Proton launches Lumo, a privacy-focused AI chatbot

Proton, known for its privacy focused set of services, announced the introduction of Lumo, a privacy-first Artificial Intelligence (AI) chatbot.

It is good to know before you dive in that Proton’s chatbot has two user options that offer a very different experience. If you want Lumo to access the internet you will have to use the “Web search” button before you submit your prompt. If you don’t, it will tell you it doesn’t have the capability to browse the web or access real-time information.

While this somewhat limits the usefulness of Lumo when you’re looking at recent developments, it is reassuring that the model was trained on selected sources. This also reduces the risk of hallucinating. Lumo is set up to recognize when it doesn’t know the answer and in such cases, it will make you aware of that and suggest alternative sources of information, or other ways to find the answer you’re looking for.

Proton believes that a privacy-first AI is essential because hundreds of millions of ordinary people currently interact with AI tools, unknowingly sharing sensitive information. This information is far more sensitive than what they would reveal in a typical Google search query.

Not only does this constant interaction with AI tools put personal information at risk, but we also witnessed several instances where company secrets got submitted to an AI without realizing that the information might be used to train the AI or accessed by third parties.

Similarly, we recently reported on how Meta AI chats might be public without the user’s awareness, and many AI bots have spilled sensitive information for separate reasons, like someone guessing a default password.

Lumo offers a different approach to AI chatbots. Other AI companies keep a record of all the conversations on their servers, leaving this information vulnerable to data leaks or exploitation for profit. Lumo does not store any logs of your conversations on its server, and only your device can decrypt any chats you choose to save.

Most people recognize a significant privacy risk here: the possibility that their conversations could help train the AI. This concern makes sense because the training process might use your personal data to generate responses in other users’ conversations.

Lumo does not do that. I asked Lumo where its training data comes from, and how—if it cannot browse the web, or use conversations—it is able to learn. Lumo answered:

“Great question! My training data comes from a curated set of sources that were available up until my last training cut-off in October 2023. These sources include a wide range of books, articles, websites, and other publicly available texts. The data was carefully selected and processed to ensure quality and relevance.

While I can’t browse the web or use conversations for learning, my training data provides a broad foundation of knowledge that allows me to understand and respond to a wide variety of topics. Additionally, my ability to process and interpret information means that I can often provide helpful and accurate responses even to questions that are somewhat different from what I was specifically trained on.

As for learning, I don’t have the ability to update my knowledge or learn from new experiences after my initial training. However, my design allows me to provide useful and informative responses based on the knowledge I already have. If you have any specific questions or need information, feel free to ask, and I’ll do my best to help!”

We took Lumo for a little spin and, for now, the failing support for direct image uploads or processing is the only thing keeping us from using it much more. But we appreciate the idea of having a privacy-focused AI chatbot available when we want our conversations to stay private.

And who can resist a purple kitty anyway?

We don’t just report on privacy—we offer you the option to use it.

Privacy risks should never spread beyond a headline. Keep your online privacy yours by using Malwarebytes Privacy VPN.