I would like some opinions on “normal” AI chats, like ChatGPT, Claude, Grok, Mistral, Deepseek etc.
Which of them strike a good balance in terms of capability (LLM ‘intelligence’ and accurate and up to date web search), a decent free tier that doesn’t ask for money after a few messages, and privacy considerations? For privacy, I was thinking of stuff like that it doesn’t require personal data, chats are not used for personalised ads, no sharing of data with third parties, deleted chats are actually deleted for good, any announcements about TEE use in the future, …
I know that there are some AIs advertising themselves as private, often with technical guarantees such as TEE, but in my experience they are not very good. For example, Duck.ai doesn’t have great models and you in my experience you quickly reach the message limit, Confer.to limits you to very few messages a day and is quite expensive ($35/month) for the paid tier, and Proton Lumo is just plain bad in my experience.
So this is not really about “what AI is the most private”, but more about “among the best and mainstream AI chats, which ones are the least bad in terms of privacy?”. (Note I didn’t even bother to mention Gemini or Meta AI at the beginning.)
Privacy isn’t a factor in any of the “normal AI chatbots”, which I am choosing to define as the investor-funded, closed-source, cloud-hosted options. They offer no privacy, across the board. You’re looking for nuance where there is none
I’m not trying to be extremist, or ride the AI-hate train. They all have a powerful financial incentive to mine every scrap of data you provide, with no meaningful safeguards to prevent them from doing so, and plenty of public evidence confirming that they’ll use whatever they can get their hands on to train their models
Edit: adding some nuance
I’m defining ‘privacy’ as ‘granular control over who can see my data & how it gets used’. From that perspective, I stand by my statement above
We could have a discussion on governing privacy policies, the jurisdictional data laws, PII required to create an account on each platform, etc. I imagine these aren’t literally equal across the board
But I think you could argue in good faith that those factors aren’t quantifying true privacy. These platforms have enough $$$ such that they aren’t actually beholden to laws or policies, for example. Without visibility or accountability, I call them ‘privacy theater’
I’m not really an AI person but in terms of reducing harm to privacy, if you have a decent enough setup you could at least run a Mistral model locally, which based on their privacy policy would be better than using their cloud service (although at least they’re more likely to respect GDPR and such, given that they are located in France). That makes it the least worst choice in my view.
I mean, there are never any guarantees, but their main incentive is to get you to pay for the Premium tier rather than mine your data for personalised ads (with the likely exceptions of Google Gemini and Meta AI). And privacy policies are still meaningful, otherwise PG couldn’t recommend DDG or Brave Search over Google Search or Bing, for example.
But that requires good hardware, being at home next to your desktop PC (or RDPing in), and it probably misses lots of features like web search. It’s the most private option, but also the least “just werks” option.
The consensus here seems to be pointing toward local as the most reliable answer for privacy, which makes sense. If you’re on a Mac and want something ready to use without setting up Ollama or configuring anything yourself, Informity AI might be worth a look. Free, no account, everything runs on your machine. Excuse the self-plug: https://www.informity.ai