Was using my SO’s laptop, I had been talking (not searching, or otherwise typing) about some VPN solutions for my homelab, and had the curiosity to use the new big copilot button and ask what it can do. The beginning of this context was actually me asking if it can turn off my computer for me (it cannot) and I ask this.
Very unnerved, I hate to be so paranoid to think that it actually picked up on the context of me talking, but again: SO’s laptop, so none of my technical search history to pull off of.
Is it possible that your chain of questions is very similar to other “paranoid” users who inevitably question copilot about privacy, so this is a learned response?
I’ll pull the rest of the context when she’s back in town, I doubt she’s used it more so it should be saved still. She looked at me when this typed out and said “you’re fucking with me right?”. I am still just as shocked, I wish I was fucking around and I have no other explaination how it would remotely key onto saying this given the previous interactions.
Absolutely amazing.
My guess is that at this point there are so many user prompts its received so far in its training set that bring up both Copilot and privacy concerns that it first interpreted the question, then searched for the most common topic associated with itself (privacy), then spit out a hardcoded MSFT override response for ‘inquiry’ + ‘privacy’.
I want to believe that is the explaination, I really would’ve expected at least a hardcoded “features and capability” response, or for it to be more than a neutered chatGPT that im sure neither of us are going to use
Occam’s razor dictates that it’s just overly permissive settings by default and an owner who doesn’t know how to turn off mic access
ChatGPT has a short but distinct history of encouraging paranoia in people who use it.
Asked for help with a coding issue, ChatGPT wrote a long, rambling and largely nonsensical answer that included the phrase “Let’s keep the line as if AI in the room”.
I have to credit to the novelty of the technology, there’s certainly a reason I’m wanting to self host models, my concern really is with what data is being used, and how these models are being trusted.
My goal is to contribute the least useable data to the likes of OpenAI “in the puruit of AGI” because it will inevitably become as did MS Tay did, especially if something can change on their end and suddenly have it spitting out garbage for users who may be potentially at risk of bad advice or actually paranoid.
That also doesn’t mean I havent and wont use chatGPT, it certainly has been a useful tool, knowing its limitations, but OpenAI has their head in the clouds and it only leads to greed in pursuit of an end goal. /Imho
I think AI is humanized and otherwise designed so that people will feel encouraged to give private data to it. The Kagi Corporation wrote about this in their manifesto. In reality, giving your data to open AI is just as unsafe as typing in a personal search query into Google or Bing. But by changing the context, it feels like you’re talking to a friend or a person you met at a bus stop.
AI Bros always say “it’s just a tool” as a sort of thought terminating cliche (note: this wasn’t intended to be a dig at your comment). Guns are a tool too. I wouldn’t want the richest corporations in the United States to personally own the most powerful missile systems, and in terms of AI, that’s kind of where we are.
Brilliant! I love your closing point about tools and missile systems
What. The. Fuck. Please tell me this is satire. Please.
It’s not real. There’s other ways that answer would come up. We only see the ones that seem weird because boring answers don’t get posted.
There’s a real risk of survivorship bias here. Somebody asking about a car gets that and thinks nothing of it. A privacy minded person, however, would find it odd. And being the kind of person concerned about what could have been the cause considered the prior conversation.
I’m not saying its an unreasonable concern or technically not feasible. It’s just not how the LLM’s tend to work.
Id consider it more likely to be a bug, or general inquiries like you said, or that SO had a bunch of documents locally that reference privacy or browsing history (anytime really) that MS could have used as a kind of “here’s more about the person asking you a question”
I will tomorrow, I understand where the skeptisism comes from, I still very much doubt that it is listening, I do have my Firefox account on her laptop, but regardless it leaves a nasty taste in my mouth
Edit: this is no more than about 6 messages into using it, first few were garbage my SO tried out, then I was curious of its actual utility, not really coming at it to find a problem
Here is an alternative Piped link(s):
https://piped.video/X4HQyqc-aVU?si=T1s6HJ6QwacPOOeR
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source; check me out at GitHub.
Fuck a Repo Man reference in the wild, nice.
“The more you drive, the less intelligent you are.”
I doubt that it’s sending audio data back to Microsoft although it probably does have access to your search history if you’ve used bing / the inbuilt search bar.
If its anything like Cortana’s permission it’ll have access to all your web searches. Cortana also had speech and typing personalization, so Microsoft is definitely giving copilot at least those permissions.
I will post the full context tomorrow when I can use the laptop again. No previous chats had anything to do with privacy and this was the first chat since the update. The first chat was something like “shit fart” that my SO had scientifically gauged the model with
If the response is not related to listening in on your convo then it smacks of a buddy processing a personal insecurity.
Actually my last girlfriend said I was “nicely accommodable.”
It’s a LLM. You asked it “what can you even do” and one of the most hot topics with AI is privacy concerns. With Copilot being neutered by MSFT to produce curated responses asking it what it can do, and it branching to privacy concerns first, seems totally reasonable
