Privacy @lemmy.ml will_a113 @lemmy.ml 5d ago

What Kinds of Data do AI Chatbots Collect?

A chart titled "What Kind of Data Do AI Chatbots Collect?" lists and compares seven AI chatbots—Gemini, Claude, CoPilot, Deepseek, ChatGPT, Perplexity, and Grok—based on the types and number of data points they collect as of February 2025. The categories of data include: Contact Info, Location, Contacts, User Content, History, Identifiers, Diagnostics, Usage Data, Purchases, Other Data.

Gemini: Collects all 10 data types; highest total at 22 data points
Claude: Collects 7 types; 13 data points
CoPilot: Collects 7 types; 12 data points
Deepseek: Collects 6 types; 11 data points
ChatGPT: Collects 6 types; 10 data points
Perplexity: Collects 6 types; 10 data points
Grok: Collects 4 types; 7 data points

Privacy @lemmy.dbzer0.com cm0002 @lemmy.world 5d ago

What Kinds of Data do AI Chatbots Collect?

82 comments

Duck.ai
Locally run AI: 0
- Are there tutorials on how to do this? Should it be set up on a server on my local network??? How hard is it to set up? I have so many questions.
  
  I recommend GPT4all if you want run locally on your PC. It is super easy.
  
  If you want to run in a separate server. Ollama + some kind of web UI is the best.
  
  Ollama can also be run locally but IMO it take more learning than GUI app like GPT4all.
  
  https://ollama.ai/, this is what I've been using for over a year now, new models come out regularly and you just "ollama pull <model ID>" and then it's available to run locally. Then you can use docker to run https://www.openwebui.com/ locally, giving it a ChatGPT-style interface (but even better and more configurable and you can run prompts against any number of models you select at once.)
  
  All free and available to everyone.
  
  Check out Ollama, it’s probably the easiest way to get started these days. It provides tooling and an api that different chat frontends can connect to.
  
  I used this a while back, it was pretty straightforward https://github.com/nathanlesage/local-chat
  
  If you want to start playing around immediately, try Alpaca if Linux, LMStudio if Windows. See if it works for you, then move from there.
  
  Alpaca actually runs its own Ollama instance.
- If only my hardware could support it..
  
  I can actually use locally some smaller models on my 2017 laptop (though I have increased the RAM to 16 GB).
  
  You'd be surprised how mich can be done with how little.
  
  It's possible to run local AI on a Raspberry Pi, it's all just a matter of speed and complexity. I run Ollama just fine on the two P-cores of my older i3 laptop. Granted, running it on the CUDA-accelerator (GFX card) on my main rig is beyond faster.
Me when Gemini (aka google) collects more data than anyone else:

Not really shocked, we all know that google sucks
- I would hazard a guess that the only reason those others aren't as high is because they don't have the same access to data. It's not that they don't want to, they simply can't (yet).
  
  Which is good (for now). Glad I don't use that shit
Gemini: "Other Data"

Like, what's fucking left!?
Wow, it’s a whole new level of f*cked up when Zuck collects more data than the Winnie the Pooh (DeepSeek). 😳
- The idea that US apps are somehow better than Chinese apps when it comes to collecting and selling user data is complete utter propaganda.
  
  Don't use either. Until Trump, I still considered CCP spyware more dangerous because they would be collecting info that could be used to blackmail US politicians and businesses. Now, it's a coin flip. In either case, use EU or FOSS apps whenever possible.
DeepSeek at home: None
- How much VRAM does your machine have? Are you using open webui?
- Doesn't the official local app still have telemetry? I might be remembering wrong
  
  You just use the model in an opensource program, not theirs.
Who would have guessed that the advertising company collects a lot of data
- And I can't possibly imagine that Grok actually collects less than ChatGPT.
  
  Data from surfshark aka nordvpn lol. Take it with a few chunks of salt
  
  Skill issue probably. They want to collect more but Musk's shitty hires can't figure it out. /s
Anyone has these data from Mistral, HuggingChat and MetaAI ? Would be nice to add them too

Edit : Leo from brave would be great to compare too
And what about goddamn Mistral?
- Its French as far as I know so at least it abides to gdpr by default.
  
  All services you see above are provided to EU citizens, which is why they also have to abide by GDPR. GDPR does not disallow the gathering of information. Google, for example, is GDPR compliant, yet they are number 1 on that list. That’s why I would like to know if European companies still try to have a business case with personal data or not.
Who TF using Grok.
- Fascists. Why?
- Most of my workforce strangely enough. They claim it’s the best for them in terms of mathematics, but I can’t find that to be a good reason.
I'm interested in seeing how this changes when using duck duck go front end at duck.ai

there's no login and history is stored locally (probably remotely too)
Note this is if you use their apps. Not the api. Not through another app.
- Not that we have any real info about who collects/uses what when you use the API
  
  Yeah we do, they list it in privacy policies. Many of these they can't really collect even if they wanted to
anyone whos competent in the matter: what about the french competition chat.mistral.ai
- +1 for Mistral, they were the first (or one of the first) Apache open source licensed models. I run Mistral-7B and variant fine tunes locally, and they've always been really high quality overall. Mistral-Medium packed a punch (mid-size obviously) but it definitely competes with the big ones at least.
Back in the day, malware makers could only dream of collecting as much data as Gemini does.
I have a bridge to sell you if you think grok is collecting the least amount of info.
- Ikr XD
Almost none of this data is possible to collect when using Tor Browser
- Nope, these services almost always require user login, eventually tied to cell number (ie non disposable) and associate user content and other data points with account. Nonetheless user prompts are always collected. How they're used is a good question.
  
  Use a third party API. Pay with monero.
Or you could use Deepseek's workaround and run it locally. You know, open source and all.
Is there away to fake all the data they try to collect?
- I just came across this article which for people who are into self hosting can take a look and participate. It's basically a tool that generating never ending web pages with non sense that load slow (but not too slow the AI tools move on) to slow down and thus cost them more to scrape the internet if enough people are doing it. You can also hide it in a way that legit user would never see this on your site:
  
  https://arstechnica.com/tech-policy/2025/01/ai-haters-build-tarpits-to-trap-and-trick-ai-scrapers-that-ignore-robots-txt/ https://zadzmo.org/code/nepenthes/
- Pretty sure this is what they scrape from your device if you install their app. I dont know how else they would get access to contacts and location and stuff. So yeah you can just run it on a virtual android device and feed it garbage data, but i assume the app or their backend will detect that and throw out your data.
  
  How about if I only use the web version?
  
  Root, install xprivacy (or xprivacylua if your phone isn't 10 years old).
Am I missing something? What do the numbers mean in relation to the type? Sub types?
- It's labeled "Unique data points". See the number 2 - Usage Data for Gemini, there's an arrow with label there.
  
  Thank you I totally missed that.
- perhaps it's the limit imof each data type?!
  
  gemini harvests only your first four cobtacts, your last two locations, and so on.
  
  how does one defeat that? have fewer than four friends and don't go out!
  
  Ask people for their phone number to add to your contacts and give them your phone for a day
Just clarifying, does this report mean it's collected while a user is using the tools, or data that is generally scraped from the internet?
- They're talking about what is being recorded while the user is using the tools (your prompts, RAG data, etc.)
  
  Does that include generated responses?
Aren't they supposed to collect data?

82 comments