Proton said even seemingly harmless conversations can accumulate into a detailed profile covering a person’s work, relationships, identity, habits, and personal interests.
5 Likes
Personalization and memory is made for this purpose. Why Proton is surprised about it?
Before critizing privacy issues of other AI platforms they should fix their own platform first. It can’t even count properly for example. To test it, give it a task, like creating a background for a character and give 1000 or 2000 character limit. It fails to match the limit all the time and it insists its answer is correct.
1 Like
Proton is not surprised, they document a fact with research standards.
The counting issue has been present in very many models, not only Proton’s. As a pentester often generating word lists, I’ve had this problem with most AI chats for a long time.
ChatGPT found a workaround by prepending the line number to every line, but if you ask “make a file with 765 words beginning with “a”, without numbering”, it still fails (I’ve just tried it and the result had 916 lines).
So blaming Proton specifically on this issue is not fair at all.
5 Likes
Proton uses GLM-5.2 afaik as a base for Lumo. Sure they could try to fix the counting problem but there is only so much that can be done without creating the models themselves and when switching to another model those fixes might not work anymore or might even have negative effects.
Great news and great tool. I casually used both for about a year (along with other unsupported AIs), before Proton switched to a decent model. So I went back to my (essentially abandonned) accounts on Claude and ChatGPT to run the test. Well, even having tried to be careful was clearly not enough…
1 Like
Contrary to that, i was surprised how little Anthropic knows about my devices when i requested a full export.
“Firefox” browser for Tor, IP address, timestamp, “Linux” for web login sessions.
I only use claude cli, had artifacts, global memory, AI training off, and the entire the entire memory and other personization related directories were completely off. They haven’t even fingerprinted my claude cli devices, to my surprise.
Proton should’ve been more honest. Seems like many people assume that OpenAI, Anthropic contracts force you into datamining.