The rapid development and spread of AI has affected so many areas of our lives, from how we work to how we learn to the products we’re advertised and content we’re fed on social media. But as AI grows more powerful and pervasive, questions remain about how much trust we can or should place in this technology.
Kate Kaye is a Portland-based independent journalist who spent nearly a year researching and interviewing experts about trust in the age of AI. That work has culminated in “Trust Control,” a seven-part podcast series she created and hosts that explores concepts such as efforts to quantify and measure trust that go back to World War II and still reverberate today. The series also examines trust in the media and the limitations of trying to score or certify the trustworthiness of news content or images shared online. Meanwhile, the rise of AI agents, or bots assigned to do tasks human employees do, present new challenges around trust in the workplace.
Kaye joins us to share more details about “Trust Control.”
Note: The following transcript was transcribed using AI and validated for accuracy, readability and formatting by an OPB volunteer.
Dave Miller: This is Think Out Loud on OPB. I’m Dave Miller. Do you trust me? Should I trust you? Should any of us trust the AI agents that we increasingly have to interact with or perhaps work with? What does it mean to measure trust? And is trust now no longer a feeling? These are some of the big questions that Kate Kaye has been wrestling with. Kaye is a Portland-based independent journalist who spent the last year or so researching and interviewing experts about trust in the age of AI. That work has culminated in “Trust Control,” a new seven-part podcast series. Kate Kaye joins us now to talk about it. Welcome back to Think Out Loud.
Kate Kaye: Thanks so much, Dave.
Miller: Why did you want to spend a year focused on these questions of trust?
Kaye: If you asked me a year ago why I would spend a year doing this, I’d probably be like, what? Really? This was part of a fellowship, an Oregon State University fellowship that was also done in conjunction with Oregon Humanities. And the overarching theme of the fellowship that I applied for to do this work is trust in the age of generative AI. And all the people who were awarded with the fellowship, we all came at it with different ideas and my idea was, well, I know from some separate work that I do, separate research work that I do, that this idea of scoring trustworthiness of media, especially in relation to AI and AI slop and all the problems we have with disinformation was happening.
And I thought, trust scores, that seems… it kind of disturbed me. It seemed like I wanted to learn more about it and I said, OK, here’s this fellowship. It’s about trust in the age of generative AI. I’m going to dig into this idea of scoring and measuring trust and learn about it. And it became this kind of a saga.
Miller: How do you define trust at this point? Do you have a working definition of it?
Kaye: I think the whole time I was doing this work, I kept asking myself, why don’t I have a more solid description or personal definition? And I did grapple with that. Toward the end of the podcast, I talk about how I don’t really have a personal definition. But when I think about trust, I think about my grandmother teaching me how to make bread, and about her showing me how she measures the temperature of the water that you need to use when you’re getting your yeast ready.
Miller: To make sure that it’s warm enough.
Kaye: Yeah.
Miller: It’s a very striking image and sort of personal working definition that you come up with because it is physical. I mean, you actually, theoretically you could get a thermometer and you could say, OK, 82 degrees is right, 105 degrees is right, 50 degrees, right. So it is measurable, but you’re not measuring it. It’s a human, physical, messy thing that was taught to you by someone you love, none of which is true about algorithmic models…
Kaye: Right.
Miller:…and all of which is very different from the way trust is now scored or measured by machines that have a lot of bearing on our lives.
Kaye: Yes.
Miller: What are some of the ways that the trust is currently being scored that are actually affecting us right now?
Kaye: Yeah, well, one way that we’re seeing that was kind of the impetus for this whole project is the idea of using computational cryptographic approaches to measuring and scoring the trustworthiness of digital images, digital videos, how do these things originate? Can we trust where they come from? Can we trust whether or not this photo that’s in a news article is legit, all those sorts of questions. That’s one way. We’re seeing it in relation to measuring how and whether people, sometimes in a work situation, how and whether they trust in the automated or AI equipment or software that they have to use. In the podcast, I talk about a couple specific examples of that too.
Miller: One of the phrases that you encountered over and over, sometimes said by humans, sometimes said by non-humans, is that trust is no longer a feeling. Who in general is saying this and what in general do they mean?
Kaye: The people in general, when I’ve seen it or heard it, tend to be in the business consulting world, sort of people would think of like McKinsey & Co., a company like that, right? Or think of somebody who is an executive at a tech company that does cybersecurity. That’s where I saw it or heard it popping up.
Miller: And what do they mean?
Kaye: Well, it’s interesting because I heard that term or saw it used, often in relation to, it was kind of like a prefix for other things. It would be like, trust is no longer a feeling, it’s a metric. Trust is no longer a feeling, it’s infrastructure. The one place I saw it, they said, it’s a transaction history.
These are all things that computers want to see, right? So trust is no longer a feeling, it goes back to that idea of my grandmother, holding hands underwater with my grandma, us both feeling the temperature of the water. That’s something that is a feeling. And all of these computational ideas, computers need binary information to be able to kind of have an output, right? And to score something and predict something.
And so, this idea that trust is a metric, or it is just a history of transactions that a computer might ingest and say, well, this transaction action history shows us this entity is trustworthy or not, those are computational ideas of trust and I think that these things are very distinct, the human stuff and the computational stuff, and I talk about how these things are being blurred.
Miller: I want to play a clip from one of your episodes. We’re going to hear from a British researcher that you talked to named Bran Knowles.
Bran Knowles: I actually don’t think people are fundamentally, cognitively engaging with trustworthiness in the way that these models imply that they should be to determine for themselves, is it trustworthy? Should I use it? People are just getting on with their lives. They’re being told, well, to compete these days, you have to use ChatGPT because you need to write 10 times faster. Everybody else is using ChatGPT. Is it trustworthy? I don’t know. I don’t care. I’ve got to hurry, I’ve got to use it.
Miller: What does it mean to focus then on questions of trust if in many cases these days, we are not opting in, we’re not voluntarily choosing to use tools, but we’re somehow coerced into it?
Kaye: Yeah. Well, one of the things that Bran, Doctor Knowles, was talking about there and that we talk about in that episode and other episodes is kind of getting closer to this idea that governments and regulators are going to start asking for this stuff. They’re gonna start saying, well, is this trustworthy? Is this system trustworthy? How do we measure whether or not it’s trustworthy? Governments need to do that.
She was talking about how we’re forced to use these things very often. And at work, we might be forced to use a piece of software, or an automated tool of some kind, or my boss might say, hey, we’re using all these AI agents now. We’re spending good money on it. I want you to make sure that you’re delegating work to that AI agent. It doesn’t really matter if I trust it or how my trust in it is measured if I’m forced, if I’m compelled to use it. It’s kind of like a moot point almost.
Miller: You note that there are now a bunch of tech startups that produce agentic trust scores.
Kaye: Yeah.
Miller: What does that mean? What are they doing?
Kaye: Yeah, they’re all these little startups that I kind of encountered when I was doing this research and, for example, what they’re doing is, I talked about this idea that businesses are using AI agents to perform tasks, like they might be using an AI agent that is a procurement agent, and it’s going out and it’s trying to procure widgets from different suppliers. So one agent in that kind of negotiation process will be interacting with others, and when these agents kind of interact and need to like perform tasks, especially in relation with other agents, other computerized bots that are doing other things, they need to gauge whether or not these other systems or bots are trustworthy.
And that’s where that computational trust stuff comes in. That’s where they’re gonna be looking at all of these computational measures and factors and binary kind of information to know. And so, it’s about like, well, should we share the credit card number with this supplier? That’s like the kind of question that an agent would want a trust score to know: do I divulge this information to this other system or not, for example.
Miller: When I was listening to this, it made me wonder. I thought of that famous phrase “turtles all the way down.” I mean, is this AI all the way down? Is it some kind of AI assessing the viability, the credibility, the trustworthiness of some other AI?
Kaye: Yeah. Yes, right, that is what it is.
Miller: Yeah, at a certain point, I mean, it’s such a different understanding of what we once thought about as trust.
Kaye: Exactly, yes, and that’s why, when I first started hearing about this idea of scoring trust, especially in relation to media, which kind of brought me into all of this, I thought, “Wow. How could anybody say they can measure or score trust? What is that?” And so, in the cybersecurity world, in computers, this has been going on for decades now and I talk a lot about the history there.
Miller: There’s another version of this, not scoring the trustworthiness of AI, but various algorithms, various bots or agents scoring human trustworthiness, credit card scores, as a version that’s been around for a while…
Kaye: Sure.
Miller:…hiring decisions now increasingly mediated by AI. War targeting, is this a valid, a good human to kill with this bomb, right? All things that you get into. How do you feel about this version of trust computation?
Kaye: Well, so, I’m not coming at it with an opinion per se, but what I’m doing is asking. I’m interrogating and I’m asking the listener to think about it. And the last episode of the podcast, the last kind of big documentary-style episode is talking about this idea of, especially in the workplace or maybe you’re a soldier on the field and you are interacting with some sort of a computerized system and you have to know… Well, in these kinds of scenarios, especially at work, we are going to be forced to prove whether we’re trustworthy, according to some of the same metrics and measures that are used for computers and bots. And so that’s some of the blurred lines that I’m talking about in the podcast, that it’s not like I’m saying this is bad or good, it’s just… Wow.
Miller: Let’s understand a little bit more what we’re doing. I want to play another clip that gets to what you’re talking about here. This is from a tech startup CEO named Michael Fassett. He helps businesses use AI agents, and this is part of what he told you.
Michael Fassett: I think the new trust model is, trust is verifiable, it’s continuous, it’s machine-readable, and it’s scored. Those things I think we have to have.
Miller: So what does that look like in practice, all those buzzy words.
Kaye: Right. Well, I was talking earlier about this idea of, let’s say a logistics procurement AI agent, a bot that is going out and doing this stuff in an automated way. So one of the things that Michael Fassett talks about is this idea of the “hybrid workforce” and this idea that it’s not gonna be – and even today he purports and others purport – that some companies are thinking of AI agents that they’re using to perform tasks as employees, as workers, they’re not just thinking about them as a piece of software, they’re literally thinking about them as employees.
Miller: As a tool.
Kaye: This puts human beings in the same context as a machine. That more and more because we’re in this environment where we have so much risk because of, I mean, we hear about all these crazy data breaches and like AI agents gone rogue and all of these problems that are already happening. We’re talking about humans having to interact with the same systems, right, that a bot might. And so the humans might be, their trustworthiness might be determined according to the same exact factors that you would require of a bot or an AI agent. That’s what he was talking about.
Miller: I want to go back to the trust in media piece that you mentioned earlier. It’s one of the reasons you started looking into this and this idea that there could now be some kind of digital validation of the authenticity of content, photos or videos or text. You do point out that there are all kinds of potential problems with this in practice. What are some of the big problems?
Kaye: Well, in an episode that I focused on this particular approach, I highlight a really great story about a photographer who used this method, this technical method that employs cryptographic signing of information. What it does is it attaches information about the origin and the changes that might be made to a piece of media, to that media, to that image. She talks about these really beautiful tin-type photos that she used the old tin-type analog process to produce. Then she digitized them. And when she turned them into a digital image, she added this computational trust signal information into it, right?
And there were some problems down the road with the way that method actually works. When the images were floating through the internet, all different types of systems are at play when images are uploaded or shared into different social media platforms or whatever. And what happened was that her photos were deemed by this technical method to be invalid. And her concern was that people would see that – this kind of computational idea of is this trustworthy or not – would see that as meaning that those tin-type photos were not real, that they were AI generated. That’s what her worry was. They weren’t. Nothing changed about the photos. Nothing changed about the original photos. Nothing changed about the digitized versions. What changed was the technical measurement approach, like the technical decisions and the design of that measurement approach is really what changed.
Miller: In the big picture here, as deepfakes and generative AI get more sophisticated, I think it’s already now or it’s gonna be truly impossible for some of us to know what’s real and what’s not. And maybe even more importantly, we just, we won’t know. So whether it’s something that’s trustworthy or not, trust may on some deep level start to be or continue to be eroded. What do you think that’s gonna do to our in-person interactions?
Kaye: I mean, I don’t know, but I worry about it and I sort of like just, I couldn’t help doing this research and thinking about, man, is there like a day in 5, 10, 20 years when there are so many computational kind of appendages that we as humans have or rely on, where we encounter another human and we start to require a computational trust measurement to know whether or not we should talk to them or whether or not we should share information with them or whatever, right?
Miller: That’s darker than I was...
Kaye: I know.
Miller: I thought you were gonna say that we’ll privilege human interaction because it’ll be the last realm that’s unmeeted and sort of trustworthy, at least trustworthy…
Kaye: I think you’re right.
Miller:…in the way that we could once imagine.
Kaye: Yeah.
Miller: Yours is darker, but also, who knows, maybe possible.
Kaye: I think they’re not mutually exclusive.
Miller: Kate, it was a pleasure talking to you. Thank you very much.
Kaye: Thanks.
Miller: Kate Kaye is a Portland-based independent journalist, now the host and researcher and creator of a new podcast. It’s called “Trust Control.”
“Think Out Loud®” broadcasts live at noon every day and rebroadcasts at 8 p.m.
If you’d like to comment on any of the topics in this show or suggest a topic of your own, please get in touch with us on Facebook, send an email to thinkoutloud@opb.org, or you can leave a voicemail for us at 503-293-1983.
