrive at a decision. I think that’s worth five minutes of going through your settings.
If you have questions, or need help with how your company can use AI safely, write to me. There are no stupid questions.