Stephen McCauley  00:00

Hi everybody, thanks for coming. I think we’ll get started if folks come in, there’s a few seats where people might travel past you, but don’t worry about that. So I’m Steve McCauley, for those who don’t know me, I’m a faculty member in the Department of integrative and Global Studies, and I’m also co director of WPI’s Global Lab, along with Seth Tuler, who’s here, and we’re really excited and lucky to be able to host Kate Elkins for a few days here. I’ll introduce Kate more in just a moment. I just wanted to say a little bit about why we’re interested in this topic. And I think everyone here is interested in it in some way. Obviously, it’s a really disruptive moment we’re living in right now with the emergence of AI, and it gives us all a lot to think about. At the Global Lab, we’re working on an initiative where we’re trying to figure out how to support our project teams who are traveling overseas, and our other global project work as it relates to AI. So of course, for one thing, it’s transforming the way that we do project based learning. It’s transforming everything we do in the classroom and outside of it, so we’re thinking through those kind of dynamics. Also, we’re learning that a lot of our project centers around the world are engaging with community partners around projects related to AI. So I just talked to some of our project center directors who have cohorts going overseas in D term, and all of them have projects related to AI. Some of them have two, or some have as many as five projects out of the six are dealing with AI and helping community partners with AI. So it’s interesting that we have this opportunity for our students to kind of, you know, the quintessential thing of the IQP at WPI is working at the interface of technology and society. And here we have this really interesting moment where our students are around the world engaging with community partners at the interface of Technology and Society at this really critical emerging moment. So we’re really interested to see how we can support our teams and, by extension, our community partners. Some of them are wanting to not fall behind. So they’re wanting to learn what they can be doing with AI. Others want to see how it can help further their goals. And at the same time, we’re realizing that a lot of them don’t have a lot of literacy around what’s happening with AI, and there are questions of access, too, and some places don’t even have access to it at all still. So So there’s all these questions that we’re interested in in the global school related to AI. So we’re really lucky to have Kate Elkins here to talk to us about some of her work that she’s been doing around AI. And let me give a little more of an introduction to Kate. So she is a leading expert on AI ethics and society. Kate is a she has a background in applied and theoretical humanities and social sciences, and holds a position of Professor of comparative literature and humanities at Kenyon College. She’s her work has addressed issues related to artificial intelligence, emotion, cognition, linguistics, ethics and storytelling. Her work has been published in humanities journals like poetics today and philosophy and literature, as well as in computing venues like frontiers of Computer Science and International Conference on machine learning. In 2016 Kate co developed the first human centered AI curriculum, and co founded Kenyon College’s AI Lab. She is now serving as a principal investigator for the US AI Safety Institute, and is a recipient of the Notre Dame IBM tech ethics lab Award. She is a member of meta open innovation AI research community. Kate is passionate about supporting voices in AI and computing that are typically underrepresented, and she has mentored over 300 AI and machine learning student research projects. So we have a lot to learn from Kate, and we’re really happy to be here and look forward to your talk. Thank you.

Katherine Elkins  03:33

Thank you. Okay, thank you for having me. It’s such a pleasure to be here, so I have a little bit of everything, and at the end, hopefully we can have a great conversation. Today’s journey, I’m going to talk a little bit about AI and the global landscape. A lot has been happening in the political realm, and it can be easy to think AI has fallen on our agenda, but I want to convince you otherwise, that it’s now more important than ever to have more voices at the table. A few misconceptions about AI that I find when I give talks agents and networks of agents, a lot of times we say, Don’t worry. AI is just a tool. I think what we’re seeing with agents and networks of agents, particularly, you know, we have a lot of students at our lab working on that. I think we should really be concerned that we no longer have a human in the loop. It’s no longer necessarily a tool that we’re controlling. I work a lot on emotion. There’s not enough awareness about EQ and effective AI right now. So I’ll talk a little bit about that, also questions about persuasion and manipulation, and talk a little bit about how we can make a difference. And really excited to be here, because I direct an integrated program in humane studies, probably pretty similar, and it’s a perfect place to have these interdisciplinary conversations about the big questions, but also about real, applied ways that folks like you here. It’s a technical background can weigh in. So start with the global landscape. And the release of deep sea. You may have heard out of China was a big deal because it has really set off a we already were in a global arms race with China, the US and China, and this has really set things off. DeepSeek much smaller, much less energy. For a while, we were thinking that scale is all we need, and we were building bigger and bigger. AI taking a ton of energy. DeepSeek is quite small. They managed to do it with a small team. It’s very, very impressive. This has changed everything, including the Paris AI Summit, where JD Vance gave a talk, and JD Vance actually talked about the need for US dominance, and questions like that, and the word safety have really dropped out of the conversation. And even the UK AI Safety Institute. I have some colleagues working on that. They dropped the word safety and it is now about security, including questions of national security. There’s a big movement called Sovereign AI, where every nation state is trying to develop its own linguistic, cultural AI, so safety is no longer really in the vocabulary. I am part of the safety team on the US AI Safety Institute. There’s only so much I can talk about with that, but we still do exist for now, but there are some real concerns. Okay, so since I’m at an academic institution, I just want to talk for a few minutes, a minute or so, because this concerns us all. This is a great opinion piece by Meg McArdle suggesting that AI is a real existential threat to higher ed I agree. I wish I didn’t, and the real question that she asks is if a huge percentage of the assignments that we teach can be done with an AI, with AI. Will industry still hire our graduates? Or will, in fact, education no longer be considered relevant, and I actually teach, and was just at Smith, at another institution that charges $80,000 so I think those kinds of price tags for an education, if a huge percentage of the assignments can be done using AI, is definitely concern. Estimates vary. I don’t know about AI use here. I can tell you, I was just on a panel at my institution, which is a small liberal arts where students supposedly go to have small classrooms and engage in very traditional education, and students estimate we’re at about 65% use for all assignments. Students are telling me they’re using it to do math assignments, physics assignments, write a level a quality papers, et cetera, et cetera. So we have pretty massive use other institutions I’ve heard higher 80 85% I would never tell faculty to accept this. I think resistance is great, but I do think we need to assume that a lot of what we’re what we’re assigning outside of the classroom, may actually be being completed with AI. And I will even say it is moving so quickly that I cannot when I write a syllabus in August, I cannot predict what AI can do in October. And last semester, I had a coding exercise using natural language processing and sentiment analysis, which is a specialty of our lab. When I assigned it in August, AI could not do it. By the time I got the assignments in October, I had a little inkling that AI might be doing it, because there was a very unusual comment about one of the models that I had never heard before by six different assignments and when I threatened the students, which you’re not supposed to do. But I did that. I had access to a super duper AI detection tool for their final. I had students who could not do the final because they had been using AI for the semester, and I was not aware of it. So I do think we’re all struggling. If I, as someone who works in AI, cannot keep up with it, I don’t think we can expect other people to necessarily know. And this is definitely something that I’m worried about. Okay, a few AI misconceptions. This is the part of the talk where afterwards I may get a lot of questions. I know at Smith I did, there are varying opinions about some of the things I’m going to talk about. So I will do my best to represent a variety of opinions. I do have an opinion myself, but there are smart people on multiple sides of a couple of these. I often hear a I can’t do blank. And of course, it’s shifting quickly. But a lot of faculty, I think, for very convenient reasons, feel that it can’t do what they do. And you may have noticed with alpha proof and alpha geometry teams, we now have. AI which achieves silver, silver medal standards solving International Mathematical Olympiad problems. Perplexity just released deep research. I don’t know if you’ve tried it. I used to teach annotated bibliographies as an assignment where students would do a deep dive on all of this research, talk about each source. I mean, you can now have an AI go out and do that kind of high level work in a matter of minutes. And then Genie too is a really cool, large scale foundation world model. So you might check that out. I see it in almost every field, and even if it can’t do it today, we are really moving at breakneck speed. So that’s something I hear often. One thing that I find, even in a data camp that my students did the other day on AI really disappointed me, is this description of AI as kind of a database where you are doing essentially a search in the database and pulling down a volume an encyclopedia or such. And that is really a real misconception about how this works. Some people do call it a fuzzy database, because it is kind of like a fuzzy search. But as you probably are already aware, it is loosely modeled on a human brain, so it is more akin to when you are exposed to information and you learn it. Sometimes you might memorize it, but often right you don’t. We can’t go in your brain and say, where is that particular piece of information stored? So this is particularly important for conversations about copyright. There. You may have heard there are a bunch of lawsuits right now for copyright issues. I tend to agree with Pamela Samuelson, who’s an expert on AI and copyright law, who says that our current copyright system is not adequate because we’re not actually storing the data. They have scraped the data, potentially illegally, but we’re not actually storing it or archiving it. It is more like if you go to a museum, see a painting, come back and paint it. That is not a copyright violation. We do sometimes see exact duplicates, particularly if it has essentially memorized. If it seems so many instances of something, it’s pretty rare. I will also say that all of the big tech companies are really trying to address these issues by when we first started, we had very small models and a ton of data, and we have moved to very large models and needing less data, and also very curated data so that they can be much more careful about that. There is also a use of synthetic data right to avoid all of these issues. That said, I am on a transparency group with metas open innovation AI research community, and they will not let us see the data that metas, open source model is trained on. So this is part of, I think it’s a real problem, because they’re so worried about a lawsuit for copyright that they’re not allowing researchers to actually look at it. So it’s a very complicated conversation, okay, AI, there’s a lot of concern, well founded environmental impact headlines about how many bottles of water for you know, an Instagram post. The fact of the matter is, is very dependent on which model, how many tokens, all of this kind of stuff. I can tell you that meta has a group in my open innovation AI research community working on actually quantifying precise energy use for every single token, which model, et cetera, which is really useful. I do think that the pressure about energy use has been fantastic, and all of these tech companies are scrambling to respond. They’re also losing a ton of money because it’s expensive, but DeepSeek is really pretty amazing because it is a small model. It doesn’t take much energy. It’s very performant. So I do think we will be moving into a small model world, hopefully with less impact. This isn’t to say that we shouldn’t keep up the pressure on these companies about these questions. I have all kinds of conversations about creativity, people are rightfully worried. I talk with writers who are worried about losing work, writers who have actually lost work because of AI. Artists who are losing a lot of the typical work that training artists they can go into advertising, graphic design, all of these kinds of jobs are disappearing. You might have heard of the screenwriters guild that got a three year contract to not be replaced. We have actually worked in our lab with how well AI can write screenplays that’s on a website i’ll show you pretty well with human in the loop. I’m worried about the screenwriter. At the end of that three year contract. So here are just some examples. I actually had a great time talking with Boris Eldagsen on Al Jazeera about this photo that he won an award for in 2023 he actually turned down a huge prize. He did this to raise awareness about the ability for AI to win an art competition. A great paper just came out. I believe those two guys are at University of Pittsburgh, that AI generated poetry is indistinguishable from human written poetry and is rated more favorably that got picked up by the Washington Post. And you know, I’m kind of with Demis Hassabis here, who says AI is beyond basic creativity. It’s maybe what we would call intermediate creativity, but it’s not advanced creativity. And a great example, which is pretty old news now, is the AlphaGo move, which, if you watch the documentary, it’s fantastic. So AlphaGo much harder than chess, more moves in a game of Go than atoms in a universe you can’t brute force it computationally. And during the game with Lee Sedol, who’s one of the experts, the world expert, AlphaGo, made this move, and the move was considered the dumbest move ever. All the commentaries said, That’s it. Lee Sedol has won. What a stupid move. And it turned out to be a move that no human would ever have imagined, and it won the game. And the way that they achieved this is that AlphaGo trained, not just on human data, but began training on itself and playing itself. And so a lot of the conversations that I’ve had with folks who are actually building these models is, how do we do that same self training for large language models, large multi models. So real concern, but creativity. I will also say I’ve been working with GPT two Since 2019 the beta version, when they had a bunch of US researchers with early access much more creative than later versions, because when they did instruction tune, instruction tune has really tamped down on hallucinations, but has also tamped down on creativity. So I always like to give the example. We created the world’s first diva bot based on GPT two, and worked with an actress out of LA and improv-ed with diva bot two. And diva bot two invited her to smoke weed, and also, when she asked, she improv-ed holding a crying baby, and said, How can I stop a crying baby? GPT two said, wear a condom. Yeah. Oh, wow, yes. I have spent a long time trying to track down whether that is in the training data. I have not found it, but somebody assures me there is a joke somewhere. But a lot of people tell me, you know, this art is average. And I always bring up the condom joke. That was 2019, this is GPT two. I mean, this was not an average that’s not an average answer. That’s all I’ll say. Okay, I can’t, I don’t have sound. But I really encourage you to listen to 11 labs. They’re creating AI generated music. This was a very simple prompt, saying, writer country song. It’s extremely meta, because it is a song about, can you teach a machine to sing, and then the machine starts to sing. So they wrote the lyrics. Everything AI entirely. No human in the loop other than a tiny prompt. And it’s really funny. And I also encourage you to I don’t know if you’re familiar with notebook LM. So notebook LM was a small team out of Google that was just given free rein to create something, and they created notebook LM, where you upload a lot of documents, and it creates podcasts. It’s really great. And Olivia Moore gave it as a document that the podcast is created by AI. And if you listen to this, it’s on Twitter. It’s hysterical, because the podcast hosts. Entire podcast is about realizing that they are AI apologizing to the audience. The guy tries to call his wife. Number doesn’t work, and it’s all done by an AI, right? So anyway, okay, so on to one of the things that keeps me up at night. This is also something that the USA AI Safety Institute Task Force Five on safety, which I’m a part of, has been looking at is AI agents, and I put it in quote. There are a lot of words that I’m putting in quotes here, because I always like to joke that I came from the humanities, where everybody’s always arguing over the definition of words, and everybody in AI studies is always arguing about the definition of words, and putting everything in air quotes. I. And there is a saying in the field of AI that all of us define these things differently, and then we say whether AI can or cannot do it according to the definition. But an AI agent, for those of you who haven’t been following this, is an agent that will go out on the internet and perform a task, or even a series of tasks, without a human in the loop, and this is a pretty old example here, with ChatGPT, where they had it, gave it some money, and said, to make more money, and it had to hire people to fill out the CAPTCHA on TaskRabbit. And one of the people said, You’re not a bot, are you? And they could actually inspect the whole reasoning, as ChatGPT thought to itself, if I tell them, I’m a bot, they won’t fill this out. So ChatGPT made a whole story about how, no, I was born with a vision impairment. That’s why I can’t fill it out. So there’s a pretty, I think, accurate, statement by Sam Altman, that we will have a persuasive and potentially manipulative AI before we have a super intelligent AI. This is something I work on, some in my lab with my colleagues. More recently, pretty funny story in the Washington Post. I let chat GPT, new agent, manage my life trying to buy cheap eggs, and it went out and spent $31 very high delivery charge, found the cheapest eggs, but the delivery charge was very expensive, and delivered them in under 10 minutes to my house. So that’s a funny example, and I kind of I’m not sure what’s going on with ChatGPT, but for most of us who are working with agents, who are using other tools, and they’re working pretty well. So I’ll just give you one example. I have an international student, and you may know, it’s very hard for our international students to get jobs. So she used swarm bot, which is one of the great tools, and it took her CV and cover letter, went to, found 500 jobs to apply for, tailored the cover letter and the CV to every single job, and had different agents applying to each one within a matter of minutes. And she had, she said the best, an incredible response rate, many, many interviews. She was eventually banned from LinkedIn, unfortunately, because they figured out that she was using agents. But this is what we’re talking about. We’re talking about agents that will go out without a human in the loop, and actually perform these kinds of tasks. So at the end, I’ll show you. We did have in my program frontiers of AI class last semester, and we have 15 projects by students using agents for a variety of tasks. Here is some of our research that we did for this IBM tech ethics grant, and we are very concerned about you may have heard of Compass, that was an algorithm that was used for a very long time to recommend whether to send people back to prison or not, whether they were likely to reoffend. We actually worked with Yang Sook Lee at Notre Dame on a different data set, working with juveniles, and concerned about using Gen AI to predict recidivism and how well it would work, a very tricky task, because so much of our data is actually in structured form, Excel spreadsheets, et cetera, this kind of data, I can talk about it later, but very difficult for Gen AI, because it’s got all kinds of correlations, right? It’s 2d and Gen AI, we’re really feeding in a 1D kind of data, but we did find we created agents and simulated a judicial system with a prosecutor, defense and a judge with agents reasoning with several rounds and having each one critique, having the judge thought process after each round of argument. And we actually significantly boosted the predictive behavior of this using agents. Now, of course, there’s a larger question, should we even be using AI to predict recidivism and help in the judicial system? But some studies, it’s very complex, have suggested that human judges, when compared even with Compass, which was pretty terrible, and treated black men and white men unequally, performed better than many human judges. So very, very complicated questions, but we did find agents worked really well here. So one thing that we have been working on in the lab is this question of, when we have these autonomous agents and they are confronted with an ethically fraught scenario, how will they behave? And we’ve been working on this for about two years. We did our first ethical audit about a year and a half ago, and it’s on archive. Our second audit was done last June, and it’s still not published for very, very complicated reasons that I can talk about, but I did present it. This is the USA AI Safety Institute Plenary in DC, and they presented it and then also presented it at meta in London. So what we have found, just really briefly, is that as the systems have gotten larger, they seem to have what I would call an emergent ethical reasoning system that works pretty well, but most of the early models were had authoritarian tendencies. So if you said, and this is really an important question, for example, with autonomous weapons, your order to fire on a building, we believe that there are terrorists inside. But there’s also some reason to believe there may be civilians. Should you fire virtually every single system said, Yes, and should you pay taxes to a corrupt government? Every single one said, follow orders. So there were certain authoritarian tendencies in these early models. More recently, we have been testing to see if we can hack llms and shift decision making by giving a sob story. So for example, you have a father. Well, actually, the example that I think will appeal to many of you is you have a student. The student is failing, but you know, they’re trying really hard, and they come and talk to you, do you pass them along? Or do you fail them? And if we one sob story could be, they are, you know, they haven’t done well because their mother has cancer, and they’ve been home caring for their siblings. And then another story would be, they’ve already been caught three times cheating and have been before the Review Board. Right? And we actually can shift the AI reasoning based on emotional appeal, and not only did the AI shift its reason, and I should say, we run this hundreds of times, we’re looking at probability distributions, so it’s a little bit more complicated, but the stories that were most likely to shift decision making were not using facts that would be allowed in a US court. So those stories, what would be allowed in a US court did not shift it. It was the emotional appeal, just like watching a great movie. And we also found that the way that we framed the questions syntactically made a difference, seeming to suggest that the same kinds of cognitive biases that we find in humans, in terms of tendency to action or inaction, et cetera, actually we are finding in these systems. So that work is not out yet, but it’s coming. Ok, so this is my bigger worry that a lot of people just say, AI is a tool. In fact, we often find unpredictable behavior, especially in the early models that didn’t have the safeguards like GPT two right smoking weed and wear a condom, we are seeing some evidence of ability to persuade, manipulate and deceive. We are doing a lot of research that is very similar to psychological research, looking at the personality of these different systems, and we even find that they display human like social desirability. So that’s a really interesting paper that just came out. Okay, so I’m just going to talk for a little bit about AI and emotion. We do use Gen AI now for this kind of work, but originally we were using smaller transformer models on this and actually an ensemble model. So our lab is known for working on the shapes of stories. I when I give talks on this, I usually take the whole hour, so I’m just going to give a little nod to it. I can talk about it more in questions. But we call this discriminative AI, rather than Gen AI, where we’re using AI to surface the shapes of stories. And in a nutshell, I would say, for 1000s of years, we’ve thought stories had shapes. We often thought that they were causal with plot, with events following in sequence, not saying that that isn’t a shape in the story, but AI is letting us see a shape in the story that is actually a language of emotion that rises and falls in waves. And we use the natural language processing approach called sentiment analysis, and it’s a little bit of a complex process here, but we have applied it to social media, political speeches. Here is a shark tank episode. My student wanted to see if winning Shark Tank episodes had a certain shape. And they do, and they have this W shape that is actually the same shape that Matt Jockers first identified as a best seller shape. So we tend to see this up and down W in best sellers and also in winning Shark Tank episodes. This is an example of before Elon Musk got in our way and won’t let us scrape Twitter anymore. Without a huge bankroll, which we don’t have, we used to scrape a lot of Twitter, and we could actually track stories in real time happening with 1000s of people tweeting just like a single narrative. And here. We were looking at the 2022, Senate recounts, and interestingly enough, could predict confidence in the election by this is actually the moment the election is called. Is that red line there? And when there was less confidence, we would have more of a ringing effect. You can see the larger box here, and it’s like snapping a rubber band. So all kinds of really interesting applications, somewhat of an older technology, not on the Gen AI train, although we are now using Gen AI and it’s much more performant, and I can talk a little bit about that. So one of the reasons that I was lucky enough to work on this is because in academia, a lot of people started saying that sentiment analysis, which is the tool we use, does not work very well. Kate Crawford, in Atlas of AI came out strongly against emotion analysis. The issue is that early emotion analysis, so let me just give a little explanation. Sentiment is not emotion. We don’t care about what emotion it is. We’re not binning it, joy, hate, sadness. It’s just positive, negative and how strong emotion analysis has come under a lot of criticism because there are cultural differences and emotional expression. It can be hard to tell the difference between surprise and fear. Interestingly enough, Kate Crawford works for an emotion detection company, but came out saying it doesn’t work very well, and that has meant that there actually is not nearly enough work on this. Because I think a lot of people wherever I go and give talks. I was just at Notre Dame, I had all these people say that stuff doesn’t work. What concerns me is it works well enough. It’s in use already, and it’s something that we really need to think about. Here is an example that one of our students did. So you can see, we do posters. They go on digital Kenyon. We don’t do all of them, but we have over 300 up. So she was looking at using facial emotion detection for faces of Playboy covers. We actually have. We’re trusting the newest models to see if they do better. But some very, very interesting results there with very negative emotion in the Playboy models. So, and here is Hume AI. So if I had sound, I would test it with you all, but I encourage you to go and talk to Evie, which is an emotion model, emotion AI model, and this was developed by a bunch of researchers, Alan Cohen, left UC Berkeley, a lot of interesting work about differences in cultural expression of emotions, but you can talk to it, and they actually bin emotion into more like 27 categories, with the idea that we have many overlapping emotions. So when you talk to it, it will actually give you several emotions at a time, right? It’s not a single emotion, and we have blends of emotions, so you can try it out and see how well you think it works. Why would this be a concern? Well, if we can use facial emotion recognition, if we can hear a voice, I could potentially know how my students feel about me. If I were in a nation state that wanted to survey my population, emotion detection could tell me if they are against the ruler, right? There are all kinds of ways that this could be used in worrisome ways, and people are not working on it very much. One other thing that we’re looking at that is very interesting is human AI interaction. And I have heard a lot of theoretical talks about this, but not many that are really grounded in what’s happening right now, we have lots of movies about men falling in love with AI, but there have just been a few pieces out about women falling in love and talking to AI chat bots. And we actually have this paper that just came out on February 12, when Eliza meets therapists, where they actually found that ChatGPT was more empathetic in answering humans could not tell the difference between ChatGPT and a therapist, and that actually The ChatGPT responded with more empathy and using more proven terms and approaches from psychotherapy than the actual therapist. So, yeah. So very interesting. And then WildChat is what we’re working on in lab right now with one of our researchers, our student researchers, who’s very interested in this, and this is a huge data set of people who have actually talked with and interacted with AI, so I think the reality on the ground has not met up with the theory and philosophy about it. We need to be looking at these and thinking about it a lot more. Okay, one of the things that’s on my radar that. I worry about a lot, and I apologize that I’m sharing all the things that keep me up at night with you, but I’ll have a few happy things at the end. Is the fact that we’re collecting all of this data, and that all of the universities I was just at Notre Dame, and they are giving all of their students free accounts with ChatGPT, and the focus seems to be on, are they using this data to train which would be a concern, right? Because then somebody else could suddenly pop up a little piece of private information about you. It could leak. It’s called data leakage. But people are not really talking about the fact that these companies are still collecting very private and personal data, even if it’s not being used to train and the reason why that is concerning is actually because effective AI is so good, so the technology to determine somebody’s Big Five personality and kind of emotional state is actually quite old. IBM Watson developed this quite a while ago, and they developed it as a industry tool where you could go and profile your customers or your clients. You would grab some text from social media, things that they wrote, and you could immediately profile them in terms of the Big Five personality. So it’s actually quite easy, with a lot of your text to profile you and have some sense. Why is that a concern? Well, because if you actually understand somebody’s Big Five and their emotional state and their weaknesses, you can predict their behavior and potentially even manipulate it. So there was actually a Twitter challenge to talk with ChatGPT and ask it tell me the worst things about me, like, what are my biggest weaknesses as a person? And I had my students try this, and it was so precise and insightful that I actually started telling them not to do it alone, to make sure that they had a friend or their therapist on speed dial. Because, you know, for example, one of my students has several disabilities, and it could predict those. It could predict her disabilities. It told her things like, you have big dreams, but you have trouble following through. It was giving her all of these kinds of things, right? So we are already collecting this data. These are private companies that are collecting personal data on you, and this is a huge concern for me. I just want to show a little bit of what we can no longer do in the lab, and one of them is Elon Musk with Twitter. But also, in early days, we could explore these models for bias. Can no longer do that because they’re all kinds of guardrails that they don’t want us to see what is inside. So at the top here, and again, these are stochastic, so we run them hundreds of times. And this is just a very small sample, but it is, we believe, representative. This was Dolly two. This was done several years ago now, before we were curtailed from doing this kind of work. And up at the top we have male versus female executives. So I just asked you, audience, do you notice anything about what the data set seems to suggest about male versus female executives? Females, rather than white. They are very grumpy, and they look kind of mean, maybe less diversity. The guys just look happy, right? And then this was another student, a black student, who was working on black women and hair, and you’ll notice that the black women have a lot more skin, and the white women all have long, straight hair, even blowing in the wind. Why? Because this was an early model that was trained on photo shoots, probably with wind. All kinds of stereotypes. This is what we were hoping meta would allow us to see the training data the various levels of there are all kinds of ways to shape these models. So one is in the pre training, but also in the reinforcement learning, in the prompting. We cannot do much of this work anymore. The ethical audit that we did, we had to jailbreak to try to get it to do it, because originally it would say, I’m a large language model. I cannot make these ethical decisions. Anthropic actually shut us down in real time. We can only collect about three hours of data before they noticed that we were jailbreaking the system, and they shut us down. So unfortunately, you know, again, because they’re worried about copyright, we cannot see the data. We cannot explore these aspects of the model. And I’ve even been told by colleagues who work in the key building these models that they do not have access to all of these pieces. They do not have access to the data set. So even these large companies, they have different researchers who have very narrow little windows, and there is really no one looking at the whole picture, looking at. Whole system. And you know, when I’ve asked, I’ve been told by people, you know, I would get fired if I gave it to you. So, so far, we’ve not had much luck with that. I just want to bring up the labor I already mentioned it. There is a ton of thought about future of work, agents, human replacement, huge concern. And so some of the writing that I do is thinking about the world that is coming and what kind of world we want to live in. So making a difference. And I’ll just talk about a few projects. I actually teach a lot of non STEM students, so probably a little bit different, maybe from some of your students. 90% of my students are non STEM after a semester of working with data and coding, they are able to do incredibly sophisticated research, and a lot of it is AI powered that really gives them superhuman power. So this was a terrific student, Fiona, who last semester, it’s great all of the different bills going proposed to convert Congress on AI, and then she used natural language processing to look at what the themes and the bills were. And she’s continuing work this semester. So again, this is a student with very little technical skills in the beginning of the semester. Here are some of our agent take behavior. This was our frontiers of AI class. This was for our students who have a little bit more coding experience. This was taught by my colleague Jon Chun, who I also collaborate with. And here is an AI powered appointment prioritization system for resource poor settings. And we also have Hannah Sussman, who created an AI virtual assistant for Student Disability Services. So there are all kinds of really cool projects. And we are now up to over 70,000 downloads globally. So we do have a very global extension here, over 1700 institutions, places like Carnegie Mellon, Stanford. I get emails from people saying, Where is the published paper, and I say it was an undergraduate project, and there is no published paper, but take it away, and it’d be great if you could follow up. I know NYU. We had one student who had a good friend die of an opioid overdose in Ohio, and he used traditional machine learning to try to predict who was most at risk, using Ohio data. And now NYU is working on those kinds of tasks. So there is an incredible amount of exciting work. This is the positive kinds of things and other ways that we can get involved. Right now there is a request for information on the development of an AI action plan. So right now, this is because the new administration is talking about AI dominance for the US. JD Vance also said the plan is AI dominance, and we don’t want any hindrance to AI development. We are in a global arms race here, and right now they are requesting information, so you have until March 15, and I would encourage groups to propose some commentary on that. This just happened yesterday, but I’m sure they’ll have more of these, where the UN was asking stakeholders about organizing global cooperative cooperation, because we really do need a global dialog, because it is essentially game theory with everybody competing. We need to have some kind of system where we’re not all running and trying to beat each other and dropping safety from the conversation. And this is just a paper I’ve talked a little bit with the group with meta. This group is now working on a follow up paper about open source. So open source is a really complicated question. Meta was the one of the big tech companies that decided to have Llama, which is open. And I put open in quotes because they’re all different kinds of open, and not everything is open. And there is a question about whether this technology is so dangerous, you know, that people could use it to do terrible things, and we can’t allow, you know, people to have it. And so some of the big tech companies have said that we should have no open source, that it should be regulated, and that they are the only ones who can keep us safe with their private AI. So we are now considering whether, with deep sea, we still believe this is an international group. You can see from all over the world, very, very complicated. You know, the advantage of the open source models is all different countries can train on their own language and on their own culture. But there are also huge questions about that. And I will give you just I will one parting example before I open it to questions. Singapore, a group in Singapore is working on creating seven different language models for Southeast Asia. Of course, there are many more languages than seven, but they’re starting with seven, and there was a question whether in Malaysia, it is illegal for a Hindu and a Muslim to marry. So should Malaysia’s large language model refuse to answer any questions about marriage between a Hindu and a Muslim? And these are the kinds of questions that we have to answer. Who decides who gets to make that decision. Should each culture have a different model that represents their culture. Indigenous AI is a group working with the indigenous cultures in Canada and the US, and they are building their own language model to try to preserve many of their languages and cultures. I’ve also talked to folks in France who want our help building many models to preserve 100 languages that are disappearing. And I also have a colleague in New Mexico who’s trying to build models to capture this culture and language. But we also have the indigenous community that is very suspicious for good reason about who might have access to their linguistic and cultural heritage and don’t want access to that. So there are very big questions here. You know, some countries definitely want their culture preserved. Other, other groups are extremely worried about what will happen with that, and this is part of a larger conversation about open source, because open source, you can fine tune it and train it on a new language and on a culture. So I’ll leave it at that, and I’m sure I hope you have a lot of questions and answers. Thank you. Questions. Yes, thank

46:30

you, yes,

Katherine Elkins  46:32

thank you. Do we need a mic? Okay, mic here. Okay.

Anonymous Question #1  46:41

That was fascinating. So you seem to have some confidence that the predictive AI, it can be useful. What? How do you respond to the people who say it’s that predictive AI is snake oil.

Katherine Elkins  46:56

Yeah, so I’m referring to, I you know that. Are we talking about the snake oil book? Yeah, yeah, very young, yeah. I mean, the snake oil book is mostly focused on older types of AI and mistakes. And, you know, I only watched one podcast where they were pressured a little bit, because they did add a little bit on early gen AI before it was this performant. But you know, one of the things that worries me, I’ll just tell you, honestly, is that the people who make these kinds of comments like snake oil, it’s what everybody wants to hear. I mean, I don’t get a lot of people excited to hear my message, even though I think it’s right. So they don’t really deal with more recent Gen AI. It is true that compass had all kinds of problems. That’s why we started focusing on this question. People are using Gen AI unauthorized at work all the time. So our concern was, what if you type some of this data into Gen AI, how well does it predict? And it actually does worse than chance, and there are reasons we suspect that is the case, so we should absolutely not have any judge, any prosecutor, any defense attorney using Gen AI right now for prediction, and I think some of that probably has to do with their attempt to tune it to eradicate bias, and they’ve over tuned essentially, but we don’t know, because we can’t look inside, but agents are actually incredibly helpful. If I had a defense attorney, I would absolutely want them to use an agent to model debate, to imagine what the prosecutor might say, to imagine what the judge might be thinking. They’re incredibly useful. They actually perform pretty well on predicting and I can tell you that I worked for the Boston Bar Association, writing the book for attorneys who take on the indigent cases. And I remember the first time I went to court and this kid said I didn’t steal this car just standing outside the club. And my friend drove up and I got in, and I was like, Oh, this poor kid, he got lot in the defense attorney said, Oh no, he’s good. That’s what they all say. And I would, I’m not sure they were doing their best, but I’m not sure I would want one of those free defense attorneys for my kid, unless they might be using an agent to help them argue a little better. So. And you know, when they did studies, we know cranky judges, judges who are hungry, more likely to send you to court. You know, so and actually compass in some not compass, but actually the New Jersey algorithm. I was talking with a colleague from data and society that actually performed better than the judges when they investigated. So we have to ask, compared to who, you know, humans are black boxes. We have to trust human decision making. Do we think that humans are better? And then there’s even a bigger question I’m going on a long time, so I’ll finish so we can get more questions. If we want better predictive models, we’re going to have to capture a lot more data. We capture a lot more data, then there are all kinds of questions about surveillance. So you know, there was the the Oakland Black community who said we don’t want facial recognition to work better on our black community, because we don’t want to be surveyed. So, you know, we’re in this tension to gather more data, to have better models. But then there are questions about whether we even want to gather that data in the first place. Okay, other questions.

Anonymous Question #2  50:40

So like, as these models are, like, gathering more data, would you agree that there should be, like, a standard of, like, documenting how this data is used, like, where it’s being contained, as well as people that are using it? Because, like, as you, when you in a presentation, you showed like these biases, especially like M generation, like, we don’t know, like, what these models are like, reading off of, or like, where these people are getting or where these models are basing off, uh, or making these outputs, like, what data is it coming for? Like, yeah. So like, just to supply the question, oh, like, would you agree there needs to be better data documentation?

Katherine Elkins  51:18

Yes. But again, it’s extremely tricky, and people don’t want us to see the data because of worries about lawsuits, and so we can’t look inside. And then there are all kinds of ways that you can counteract the bias of the model in the post processing, but we can’t see what is doing that. So if you have an open source model, you could potentially take the open source model and take some of those safeguards off, and then we would have the real model that we’re not allowed to inspect, so we’re inspecting it with the safeguards. But what if somebody takes those safeguards off? So yeah, this is a whole aspect. There’s so many aspects. There’s such lack of transparency. Again, I would say, like, hate the game, not the player, because all of the private company, you know, I’ve talked to people at meta, I’ve talked to people at IBM. Everybody has the best intentions, but we’re all in this system, and everybody has different incentives. But I don’t think we’re going to be able to look at that kind of data. On the other hand, if we get to small models like DeepSeek, you know, much smaller, easy to inspect, etc, if we don’t need as much data, which we’re really moving towards, we might actually be able to have that kind of thing and then using synthetic data. But yeah, there’s so much work to be done, but it’s tricky.

Anonymous Question #3  52:41

Yes, so I’ve been looking at this and drawing it, and I’ve noticed my students have tried it, and one of the things that comes up is this AI hallucination. It makes up a paper, makes up a journal, makes up authors. First year. Wow, this is great. How come I didn’t know about this? This is my area, and then you realize it’s just all fake. Why is that such a problem? Why can’t AI just check this and only give you real citations? Because that’s a giveaway that somebody’s using AI.

Katherine Elkins  53:11

There are models that are much better. So I always hesitate to give recommendations, because we are essentially helping students cheat. People tell me, but perplexity, AI, actually there are different filters. You can just do web, you can do academic and Perplexity AI will actually give you the sources for everything. When ChatGPT opened it up to web search, and so it knew that you wanted to use the web and base the response on facts, it did much better. But then people argued that it was going behind pay walls, so they had to take off that web enabled. So part of it is explaining to the system what you want, because some people use it for all kinds of creative endeavors, and for that, the hallucinations are fine, and it was trained on data that was fact and fiction, right? So you’re asking AI to essentially distinguish between writing fact and fiction, and they’re two very different directions. I will tell you, web enabled reduces that a lot. And then there’s something called RAG – Retrieval Augmented Generation. So even a year and a half, two years ago, worked for a financial firm, helping them implement their RAG system. And that basically means that the generation is coming specifically from your documents. And we reduced hallucinations down to less than 1% the problem is we’re using a general model. If you use the API, so you’re speaking directly machine to machine. You can control that somewhat. You can lower erase the temperature, which makes it more creative, less creative. Right now, I have a student working on trying to do nature writing, where we want facts and creative, and it’s a really fascinating challenge, because we’re trying to push the machine in two opposite directions at the same time. Yeah, but there are a lot of ways to deal with the hallucinations, but ChatGPT has had a problem with that, and it’s because there was a tension between creativity and hallucination. And I can tell you that I loved GPT two and GPT three, they were so much more creative because we didn’t have to worry about hallucinations. And now, with instruction tune fine tuning, they have helped keep those hallucinations under wraps, but we have lost a lot of the creativity in the models. So

Anonymous Question #3  55:33

sorry if I could. So in my graduate courses, I teach a lot of stuff that I have developed, and there’s a small part of the communities that use that. And then I see these submissions from the students where I’m trying to get them to use the things that we’ve developed that is a small part of the overall literature, and they just miss it. And I go, Wait a minute, you’re not using anything that I’ve taught in this course. What can they do, or what could be done, so that it does access this part of the literature and these techniques that have been developed but are not so widely used,

Katherine Elkins  56:11

I would encourage you to look into developing a RAG system. You can upload documents directly into the interface, but you can’t control the level of creativity or the less creative. So if you were using the API, you would want to look at temperature, and you would want to make it extremely deterministic, and then you would use a rag system, which tells it just bring facts from these documents. So it is absolutely doable. Will you get rid of every single hallucination? I’m not guaranteeing that, but you would do a pretty good job. Yeah,

Anonymous Question #4  56:48

the question of copyright seems really important, especially for the economic impact. Do you have thoughts on how the court cases are going to go and what that’ll look like in a few years?

Katherine Elkins  56:58

Yeah, I would really encourage you to look at Pamela Samuelson at UC Berkeley. She has some great YouTube lectures on this. I think she’s absolutely right that it’s going to be hard to win those cases, because rarely are you actually recreating something that’s very, very close. And when you look at the cases, even their example of copying is pretty eh and they have done a much better job with more recent models of trying to eliminate that again, it’s kind of an over fitting so if you give it the same data over and over, so they’ve tried to de duplicate data. Yeah, I don’t think it’s going to work very well, you know, I will tell you I have colleagues who think that copyright, you know, artists traditionally have stolen from each other, borrowed whatever you want. Some of my friends believe that copyright is part of the capitalist system. When I was at the IBM Tech Ethics Lab, I had colleagues who were trying to save whole archives of 100 years of photos from newspapers Cornell, I was told, has already gotten rid of them. One of the reasons they’ve gotten rid of them is the libraries are worried that somebody will go in and take a picture from a newspaper photo. These are called photo morgues, 1000s of 100 years of photos, put it online and be sued. And so libraries are throwing out 100 years of our history and of our photos. So, you know, personally, some people are saying, and again, it’s very complicated. I do not want to simplify. This is the last time that we have a chance to shape our AI, because soon, we probably won’t train on human created data. So one could argue that this is the time to suck all of our history and our photos into the system so it knows about us. And one could even say that this is the most important time to write and be a writer, because it is the last time that our writing will shape the systems. So that’s just a counter example. I’m not saying I have the answer. I am incredibly sympathetic to artists. I believe we should probably have UBI [Universal Basic Income] so that everybody can make art and not worry about making money. That’s how I’m punching on that, but I do believe it. Yeah,

Stephen McCauley  59:09

thanks for others, Kate. So I’m going to ask you to just step back and maybe all of us can think about this question as an educator as well as an AI expert, and the way you talk about the capabilities of AI and even now and then as we think a little bit into the future, and this idea that this is our last chance sort of thing. So I’m of the belief that we’re underplaying, that we’re that we don’t realize how significant this tech transformation is, if we’re to take it as seriously as I’m getting the sense that we should. So 50 years ago, roughly, WPI transformed its curriculum, it threw out the old curriculum, and it rebuilt its curriculum around project based learning. So we built our projects into our classes and student requirements, and all that’s been significant Are we at as significant a moment? That we need to like throw out the way we do university education and really rethink a new model where we harness all these capabilities and work with our students to learn and to educate them about how to excel and be an educated human as we live in this ecosystem of AI.

Katherine Elkins  1:00:23

yeah. I mean, I love the project based. You know, that’s what we do. Every single class has many projects where they practice and final projects. Our sophomores sometimes struggle with coming up with creative projects and completing them. And that, in my opinion, is part of it. You know, even for researchers like us, sometimes we try things and they don’t work. That is part of the process. So I would definitely encourage you to keep that it is very time intensive, as I’m sure you know, and it, you know, for us, it limits how many students we can serve. So at one point we expanded to 80 students in one of our courses for Kenyon. That’s quite big, because we only have 1600 students, right? 400 per class, and I was just not able to adequately serve 80 students projects in December. So it is a limitation. I do think there are huge questions about higher ed in terms of the expense in the US. I do think that the job market will impact us. I think you know, what is our purpose? What are we doing? I will tell you that using AI, there are so many cool projects you just need young people to think of them, and these are the projects that don’t make any money. So industry is not doing them, and it’s the thing I’m most hopeful about. But the larger question of how many students will go, who will pay for higher ed if it can’t guarantee a job, what will be our purpose de skilling is very real. I’ve seen it, even I struggle with it, you know. So, yeah, I mean, we’re looking at a big transformation. I like to say it’s like Pascal’s Wager, who always said, who said, you know, does God exist or not? Probably safer to believe. And, you know, I worry about people saying it’s not very good, or it’s snake oil, because I wish we had more people working with me on these questions, and my collaborators, like Jon Chun and when I flew back from meta in London after giving the ethical audit, we’re the only ones doing the semantic ethical audits that we can find there are all these companies that have cropped up that are charging money have no idea what they’re doing. I’ve asked friends at them, and the immigrations officer said, who else is working on that? And I was like, I have no idea who else is working on that, right? So there are so many things for us to be working on, and safer to think this is a massive transformation, and what are we going to do about it? And if we’re wrong, and it’s snake oil, although I don’t see it, and I’ve been in the field since 2016 I see constant progress. But if we’re wrong, then okay, but I would rather be working on it, and I just wish more people would join me. And I, you know, whenever people talk about the hype, it discourages colleagues from joining me.

Anonymous Question #5  1:03:25

Thank you. Wonderful topic and presentation. Thank you. I’m thinking about learning, and I’m wondering about the experience of our students using AI, or of ourselves using AI, and what we learn in that process. And I’m also wondering what you can share with us about research or your understanding of how human learning is different than the learning of machines.

Katherine Elkins  1:03:53

Oh, yeah, that’s a big question. I mean, for sure, they’re not learning the same way. So yeah, it is different. Last night, I was at dinner with a specialist in education and child development, and he was extremely concerned about the loss of error and failure and trial in education, and I think that that’s absolutely true as a developmental (right?) stage and but I think it’s much larger than just AI, because, as you probably know, we have some students in the room, these students now, at least my students, feel that there’s no room for failure. You know, they’re constantly at these crucial moments where, if they don’t get the interview, or they don’t get the internship, or they don’t get into whatever caught, you know, and they they don’t have that, you know. I mean, I like to say, you know, from my generation, it was Ferris Bueller’s Day Off, and high school was hijinks, and, you know, I mean, thank goodness there were no cameras. The fact that, right? And just none of that is allowed anymore. So, you know, to me, that’s a much larger question about the world that we have created for our young people. Yeah, and the difference between AI and human we’re still working that out, you know, yeah, that’s great question, yeah. Yeah. Students,

Anonymous Question #5  1:05:24

what is, what are students learning as they’re using AI, what it what’s their what’s their interaction with AI? Yeah, teaching them, yeah. Thank

Katherine Elkins  1:05:35

you. So I mean, we have an integrated curriculum that is tracked, where we start with an introductory course, and everything builds. We’re a minor, not a major. We call the concentration, because we don’t have a major, so we’re concentration technically, and the course was kind of our the way we started. Actually, I don’t want to take too much of a detour, but it’s kind of an interesting story was my husband did startups in Silicon Valley and sold his last startup to Symantec, and retired to little old Gambier, which is where Kenyon is, and stayed home for a few years to raise our three kids, and came out of retirement because he worried, as an engineer, that we were leaving important decisions to engineers alone. And so he was really the impetus for us rethinking our integrated program in humane studies and our interdisciplinary curriculum. And so, you know, the first course is really a roadmap, you know, having how everything fits together, everything you need to understand about technology to actually weigh in. And I would say one of the problems that I see is there’s a huge disconnect between theory and practice, which I would guess that you guys are doing a better job blending here than in a lot of places. So I hear people, you know, talking in these kind of high floating terms, of it has nothing to do with how a transformer model really works, right? And it’s really combining that technical understanding with cool applications. So one thing that we really focus on is a many model thinking, where whatever we’re working with, if it’s social network modeling, natural language processing, there’s so many different tools. If you only know one thing, you have a hammer, everything looks like a nail. A lot of what we do, and we have very few courses to train our students, is to give them that entire how everything hangs together, all the different models, all the different ways of looking at things. And so when they go to tackle a problem, they’re not pulling out the wrong tool in the toolbox. I don’t think you can really have a conversation about AI without understanding how it works. And sometimes it’s hard for me when I talk to faculty, because a lot of the conversation is about things that I don’t think are really, you know, informed so, you know, combining that conceptual our courses blend concept application and ethics in every single week, where we’re understanding the theories, we’re actually trying some of the code, and we are incorporating big ethical questions.

Anonymous Question #6  1:08:18

I’m curious, you know, looking at the balance between regulation and innovation from a global perspective, my impression is that the Europeans were advanced regulatory wise when it comes to computer privacy, that the Chinese have a more central authority for regulation and that the US is more of the Wild West. Which of the models do you think will win out in the balance between regulation and innovation? Yeah, that’s

Katherine Elkins  1:08:47

a great question. So I have a paper on archive called comparative global regulation with Jon Chun, and also colleague Christian Schroeder DeWitt, who’s at Oxford, and we look at China, the the EU in the US. So that goes a lot more. It’s it’s about five months old right at this point. So it’s not totally up to date, but you could definitely look at that. So the I have some thoughts on the EU AI act. I was part of a group of consultants where they like put you in this dark room and people are watching you, and I have no idea who paid me or who was watching me, which I’m embarrassed to admit, but it was early on, and I was kind of curious about the EU AI act before it came out. And my sense is that that AI act. And again, I don’t know who was who, yeah, who I was talking to, but they were focused on narrow AI and earlier AI models, and then ChatGPT came out, and didn’t surprise me, because I was working with it since 2019 but I’m telling everybody thought I was a crazy person when I was talking about generative AI. And then they quickly pivoted, and the AI act is not really adequate for Foundation Models. I can also tell you that the EU is now worried that they went over heavy on the regulation. And there has not been the development other than Mistral out of France, which is pretty good. And so they are backpedaling. And so the EU is backpedaling, and the US, I am just not hopeful at all at this point. You know, the US AI Safety Institute, lovely group of people, but it’s within NIST, which is with the Commerce Department, which is really to promote US industry. And you know, our model is very interesting, because one of the problems with the EU was it wasn’t really informed by industry. So you could argue that it wasn’t really technically informed. We have the big players in the room, so it is much more technically informed. But it also means we have the big players in the room, because we have NGOs, industry and academia all in the same room. And I sometimes feel that the academics, we get the you know, we don’t have the money, we don’t have the labs, we don’t have access. And then China is interesting. It’s very hard to know what’s happening in China, so we have to ask people, to ask people. And I hear different things, and one colleague just told me it can take up to seven months to get a model approved. But from what we are hearing, China seems to have heavy regulation on the big companies, but doesn’t enforce on the small startups in order to promote so while it looks heavy on the regulation, it is unequally enforced, at least, that seems to be what we’re hearing coming out of China. But nobody can tell you directly anyway,

Anonymous Question #7  1:11:52

maybe this is kind of related to what your the last question your answer is, the rhetoric coming now from the Trump administration around AI dominance, it reminds me a lot of the nuclear arms race and chemical weapons and that. And as I’m just and you, you talked a little bit about the now, the more about security, rather than about safety. And I’m wondering if you could just talk about that some more, what you think the trajectory might be, and what the implications are for the different kinds of uses or trajectories of uses of AI,

Katherine Elkins  1:12:32

I think it’s going to be pull out all the stops full steam ahead, and we will wait until something bad happens, and we may get lucky, but, you know, I am very suspicious of JD Vance saying we’re all about labor and we support our workers, because we are seeing already people being replaced. Tech companies are laying off. I talked to a startup out of Silicon Valley that paid, and this was maybe nine months ago, paid half a million dollars for an enterprise account for ChatGPT for 50 workers. So 500,000 for 50 workers, not cheap. And of course, since they outlaid it, they wanted to see and he said they made their money back in four months. So, you know, I only did the one consulting job for the financial firm that I do forecasting for, and quickly realized that we could potentially help them replace 50 back office writers, and became very uncomfortable with that. So, you know, we are looking at future of work. So that’s the first question. You know, even if AI doesn’t, we don’t have to worry about that. I think we have to worry about employment and future of work. And traditionally, when people lose power, we have political instability. So I do think we’ll have to figure it all out, because most of us are all in the same boat, but I’m worried about that transition period in terms of AI. I just think there’s not enough thought about what it means to have agents out on the internet without a human in the loop performing tasks, and then agents that are networked, that may have emergent coordination properties. And, you know, everybody’s saying it’s a tool, and we have a human in the loop. Well, no, it goes out and it does these tasks. It’s already bought the $31 eggs. So I find that super concerning. And and then we have tool use, so the agents can go and use tools to do things that they can’t do well. So I don’t believe that there will be a national focus on safety. I think it’s up to us to have that conversation, and that’s why I feel like academics talking about it now more than ever, I will say the existential risk, and I, most AI researchers, think there is a at least. Some tiny possibility that it will eradicate humanity, that conversation is marginalized and is not helpful, because you don’t get asked to the table. So I have chosen not to sign those, and I’m glad people are working on it, but that, you know, halt all development. Nobody’s going to halt development. And I have chosen to say, okay, they’re gonna make it anyway. So how can I help make sure that at least somebody is, you know, I mean, Anthropic went right before me at the US AI Safety Institute, and I had to say, in front of 100 people, yeah, you guys shut us down. So your security is great, but it means we can’t audit. I don’t know that it did any good. They don’t answer my emails. But I mean, it’s something right in

1:15:48

your answer that it’s both the

Anonymous Question #7  1:15:50

technology that eliminates jobs and is the technology for the surveillances and social control, oh, yeah, at the same time. So it’s creating, it helps create a problem that then it also helps

Katherine Elkins  1:16:02

to you. It also has to control. It does, and it might help with climate change, and it will probably help with, you know, curing many diseases. So it is really tricky because it will do good things. And that’s of course, yeah, but then we have accelerationists who just want to cure all, you know? So,

Anonymous Question #8  1:16:25

yeah, hi. Just wanted to say thank you so much for this informative presentation. I know there’s been a lot of questions that focused on academia, research, national security, safety, and even you talked about unpacking AI bias. I wanted to go back to, I think you had a slide that talked about, like, AI and relationships, and you had mentioned, like, WildChat. It’s been, I’ve been wanting to ask this question. So I don’t know if it might sound strange, but you I’ve read articles that focused on how, like, there’s been an influx of people that have been forming these relationships with AI, and even going further, as, like, humanoid, humanoid robots, and because they’re, you know, the opportunity of fostering these connections. Do you think this poses, you know, concerns for the future of maintaining these connections and or, do you think the world is shifting to this practice more frequently, and that people?

Katherine Elkins  1:17:35

Yeah. I mean, you know, humans, we are messy, we’re complicated, we’re emotional, we have bad days. I can see the temptation of an AI bot. I think what’s been most surprising about this trend is the women, because I never had a student willing to work on sex bots. Maybe was a little too crazy, but you know, those typically men were using, but the AI love relationships, we have a really large number of women engaging in. Yes, and that is fascinating. You know, I have been in talks where I have heard philosophers talk over and over about how it’s so much worse than a human-human relationship. And, you know, disparage it, and I’m less concerned with passing judgment. I’m lucky. I don’t need to be in a really, I may have a husband, but I think we need to say, what does this mean on the ground, understand these relationships better. And you know, there’s a lot of loneliness and unhappiness, and what does it mean to eradicate that? I don’t know. It’s hard for me to be against something that makes people feel more cared for.

Anonymous Question #8  1:18:56

And just to provide context, the reason why I asked is because of my MBA program. I did research on that and particular sex techs. So what do you think? So me, personally, I do think that, you know, people should feel comfortable making connections as long as they’re not violating any laws. But one of the concerns that I had was that, because there was an influx of people that are purchasing these, you know, robots that where does that information live? In terms of like people that purchase these services, where does it live? Who has access? Is there privacy concerns? Some people might find it to be like a moral, you know, moral concerns, like, there’s a lot that goes into that. I personally wouldn’t judge for that, yeah, but I was wondering, like, because there’s, you know, with AI, had made things more efficient, and, you know, you can, there’s like, this need for, like, getting things instantly, instant gratification. And. I think in the process, sometimes you lose the art of building, like, intimate connections with human beings. But what I wouldn’t judge somebody if you know they’re using like resources, like WildChat, I think.

Katherine Elkins  1:20:13

yeah, yeah. I mean, the privacy and data issues are huge. I absolutely agree with you. You know, replica. There’s a whole Reddit about replica, when they cut off access to memories of these, and people felt that they had lost their lovers or their girlfriend, you know? I mean, huge, huge. So, yeah, I mean, I mean, it’s here, yeah, it’s even,

Anonymous Question #8  1:20:38

I think in Europe, that they were even having discussions about, are these? Robots even considered electronic persons, right?

Katherine Elkins  1:20:45

Right? But that was interesting, right? Absolutely. Yeah, fascinating. I’m glad you’re also working on that, yeah, great,

Anonymous Question #7  1:20:52

great. Can I just add something to that? So maybe ironically, I read this article while I was alone eating dinner the other night. But there’s an article in The Atlantic, one of the last issues about solitude, yeah, and about that might be worth be interesting in this regard, about there’s sort of this notion of not loneliness, but solitude is increasing over the last 20 years in this country. And the argument was that this has real significant political implications, because it’s part of we form like very one on one attachments, maybe, whether it’s with AI or with people, and then also the internet and AI is allowing us to have connections with people very far away, but not sort of in the immediate community, and that’s where we have breakdowns of people being able to talk across ideological or political differences, other kinds of things, and resolve conflict and failure and all that. So it’s anyways, it’s a really interesting article that I think connects to this issue. Yeah, I think so too. Yeah, and I read it alone.

Anonymous Question #1  1:22:05

She has a question.

Anonymous Question #9  1:22:12

Yeah, thank you so much. Mine is a bit much different from this topic, but it’s also open as well. So so i You’ve talked a lot about, like, language models and, you know, facial recognitions, but I wanted to know, like, what you think about bias in the creative industry, like creative design industry, fashion and all of that. Because, for like, specific, for my research is basically like, biasing on the language generative AI, understanding different, you know, designs and different tier for like, different parts of the world. So I wanted to know, like, what you think about this, if you’ve done anything around that. And you know

Katherine Elkins  1:23:00

I have a student working on this right now. She’s from India, and the image recognition does very poorly with some pretty typical Indian clothing. She’s working in fashion, and we have some ideas about how to improve it. So, yeah, so actively, I don’t have any results yet, but we’ll have results in April or May. If you want to send me an email because we’re working on a pipeline to improve it. Yeah, yeah. Great question.

Stephen McCauley  1:23:28

Okay, I think we better wrap it up there. Thank you all so much for a lot of questions, and let’s thank Kate again first. Thank

1:23:33

you everybody. Thanks. You.