Proton AI Paper Trail Shows How Much Chatbots Know, TechRadar Review Finds

The tool was built by privacy software engineers at Proton and runs on Lumo, the company's privacy-focused large language model. Proton says Lumo deletes user data immediately and does not retain it, TechRadar reported. The reviewer uploaded a ChatGPT export to test what the chatbot had learned.

The resulting report sorted the reporter into categories. A "Financial Background" section contained sensitive information that the reporter said was never explicitly mentioned. A "Health Limitations" category tracked health changes over months, including injuries and facilities the reporter had turned to for help. A "Family" section was fabricated, showing that LLMs can still hallucinate, though TechRadar noted the attempt was real and could become sharper with more data and time.

The report rated the reporter "easy" to profile and assigned an "ad value," a four-figure dollar amount, along with a list of sectors that would bid for the reporter's attention. TechRadar said that only a few years ago building such profiles would have required extensive research; now it takes only an upload.

Much of the data users share comes from accumulated context they build while using chatbots, according to Eamonn Maguire, Proton's Director of Engineering. "Even when a user never disclosed certain facts directly, these can often be estimated from patterns in their interaction — approximate age, income bracket, education level, or political leanings," Maguire told TechRadar. Writing style and tone can also reveal socioeconomic profiles and psychological states, including stress levels, mood, and moments of vulnerability, the article said.

Once personal data leaves a platform, TechRadar reported, it risks being permanently compromised if it is combined with other datasets for advertising, training, or sharing beyond what is necessary for the basic functioning of LLMs. That leaves users with few means of protecting it, the review said. The reporter's data carried a specific and substantial price tag with intrinsic value that companies can profit from.

Incogni's data privacy ranking 2026 scored 13 AI platforms based on privacy risks. Bogan Popescu, senior communications manager at Incogni, told TechRadar: "It’s less a case of which is the most privacy-friendly platform but rather which is the least privacy-invasive platform." Meta AI, Gemini, and Copilot were the worst, according to Incogni, which pointed to sprawling privacy policies covering many products at once and more extensive data collection or sharing practices. Others, such as Mistral's Vibe, accumulated fewer risks overall but still train on publicly available datasets without disclosing exactly where those datasets come from. "They tell users very little about what information was actually included, whose data may be in there, and how it was obtained," Popescu told TechRadar.

ChatGPT followed Vibe as "the least privacy-invasive platform" in the Incogni ranking, TechRadar reported. The review added that last week's 404Media investigation revealed humans are reading users' prompts to improve OpenAI's models. The companies were contacted but had not responded to requests for comment on their privacy policies, according to TechRadar.

The article also quoted DuckDuckGo Senior Vice President Zac Pappis: "Many companies have given the impression that they should be using AI for everything." TechRadar wrote that rather than LLM providers harvesting data, users are voluntarily sacrificing their privacy.

Editor's Summary Proton's AI Paper Trail tool analyzes ChatGPT and Claude exports to show how much personal data users reveal, and a TechRadar test found detailed financial, health, and psychological inferences. Incogni's ranking placed Meta AI, Gemini, and Copilot as the worst platforms for privacy risks, while ChatGPT was less invasive but still under scrutiny after a 404Media report on human review of prompts. The review highlights that users often give up privacy voluntarily through accumulated chatbot context.