Privacy Talk with Hao Ping(Hank) Lee, PhD student, Human-Computer Interaction Institute at Carnegie Mellon University: How has AI changed the concept of privacy?
“This interview has been recorded on 23th May 2024 and discusses AI and privacy risk”
- How has AI changed the concept of privacy?
- How does AI exacerbate our privacy risk?
- What is the practitioners’ current practices and challenges in considering privacy?
- How has AI changed the concept of privacy?
Hank: Yeah, totally. Yeah, this is basically why I talk to other my colleagues on a day to day basis. Oh, yeah, there’s really happy to talk more about privacy and AI. So I guess I could start this like dialogue in the more you know, in some way that how people normally talk about it AI these days.
Obviously there are more and more technology that are actually used utilizing AI to either deployment systems or actually building features. And you know, for example, we use facial recognition authentic film or we use a logic model to summarize documents, right, you know, write articles or we use like deficient bases with them to create images and arts, and they’re all great right?
But they’re also like, they’re also examples that we’ve found in the real world that how these AI technology actually could introduce additional harm to user for example, rhetoric, facial recognition, you could easily imagine that there’s like third party could get access to the capability that if they have your image or your photo, they basically know who you are right?
Or you know, for large language models there is always the best in terms of how they trained these mode. Do they actually get user content when they’re training, like really personal information It’s like a really unknown question right.r for diffusion-based algorithm.
There are really well known incidents of how the deepfake basically ruins the whole web society or environment right? Yeah, there’s always like, solely incidents that we observe AI systems that have how these assets could actually change about privacy, but in a way that we’re when we really want to boil down in terms of so helpless AI change about privacy is actually a recurring question, right.
So like, there’s tons of that privacy risk already out there before the modern advances of AI. So how does actually AI play a role in it? So obviously, there’s a hype of AI you know, that also made this conversation a little bit harder, because there is always a hype about what AI can and cannot do, which makes the conversation really tricky.
And it turns out that we could always think about a Black Mirror version of AI that could basically ruin the whole humans, human privacy. So I think something is curious that moderate advances in AI definitely change the capability of what computer could do like that’s for sure.
But what is clearly unclear is how these changes are advances in terms of capability, or even the requirement in order to actually unlock those capabilities change about privacy data. That is the part that is unclear, right. So basically what’s the delta here?
Well, we got, we got when we introduced AI into the system. What we do is we try to synthesize our analysis in real world cases rather than a hypothetical.
So we make a decision by looking into the cases that has already been out there. So one of the sources that we routed our analysis is the AI Incident database, which is actually public, and it’s open source and it’s actually collected actions.
It’s called AIID, this is a in that incident database, which you can imagine that a lot of activism or journalism when whenever they spot any type of you know, AI incidents that cause any type of harm, they will put those incidents into a shared document, basically a shared spreadsheet and and so why using as a collective of databases.
And they are all structured editors who actually sort these incidents, and they’ll remove the ones that maybe it’s not, not relevant, but long story short, these are a bunch of different incidents that in some way AI call some type of harms, either privacy or sometimes transparency, or you know, fairness type of things, right.
So what we did is we went through those incidents manually and pulled out those that are privacy relevant, typically, right. So we want to base our analysis based on a real incident which is actually happening in the real world already.
So very briefly, we identify the type of privacy risk on that AI either exacerbates or creates. So what do we mean that exacerbate creators, whether AI in some way fundamentally changes about privacy rights that we found in the incidents.
If that’s the case, we say that, you know, maybe AI creates a new type of privacy risks that we haven’t seen before, or whether you know, AI exacerbates the privacy risks that we’d know about.
We already knew about but in truth, it increases the skill or intensity, right, so that could be the case where AI exacerbates that no risks that we already know about, or otherwise AI could also sometimes not really change privacy at all right.
That’s also the case. But long story short, what we found is that AI does not always lead to new or its other privacy risk, but often does, right specifically put a number here.
- How does AI exacerbate our privacy risk?
Over 90% of the incidents that we’ve looked into, that we end up analyzing in the incident database, actually involve situations where AI exacerbates or creates privacy.
Obviously, I’m not. I definitely don’t have time to go through each individual incident, but we actually put the incident into a website at AI privacy taxonomy.com to encourage folks to actually check it out.
So I guess I could actually provide two examples here. One, so one demonstrate what do we mean by AI exacerbates privacy risks, , and one as the example of AI creates new ones.
So what do we mean exacerbates? So one example is surveillance, right. So surveillance is something that we already know even before modern advances in AI, right?
We talked about technology that is constantly listening to or recording or, basically, like, recording people’s individual activities that we talked about that we said that our business actually serves as surveillance issues.
What AI change about the surveillance issue is that it gave sort of the promise of with AI, we as long as you had enough data to act, a really effective model. So that’s all like the mindset of that.
Yeah, regardless of what we’re doing as long as we’re using AI, we need more data. We just collect more data with a goal that let’s say deploy or build a more effective model, right?
Which is problematic, which is actually pretty problematic, because, you know, we actually see a lot of AI systems that actually collect more data out there than they actually need to, just because there’s a promise of AI could easily sift through this data and make inferences, make reasoning about it very easily.
So let’s look at one example of that. Surveillance is happening there already, but AI kind of is exacerbate in terms of their skill intensity, right. So an example of how AI can create a new type of privacy risk could be phonology and physiognomy.
So physiognomy is a pseudo science that people had, you know, being seen like purports that one could understand one’s personality or mental states only based on their physical appearance like the face or body right.
And what we found is that AI is indeed being used in this way. You know, there are attempts, for example, to try to predict whether a person is gay or straight or whether they are a criminal simply based on their face.
So on the size of really problematic use of AI for us in this case, but there is also risk of that, you know, potentially users or people’s really sensitive information is actually being inferred, or, you know, potentially being shared by AI systems.
So yeah, just like the two examples in terms of what we do we mean by AI exacerbating known privacy risks, and creating new ones.
But as I mentioned, there’s actually more that we actually cover either in the paper and also in the website, but as a starting point, I actually encourage you to check out the website to learn more about it.
Kohei: Thank you for sharing that. There are very few resources that speak about the history of AI and privacy, then your research is very helpful to reach out to how this transition is coming in space and please check out his remarks.
- What is the practitioners’ current practices and challenges in considering privacy?
So as a next topic, you got a very interesting research with more than 30 interviewees, how they make a practice about privacy in this space. So could you share the overview of the research analysts share about any of the output that you receive from the interviewees so far?
Hank: Yeah, sure. So the private project I was talking about, in some way tried to answer this AI trained and learned about privacy and the short answer is probably yes, and, in different ways. Um, and another question that we’re also really interested in as well.
So given that AI in the change about privacy, how do those folks, specifically practitioners who are using an attribute system in some way, approach privacy in their product right, in their AI policy specifically, specifically right.
We talked to 35 industry practitioners who actually work in industry settings, ranging from startup company to like, large cooperation, and the only requirement of you know, in terms of recruitment of during the interviews that they have to be participate or have been participating in a departmental AI products or products that involve AI as a component are crucial to title opponents right.
So specifically, we asked them, you know, how do you define privacy? How do you approach privacy? How do you talk about privacy and addressing it? And also importantly, what if there are any types of tools that you would use when you do that type of privacy work, right?
Thank you for reading and please contact me if you want to join interview together.
Privacy Talk is the global community with diversified expert, and contact me below Linkedin if we can work together!
