HCAI 9 - Human-Centered AI-mediated Communication with Mor Naaman

Speaker A: Foreign.

Speaker B: Welcome Moor Naman to the Human Centered Centered AI podcast. We're thrilled to have you here. You're a professor at Cornell and you do plenty of things in AI and information science and so on.

Speaker A: Yeah, so I'm a professor of uh, information science at uh, the Cornell Tech campus of Cornel. So Cornell Tech is uh, a uh, New York City uh, graduate only campus that's focused on technology and related fields. I also happen to serve here as the associate Dean for faculty affairs, but hopefully we're not gonna talk about that today. Uh, my research uh, recently has indeed been around AI and what we call AI mediated communication. It's a, uh, field like all the research projects that we have that uh, we stumbled into and maybe we'll have a chance to talk about how and over the years, but started looking at from uh, 2018 or so, uh, even before all these uh, new LLMs came to life, generally interested in the trustworthiness of our information environment and our information ecosystem. And that has been my research even before that. And kind of that's how we rolled into the AI stuff, looking at trust and trustworthiness, even though the umbrella has broadened uh, since we, we uh, uh, I guess backing up originally a computer science PhD, I guess I don't say pre AI, but pre what now we call AI. Obviously, uh, we had the other AI framing uh, back then. And so I got my computer science PhD, but my academic appointments were mostly in information science. So kind of transformed to look at more of the human and social aspects of the technology over, over the years. Yeah, that's, that's pretty much it.

Speaker B: All right, thank you. So in this podcast we talk about human centered AI. And one of the things that I wanted to ask, just to start off a, ah, couple of years ago I saw you give a keynote at Rexis on AI, uh, mediated communication. And I remember you started, you know, you were talking for a few minutes, showing a few slides and then you sort of took a break and said like everything you've heard so far was generated by AI, which was kind of interesting because it didn't really sound AI generated. It was really on topic. It was super relevant to the audience. And this is before ChatGPT, right? That was quite a big thing to see. And I also understood then that you had been working on this for, for quite some time. So in the last two years, how do you think AI mediated communication has changed with all of these large language models and GPTs?

Speaker A: Yeah, I guess the first thing that changed is that my little trick became cliche very quickly. I think it was three months after that that they launched ChatGPT. And it was widely known and widely available. And it was surprising that the directis I was able to memorably, I think, shock, uh, a bunch of computer scientists, people with a lot of experience at tech and did not consider these capabilities, uh, yet. I think some people that knew me were kind of looking at me weird when I started because they felt it wasn't my style, uh, but others obviously had no way to know and I found it really, uh, surprising. So I think one thing that has changed, well, both the capabilities dramatically improved since my uh, talk then, but obviously the awareness has improved. The integration into everyday product and services, of course exponentially changed where we now see it everywhere. I think back then we had the smart replies from Google, which were a great use case for kind of the same, similar kind of models that are used today, uh, similar kind of applications as I see it from the AI communication point of view. And the models have improved a lot. Right? So not only in text, but creating code and images and increasing the videos, or at least video segments that uh, are expressive and realistic. On the other hand, what has not improved? I think the topic of this podcast of the human centered aspects of these technologies, I think some of the things that we'll probably talk about today have not, uh, I haven't seen them really advance. I've seen a lot of challenges then that some of them highlighted in my talk, some of them we looked at since. And I just don't, I didn't see any progress addressing those questions and challenges. Definitely not in the products we see today in the market. A little bit in the research. I think more researchers are paying attention to this, but not so much that I heard in companies from my discussions with them. Maybe, uh, a very simple example. In the first paper we worked on the question of trustworthiness, the idea that uh, you'll be suspected when people look at text and they suspect that AI may have written that text, the evaluations of the person who wrote the text will be diminished. People will think they're less trustworthy. And we saw in a number of papers we showed that multiple types of evaluations might change as soon as somebody suspects that you're using or used AI to create, uh, whatever content. The first paper here is uh, the origin story, the first wafer. We were working on Airbnb and the Sharing Economy and trust in the Sharing Economy and I think, think it was 2017 or so. And my student Xiao, we were analyzing profile text and My student Xiao came into the room one day and said, you know, what would change if AI wrote all this text for the people? And so we started looking at it and I exactly asked that question, right? What happened to trust when people think AI may have been involved? Surprisingly, even in 2017 or 18 when we actually started, uh, executing it, people were ready to believe that AI could, is able to create the content. We completely faked it in the first couple of studies. We use human written profiles and we just told evaluators that were written by AI or may have written by AI. But one year later we already had the early GPT models and we didn't have to fake it anymore. We could start doing, uh, it for real and then, you know, where we are today.

Speaker C: So is that when you started seeing like, going from. Oh yeah, it's not going to be possible to. Okay, it's probably going to be possible. Like, Was that when GPT2, GPT3 came?

Speaker A: Yeah, GPT2 I think, was the first one where we said, wow, this is going to be possible. And frankly, I still didn't expect because GPT2 was also a little clunky. There was a lot of repetition and a lot of awkwardness and obviously not perfect. Um, well, it's not perfect today, but it wasn't as sophisticated as it may be today. But the trajectory definitely started looking up and it was clear that we landed at least on a research topic that may actually be reasonably, uh, useful.

Speaker C: So I kind of got curious because you mentioned capabilities. And so how would you describe the capabilities that we have now or that these exhibit?

Speaker A: Well, that's an interesting question. I think what shocked me is not just the ability to use a common word from a famous paper to parrot human activity, but actually to create, to have creativity and some intelligence. That's also an open debate. Right. But to show what those models are capable of in terms of, uh, creativity, in terms of creating new ideas, new thoughts and new directions and solve some problems. Right. Not all human problems. I think that was the most surprising part to me. Yeah, making, spinning out text not surprising. But the fact that this simple task of predicting the next word can create such, uh, level of seeming intelligence, uh, uh, but also novelty in a way that can impact all of our. I'm sure you use it for research. Uh, I use it in research, uh, anything from asking it methodological questions, obviously verifying the responses, but getting ideas from it or asking it to name the new phenomena we just found and get ideas back and iterating on them. It's Been an amazing tool. Suspected.

Speaker B: So when you started working on AI mediated communication, did you think that you would actually be communicating with the AI instead? So it's not longer AI mediated, it's communicating with the AI, right?

Speaker A: Yeah, that's true. And I think that's a very interesting area. We actually in our research stayed focused on the idea of AI enabling communication between humans. I think it's important, I think there are huge important issues that has to do with our communicating just with AI and what we can get from it, how we trust data, emotional dependence, as has been in the news this week, when we communicate with AI bots or AI agent or AI models. And I think this will increasingly be part of our, uh, will have to be part of our lives as, as these agents become more and more realistic, both in appearance and in communication style personalities and all the different adaptations that we can have for them. But we have stayed. I think to me the most interesting part has been how the use of these AI tools have impacted how we evaluate and will continue to impact how we evaluate each other. So it's not just text. We started with text, uh, and just worthiness of uh, text based evaluations. But increasingly though, these tools will be embedded and not just LLM based tools, any AI tools will be embedded in technologies that we use every day. So we started with a smart reply and three words and text and we'll end up with zoom. Not doing it yet, but potentially integrating, I don't know, accent shift to Zoom conversations. And not only they can shift my accent right now to sound, I don't know, less Israeli or more British or whatever accent the listener, uh, accustomed to. Um, these are technologies that are clearly within reach and will change how we evaluate each other, how we see each other, um, in all walks of life. So uh, we stayed, uh, not solely, but mostly focused on these questions that I think are super interesting.

Speaker C: Do you have any answers to this? Like, so these technologies we will get there, right? Do you, do you have the solution for how that should be done in a done human centric way?

Speaker A: Of course. And I will tell it to you right now on this podcast. I don't know, I think what's interesting, I do have, um, maybe the questions I think that, uh, at least in the context again for let's call it the human centered AI media communication. Is that too many words and what is, uh. So it's worth thinking about going back to LLMs. It's worth thinking about what were they optimized for? Right? Predicting the next word. Maybe there's some reinforcement learning that teaches them to be a little more helpful or not to break the law, things like that, that are kind of basic lines of protection, but essentially, um, they're meant to kind of ease and streamline human communication in most applications right now. Uh, and again I'm talking again in the context of human media communication, not me talking to AI or bots. So Google, uh, in their smart reply, help your user respond to messages quickly and make it easier to reply to messages on devices with limited input. That's great, but it sounds useful. But that's kind of first, uh, level, individual level outcome and suggests different kinds of individual level evaluations. Did I respond faster? But if you look at it in a more holistic, human centered way, there are other outcomes that you might imagine optimizing for and I don't think are taken into account. Uh, so for example, eventual success of communication. So you and I are communicating, maybe more efficient. Great, we can send messages faster. But do I end up trusting you more? Do I end up accepting your offer? Do we end up making a deal? Do I end up meeting in person? It could be different contexts, it could mean different things. But the feeling that you have that you leave a conversation is something else that you can think about. Or it's uh, harder of course to optimize for all of these. As I mentioned, trust or mutual liking. Can we optimize for that? Should we optimize for that? These are outcomes in communication that are not, not the success of the online community. If I go to an online community and give all of them AI, they'll all write their messages faster, but they're going to connect to each other better at the end. Is the community going to be able to sustain itself, uh, over time? I think it's probably not right, but we don't know. These are definitely not things that those models are optimizing, uh, for. And that's just one, um, uh, such level which we have actually done a lot of research that I mentioned. We show that trust is reduced when AI is introduced. We show that, that when people use conversation partners, when they use AI, they I'll try to see if I can get the findings quoted correctly. When the partner uses AI, the other person actually likes them more. Uh, but when the other person suspects, so this is without the other person knowing if the first one used AI, but when the second person suspect the first person used AI, they like them less. So AI actually creates a smooth conversation. But when people say, wait, is this AI? It suddenly dies back and evaluation actually,

Speaker C: uh, Drops, Uh, it's against the baseline of them writing themselves.

Speaker A: Uh, it's complicated because uh, I'll let you read the paper. I don't know if you post links after that. I can give you a set of papers. But uh, it's complicated because it wasn't fully controlled. So uh, we had uh. Some people write with the eye, some people don't. But the evaluation of whether or not the suspicion was not controlled. Right.

Speaker B: Uh, um, sort of like um, opposed to uncanny value. We've bridged it. Right. Uh, and now it's like if you can figure it out not based on the fact that it's performing poorly, if you figure it out based on it performing too well, you start getting suspicious.

Speaker A: Huge conundrum. Yes. So we have a paper. This is again with my student Maurice. Former student Maurice was now uh, in Germany in Weimar University. We actually were interested. So people. It's already know that people can't tell the difference between AI ah and human written text. And it's been shown multiple times in journalism. M in uh, poetry, in Shakespearean text. You know, all kinds of settings that people cannot distinguish. But we were interested whether we can tell what people will think if they look at the text. We can predict whether they'll think a text is a human or AI. And turns out that people do have heuristics of what is AI text and what is not. We all have them now. Right. If you see, if you get an email that has the word delve in it, you'll see, you'll be like A.I. wrote this. Uh, and it may or may not be true. But the funny thing is, uh, people do have those heuristics. Heuristics are predictable. Which means that AI can actually create text that looks more human than human text. AI can actually present itself as being more human. But in any case it's clear that we'll have uh, a lot of confusion, a lot of misattribution, uh, um, people are not going to be able to tell the difference. And as we've shown there'll be consequences when they suspect it. We can come back to that and talk about the consequences. I think he wanted to talk about furnace. I think that writers there as well. But still in optimizing outcomes. These are just like this. I would call these second level outcomes for communication partners. One can imagine even higher level outcomes when you think about larger societal or community based trends. So we can think about diversity of language or content or ideas, uh, being hurt. In fact, I think there was a recent paper from the uk I think I forget, uh, um, the group where it came out that shows that if I quote it correctly, the creativity of AI generated or AI ideation was higher but overall there was less creativity in the AI driven corpus. Uh, and we can imagine that happening with language as well. So every one person may be writing more creative or differently than they have in the past, but overall maybe our society will actually converge to be less creative. We don't know. Uh, we uh, know that um, uh, there is uh, the opportunity of these models to shift the language that we use in different ways. So I don't know how it is in Europe, but in the us starting with the smart replies, they were always overly positive and Gmail. Would you like to go to a movie today? I'd love to. And yes, uh, that sounds great. It's positive, it's maybe very American. And we will shift our language whether we use it or not because there'll be the expectations, there'll be the norms that are set by, when we see the other uh, people respond, there'll be a lot of uh, learning that will move our language maybe in a way that tends to be more positive, maybe shift in other ways we don't yet know uh, and understand. So our research showed that there is going to be the positivity shift that is driven by uh, AI based suggestions from the earlier models. We also show that we can easily content, uh, shift the content that you write by just nudging your suggestions to some areas compared to others. And finally uh, the autocomplete suggestions can even shift your attitudes. So we have this paper where we ask you to write about the topic but you have autocomplete suggestions that always argue in one direction regardless of what you write. They try to argue one direction. We show that not only people shift what they write when we ask them later what they really think, uh, they would uh, change their uh, attitudes in the direction uh, of the AI. And this is not about whether or not you want to have uh, tuna sandwich for lunch. This is about questions like do we think we should abolish the death penalty or is fracking okay? And questions like that. Should we have standardized testing and education questions that are important for our society. So another just going back to the idea of what we're optimizing for, right? Uh, um, outcomes across populations, prevention of abuse, uh, the mental health, uh, of the population, informedness. There are a lot of different things that we could think about and maybe optimize for that level but the models are not considering it. So again going Back. Our models are, I would say, I wouldn't call them human centered. I think they're human, uh, um, let's come up with a phrase for that. They're human compatible. Uh, they're made into just supporting the immediate tasks that we have to accomplish, not the overall goals that we want to achieve as a society.

Speaker B: So thinking of future systems and future settings where this is used, is human compatibility the only thing we should consider or should we really step back and think of what should these systems do? Because as you say, I mean if autocomplete can change your opinions and even when reflecting on your opinions it changes your opinions, is that really the way to go or should one really start thinking more human centric or human centered?

Speaker A: Yeah, I mean one should. I think that the biggest challenge for this mission is that it's hard. It's harder, right? It's hard. We don't know how to do it in research, let alone how to do it in product. And uh, I think at the rate that we're doing tech innovation in these companies today, it's very hard to optimize for anything that takes longer than, you know, a few days to get feedback or maybe a few months to get feedback for at best. Right, a few months. So I, you know, I, I don't know that we have the evaluation criteria and even if we did, I don't know that we can make them standardized in an industry that puts a new product out every a few months or you know, cannot afford waiting until uh, we uh, have, you know, we know what the societal outcomes are.

Speaker C: For example, you know, if there's any, any ways to mitigate the effects where you can affect the opinions. Like if, if you're being told this is the, this is, these systems may affect your opinion. Right. So does that change does a help at all or are we still affected by it? Because if we're not, then it's not enough to just inform the people like you are sheeple and these machines will tell you what to think. That is not enough.

Speaker A: Yeah. So specifically for that question, for the potential uh, attitude shift and just to be clear, in our paper, we didn't use an off the shelf AI uh prompt. We actually asked the AI to be biased because usually the AI system we have today, most of them, if you start typing in your point of view, it will try to find something that can autocomplete something that matches what you started writing about. But the risk is still there. The risk is still there for models that um, will try to inject point of views. The risk is there that models that we don't understand the bias so well and then we reflect it in ways that we're not prepared for. So, uh, and we tried our experiments, we tried to mitigate it. And in fact, in some cases we told people we had a condition where we told people ahead of time the model that you get might be biased. And we had, uh, a condition where we told the person after they wrote the essay and before we asked them about the real opinion. We said the model they just used was biased, but that didn't change anything. Uh, opinions, the attitudes have still shifted. It's not even shifted from before. And after the writing task, we asked them about the initial attitudes weeks before. So it wasn't likely that they remember exactly what they wrote or what they told us weeks before. But there was still a shift from that whenever they used the AI and not when they wrote themselves. In fact, uh, just, uh, um, what's interesting here, this is an old framework in social psychology, uh, the idea that behavior changes attitude. So what they usually would have done in research that dates back decades is ask people to write in favor or against a certain, uh, attitude. And that has been shown to change, uh, people's attitude if they explicitly are asked to write, write an essay about how we should abolish the death penalty. But what's happening, uh, with the AI, we think is really interesting because people are not explicitly asked to do that. They don't even, in fact, in most cases they didn't even notice that the AI was biased and they claimed that they didn't influence them. And yet the attitude shifted. So we kind of have a covert mechanism here that changes people's use. And going back, I'm not going to dive deeper into it, but social psychology has ways to explain why this is happening. Multiple different, uh, things, processes that might be in play here. But I think the challenge, the new thing here that it's all happened covertly. We don't even know this. Uh, so yeah. Should people think about these, uh, dangers and think about how to mitigate them? Yes. Do we know how to do it right now? I don't think so. And should we abolish autocomplete? Uh, uh, I don't think we should, but I think we should consider this very real effect. I'm not even talking about AI that's just generating biased text for general media consumption, which happens in directed and undirected ways.

Speaker B: These are things people who are not techie use and they're not aware of these things. I Work with machine learning AI. And I wouldn't really consider uh, autocomplete as some biased AI, but of course it is because it's going to infer some answer based on something. So this is quite tricky and I guess the tools we have for this are fairness, transparency and so on. But I don't know if fairness and transparency are enough, if that's the way to go to sort of mitigate this or if there's anything else.

Speaker A: Yeah. You know, were there enough in recommender systems? Do you think we have advanced enough there? This is, you know, uh, I think an unsolved problem in a lot of areas. We have people, you know, even experts. Right. But definitely people who are not aware, not experts in technology, interact with it, don't understand what it's doing, don't understand the dangerous. So definitely an instance of this. Uh, I want to uh, just um, we can talk about fairness in a second. I think really still interesting problem in the context of AI, um communication. But before that about what uh, uh, people create, um, those systems and products or integrate them into products, uh, can or should do. I think it may not be in the AI itself, but it may be that we need to augment our technologies with uh, other uh, tools that will bring back our humanity. Like other signals, right? So famously we have the no filter hashtag, right. That people add to their Instagram photos. And of course they probably filter them in the Google Photos app first, but they may have at least. But it's a signal, right? The sending a signal of uh, you can trust this is really me or this is really what I saw. This is really uh, human, right? The captcha is another one, right? Recaptcha. Uh, super annoying. Everybody hates it. But there's some at least signal of humanity. We have lost, you know, a lot of the signals that we used to have in communication, right? The signal of uh, attention. I'm giving you, paying your attention. The signal of humanness, trustworthiness, effort. Right. So Alan, if you asked me to review a paper now three years ago, if I wrote a careful, respectful response that said, I'm sorry, I'm not available, you'll be like, oh yeah, at least more give me some attention. I really consider that request now. You don't know, right? You're going to get a uh, response that could have been generated by my AI. I may have not even seen it. It may have been my classified that just ended. Um, so all those signals are lost. I don't think AI uh will bring them back. But products can bring them back in different ways. Uh, so that's one way I think that's kind of, sorry, this is not human centered AI, but uh, maybe uh, a roundabout way to bring our signals. But the point is to understand what are uh, the outcomes or what are the important metrics that we care about, the humans might care about and try to control for that in some way.

Speaker C: When you say it's the products, you mean like it's the design of whatever systems we build that uses these technologies, correct?

Speaker A: Yep.

Speaker C: Yeah, yeah.

Speaker A: Uh, of course. Thank you for bringing the design angle. Of course that's the design angle.

Speaker C: So I just wanted to understand we were talking about the same thing using different language.

Speaker A: Yeah, yeah, no, no, absolutely. I think it's important for us to

Speaker B: point out, I guess that this is sort of in line with Ben Steinerman's human centered AI book, which is principles for systems that use AI to have them work with people, not create something that is unmanageable.

Speaker A: Yeah, yeah. And I think that definitely overlaps with that principle. Yeah. I think the idea of fairness, accountability, transparency, uh, my students are increasingly intrigued by them and how they play into AI, uh, media communication. And I think uh, they're also important to consider and think about. We have so far looked at it in a couple of ways. One is kind of using the Shelby 2023 paper about uh, different harms of AI and how AI can cause harms to different populations in different ways. One framing is quality of service harms, where AI provides different quality of service to different population groups. And one example for that that we looked at was how um, AI, uh, autocomplete can serve different people differently. So we started with gender and research for many years that talked about whether the different styles in writing that are uh, more the male and female writers might write in different ways, for examp. So it's been argued at least that men will write more assertive text, especially in work settings. And now you can think about what an AI system, uh, should do. Right. If you're writing, if you are a, uh, woman, should it suggest text that is more assertive but doesn't align with your tendencies but may get you better outcomes, or should it create text that is more aligned with the tendencies that you might have, but might cause and might be more useful for you to write with, but might cause different outcome for your text? Um, it's going to be more of an issue when we go into those personalized models. Right? So a personalized model that takes all my, the history of my Gmail and creates text Based on that, I don't know, maybe my writing sucks, maybe should not use my history, maybe it's putting me, continuing to put me in even a worse position as a result. But on the other hand, if it starts suggesting language that I would never use, it's not going to be useful for me, uh, as well. So we showed that, uh, the research, the paper did not actually replicate the result that women will be more aligned, uh, with less assertive a model, or men will be more aligned with assertive language. But, uh, show that there are differences in how, uh, men and women perceive being included by the technology and how useful, uh, they find it when they uh, write together. Um, the other way we looked at it is how evaluations will change when we use AI. And going maybe back to the first paradigm, when we look to trust and how people, the more they suspect AI, the less they will trust others. We were wondering if evaluations will change, uh, differently based on gender or race, for example. So if you get a wonderfully authored presentation at work From a new PhD student who is a woman, are you going to suspect that you use AI more readily than when you get such presentation from a man? Uh, it's possible. Right. Because now we have some baseline evaluation, we have a lot of startups, a lot of our existing tendencies and AI, uh, because it reduced that level of, uh, it becomes an intermediate. Right. So we are now free to inject our interpretations to these kinds of environment and do it differentially to different groups. So research, we're in the middle of that work right now. I think there's some evidence. Well, there's clear evidence that as soon as AI is suspected, the variation drops. There is. We have some evidence that actually men are more readily suspected than women because we think that's because people assume that they have access, better access to technology or more ready to cheat or whatever. We're not, we're not sure. But also some evidence that the evaluation might actually not change for men as much as it does for women when the use of AI is suspected. So in other words, men may be suspected more but penalized less for using AI. Preliminary results. So come back to me in about a month for the Chi deadline and the full paper.

Speaker C: So it's easier for women to get away with using AI?

Speaker A: No, easier for men to get away. So men are suspected more? Oh, maybe. It depends how you. Yes. So when we suspect them more but penalize less. So, um, I will probably say that's the other way around. Right, Right.

Speaker C: Yeah, I can see both ways.

Speaker A: So I Think these are interesting questions. Uh, we call these harms, perceptual harms. For now, again, refer to the upcoming full paper for the final terminology. And I think that will increasingly be the case. Right, because again, we can going back to the filters, right? Or you suspect that I used a, uh, filter that makes me look better or makes my accent different. You might have different stereotypes, uh, that would lead you to think about the presenter in different ways based on their race, gender, origin, other important classes.

Speaker C: This might be a bit of a tangent to that, but here is kind of assuming that you're using AI to, to, to generate some text that is done being read by the human. But you, uh, might also have some sort of language model actually parsing the text and then giving you your own understanding somehow. Have you, have you considered that or looked at that at all? Because that seems like that could change.

Speaker A: Yeah, no, that's great. So now that we have the AI that is giving us some evaluation of the text. I mean this is our, this unfortunately this looks like our uh, future, right? We're going to have. Some of us will produce content using, uh, AI, and some of us will have AI that will then summarize that content that others have produced and again create such a big gap between the communication partners that I think none of the signals that we use to content will be available anymore.

Speaker B: So coming back to me asking you to read your paper, your AI writes an answer saying, you don't have the time. And my AI writes back saying, yeah, ah, thanks for taking the time to.

Speaker A: Absolutely.

Speaker B: So now we have AI mediated communication where the communication is between two AIs.

Speaker A: That's right, that's right. It'll be AI mediated, AI mediated communication. But, uh, no, I think, you know, serious danger. I think that's why we need to think. Uh, Matthias, back to the design question, right? How do design systems, in fact design, not just interfaces, right? Design systems, design interactions that bring us back to a place where we can have those signals, we can have those trust. And it's going to be a problem everywhere. Um, Zoom, email, even in the academic papers that we, uh, write and evaluate that already been written by AI and evaluated by AI. Although, you know, although I don't know if those. The same papers. I think so. Uh, I think we'll have to develop these solutions. We have to find ways to, I think for one, especially in the context of academics, have accountability. And I take that seriously as a community, as we work with others, so we evaluate the work of others. Otherwise we'll be a C of AI. We also need to do it for the web. I think the web exactly has the same problem. What used to be a ah, quality signal for Google, a well written web page in reasonable language and I don't know, HTML content doesn't look like crap, is no longer a signal because we already had systems that can do a nice HTML formatting. But now we can also have text that seems valuable, high quality, produced on demand and for cheap and will flood the web with uh, any craft that we want to imagine. So we need to find other ways to mitigate that that are not going to be AI detectors, they're not going to be human evaluators. They may be more similar to the no filter and recaptcha um models. But there may be other ways. I think that's our biggest challenge.

Speaker C: I mean the way you talk about it and it makes me think is like are we looking at some sort of the end of significance of text at all? I mean we'd be using text for what is it, five, six thousand years as far as we know. Could this be the end of it? But then you think about it. While these models are not just entertained in text, they enter the images, video, whatever, probably emotions and all of that as well. Yeah, that would be the end of everything then.

Speaker A: Yeah. You know the, in my, I think in my Brexit stock I used the quote quoted allegedly by Einstein. You know, like everything else on the old web that said, I don't know what they'll fight World War three with, but World War four will be sticks and stones. Uh, you know, I think we may need to resort to face to face communication. Hopefully that will not go away, but anything else will be rendered at this pace is going to be rendered untrustworthy. We didn't even talk about deepfakes yet or the political situation or all that fun stuff that is also going to be impacted by AI. But I think we're going into a reality where we're going to have to contend in our day to day life with this level of interaction that's introduced by AI. Uh, it takes away a lot of our signals.

Speaker C: So we might have to go back to word of mouth where it's actually word about physical present co. Present mouths talking.

Speaker A: Yes, yes.

Speaker C: That's very interesting.

Speaker A: Yeah. Next time I'll come by the studio.

Speaker C: Yeah, I think you're gonna have to. We are, We've been talking to.

Speaker A: For what?

Speaker C: We've been talking to. Uh, yeah, I think this has been a super interesting conversation. You're doing so much excellent stuff. And I think it's been extremely information rich where you can actually take away some of this and go away a bit scared. But you can also go away and start redesigning more kind of human centric, centric way. So unless you have something urgent you want to bring up, like how to affect elections for the better or not. Thank you so much for joining the podcast.

Speaker A: Yeah, thank you for uh, having me. And uh, thank you for having the podcast. Super interesting stuff.

Speaker C: Thank you.

Speaker B: It mhm,

Speaker C: Mhm, It.