People already tell AI chatbots things they would hesitate to tell a colleague: money worries at 2 a.m., an embarrassing symptom, how to ask someone out. What they do not do is trust the companies on the other end. That is the headline of a survey from Proton, which asked 4,014 current chatbot users in France, Germany, the UK and the US what they share and why.
What the survey found
About two-thirds of respondents, 66%, had discussed at least one sensitive topic with a chatbot. The eight topics covered personal finances, relationships, mental health, sex, secrets, romantic attraction, work and family conflict. Americans reported the highest rate at 68.4%, and French respondents the lowest at 60.3%.
Trust did not keep pace. Only between 11% and 20% of respondents said they highly trust AI companies with their private information, and 42% said they have little or no trust. Distrust was strongest in France at 47.7% and lowest in the US at 38.8%.
Why people confide in AI anyway
Money, mental health and career problems were the three most common sensitive topics. Mental health ranked first in Germany, career problems in the US, and personal finances in France and the UK.
The reasons are practical. Convenience and freedom from judgment were the most cited reasons for choosing a chatbot over a person, while privacy ranked lower. Convenience won for work and money questions, and judgment-free advice mattered more for personal subjects such as sexual health.
AI has not replaced human support. Friends and family stayed the first choice for every sensitive topic, ahead of chatbots, therapists and religious advisors. The gap was narrowest for money: in France, AI came within 2 percentage points of friends and family.
What worries people most
Respondents were more concerned about their conversations being used to build advertising profiles than about their use in training AI models. The gap was widest in the US, at 65.3% versus 57.6%. Proton reads this as a sign that people worry less about chats being saved than about them becoming a commercial asset.
The finding lands at a timely moment. OpenAI said this week it will test labeled visual ads in ChatGPT, and states that ads do not influence the answers the chatbot gives.
What would change behavior
A guarantee that conversations would not be used to train models would make 44% of respondents more willing to share sensitive information, according to the report’s detailed results. That was highest in the US at 47.7% and lowest in France at 40.7%. A summary list on the same page gives 48%, so treat the figure as roughly four in ten to nearly half.
Interest in a chatbot designed around privacy was higher still. More than 70% of respondents said they would be more likely to use one, from 71.1% in Germany to 82% in the UK.
What the survey cannot tell you
Proton sells a private chatbot, so it has a commercial interest in these results. The company says respondents were not told who commissioned the survey, which was meant to limit bias from brand reputation.
The data is self-reported. It does not observe what people actually typed into chatbots, and it covers only current chatbot users in four countries, so it should not be read as representative of other regions or of people who avoid AI. It is a snapshot of attitudes, not a measurement of behavior.
What you can do today
The simplest protection is free. Most major assistants let you switch off the use of your chats for training, and most offer a temporary or incognito mode for one-off sensitive questions. Beyond settings, a good habit is to keep identity documents, account numbers and medical details out of any chatbot you do not fully trust.
A private alternative: Lumo
If you want a different approach, Proton makes Lumo, its own AI assistant, built around the idea that your conversations should stay yours. Proton says Lumo uses zero-access encryption, keeps no logs, does not share data and does not train on conversations. The code is open source, so anyone can inspect how it works instead of taking the claims on trust.
The privacy design covers more than text. According to Proton’s description of Lumo 2.0, it runs on Proton’s European infrastructure under Swiss privacy law, and the same protection applies to images you upload and images Lumo generates. Proton uses the same zero-access architecture that secures its Mail and Drive products, so the approach is one the company has applied for years.
Pricing is straightforward. The free plan offers core capabilities, which makes it easy to judge the quality of the answers before spending anything. Lumo Plus adds unlimited chats, Projects, advanced image generation and priority access to the fastest models, and Proton lists it at €12.99 per month or €9.99 per month on a one-year plan. There is also a Lumo Professional plan for teams. Prices can change, so check the current offer.
It fits the situations the survey describes best: personal questions about money, health or relationships, where privacy matters more than squeezing out the last bit of model performance. An independent review on PrivacyTools calls it the most credible privacy story among hosted AI assistants, while noting that your prompts still go to a server and that the free tier is limited for heavy use. Those are fair limits to know before you start.
You can try it through this Proton link.
The takeaway
The survey does not show that people are careless. It shows that convenience gets people to start talking to AI and that trust decides how far they go. Providers that can prove their privacy claims, not just state them, will be the ones people confide in.
