"We're pouring our secrets into chatbots we don't trust"

Late at night, when someone needs to vent about a relationship or quietly check a worrying symptom, a rapidly growing number of people turn not to a friend but to a chatbot. A new survey from Proton, the Swiss privacy firm behind Proton Mail and Proton VPN, puts a number on it: 66% of British AI users have discussed at least one sensitive topic with an artificial intelligence assistant. Personal finance leads the list (33%), followed by work and career problems (32%), mental health struggles (27%), and relationship issues (20%).

The striking part is the contradiction underneath. The same people pouring out their most intimate thoughts say they do not actually trust the machines holding them. Some 42% report little or no trust in AI companies to protect their private information, and 11% say they do not trust them at all. Yet 55% admit AI has fundamentally changed their everyday decision-making. We are, in effect, handing over our inner lives to services we openly distrust.

The reason, the research suggests, is freedom from judgment. Friends and family remain the preferred confidants for the most delicate subjects, but a chatbot is available at 3 a.m., never looks surprised, and never blushes. It is a 24/7 listener with no emotional reaction — and that neutrality turns out to be exactly what makes disclosure easy.

This is a textbook case of what researchers call the privacy paradox, the gap between what people say they care about and what they actually do. But generative AI deepens it in a way social media never quite did. A search query is a handful of keywords; a therapy-adjacent chat is a full emotional biography. As Proton's Director of AI Engineering, Eamonn Maguire, put it, "AI is learning far more about us than a search engine ever could." The intimacy of the medium raises the stakes well beyond the old privacy debates.

There is also a hard business logic at work. A chatbot that learns from your conversations to get smarter is a chatbot whose economics depend on your disclosure. Maguire frames it bluntly: people are "handing AI companies their inner lives and biggest secrets on the assumption that they will protect them, when it's actually the opposite." The moment a model is trained on your confessions, your privacy isn't a bug being managed — it's the input.

The encouraging part of the survey is what people say they would change. A full 82% said they would be more likely to use an AI chatbot specifically designed around privacy, and 44% would share far more sensitive information if they had a cast-iron guarantee their chats would never be used for training. That reads as a genuine market signal, not just an abstract complaint.

Proton's answer is Lumo, an open-source, privacy-first ChatGPT alternative built with the same zero-access encryption the company uses for its email. The firm also recently shipped AI Paper Trail, a tool that visually demonstrates how much personal information an average AI conversation leaks about a user's inner life — a clever way to make an invisible risk tangible.

Building privacy into a chatbot is not a checkbox; it changes the architecture. End-to-end encryption means the model provider cannot see your conversation history, which in turn constrains what the assistant can remember across sessions and how much it can personalize. That is a real product tradeoff, not a marketing line. Proton's security model documentation walks through how zero-access encryption can still wrap an LLM's responses while keeping the provider out of the loop — a genuinely different technical approach from the default "log everything and train on it" model.

The unresolved question is whether people will actually pay for the privacy they say they want. A zero-access chatbot has to make money from subscriptions, not from training on your data. That flips the bargain from "free, but you're the product" to "paid, but you're the customer." The survey says the appetite exists; whether it survives the moment a credit card is involved is the real test.

Stepping back, the story isn't really about one chatbot. It is about a public that has already decided AI is too useful to give up, while remaining aware it doesn't trust the bargain it has struck. That tension — usefulness versus trust — is likely to define the next phase of consumer AI, and privacy-first tools are the first serious attempt to resolve it rather than ignore it.

Further reading:

• Proton's announcement of Lumo
• How Proton's Lumo security model works
• The privacy paradox in academic research