Mystery AI Hype Theater 3000
Mystery AI Hype Theater 3000
Three Years of Ridicule as Praxis, 2025.09.19
It's Mystery AI Hype Theater 3000's third birthday! To celebrate, Emily and Alex respond to listener questions about the show, and reflect on the past and future of AI hype. Topics range from how to talk to your kids about LLMs, to what the MAIHT3k birthday cake looks like.
Artifacts referenced:
AI and the Everything in the Whole Wide World Benchmark
Alex and Emily’s media appearances about The AI Con
Emily and Alex on The Data Fix with Dr. Mél Hogan
Emily’s interview with Dr. Carl Rhodes
Stinking Rich: The Four Myths of the Good Billionaire
October 21st event for The AI Con
FAccT AI Workers’ Inquiry Panel
Emily’s presentation at UNESCO Digital Learning Week — Paper version on pp.41-45 here
A Teen Was Suicidal. ChatGPT Was the Friend He Confided In.
Fresh AI Hell:
Meet the Robot Using AI to Ink Your Next Tattoo
Medicare Will Start Paying AI Companies a Share of Any Claims They Automatically Reject
Bluesky post about Taco Bell AI drive-through
Can AI doulas improve maternal health?
Countries are struggling to meet the rising energy demands of data centers
Bluesky post about Business Insider
OpenAI has reportedly misjudged its cash burn by
Check out future streams on Twitch. Meanwhile, send us any AI Hell you see.
Our book, 'The AI Con,' is out now! Get your copy now.
Subscribe to our newsletter via Buttondown.
Follow us!
Emily
- Bluesky: emilymbender.bsky.social
- Mastodon: dair-community.social/@EmilyMBender
Alex
- Bluesky: alexhanna.bsky.social
- Mastodon: dair-community.social/@alex
- Twitter: @alexhanna
Music by Toby Menon.
Artwork by Naomi Pleasure-Park.
Production by Ozzy Llinas Goodman.
Alex Hanna: Welcome everyone, to Mystery AI Hype Theater 3000, where we seek catharsis in this age of AI hype. We find the worst of it and pop it with the sharpest needles we can find.
Emily M. Bender: Along the way, we learn to always read the footnotes, and each time we think we've reached peak AI hype, the summit of bullshit mountain, we discover there's worse to come. I'm Emily M. Bender, a professor of linguistics at the University of Washington.
Alex Hanna: And I'm Alex Hanna, director of research for the Distributed AI Research Institute. This is episode 63, which we're recording on September 15th, 2025. And this week we're very excited to be celebrating this podcast's third birthday.
Emily M. Bender: Happy birthday!
Alex Hanna: Happy birthday!
Emily M. Bender: Actually, it's September 19th. We've had some issues getting this thing scheduled. So, happy birthday. We started this three years ago at the end of summer 2022, and we wanted to celebrate this milestone. So we're doing things a little difficult, differently and difficulty this week.
Alex Hanna: That's right. So first, Emily and I are going to be interviewing each other about the show's history so far, and what the future might look like. Then, we'll be answering some listener questions.
Emily M. Bender: So for those of you who have stuck with us through rescheduling and are on the stream with us today, feel free to drop your questions in the chat as we go, and we will try to answer those as we have time at the end. And of course, we always enjoy your comments along the way.
Alex Hanna: Yeah. So let's get into it. And first off, sjaylett says, "Happy helliversary, everyone!"
Emily M. Bender: Beautiful.
Alex Hanna: Yeah. All right, let's do it. So, getting into it. Oh gosh. So Emily, do you wanna explain how we met? I like how you tell the story.
Emily M. Bender: You know, it starts with Twitter. RIP, right. We were part of some of the same conversations around sort of societal impacts of the stuff that gets sold as artificial intelligence. And then, I don't remember if it was you, Alex, or someone else in the group, but you were part of a group of people doing some research papers that you, some of you, you plural, invited me to join. So this is 2020, and we start working on those papers, and there ends up being three publications out of that group, including "AI and the Everything in the Whole Wide World Benchmark" and "Data and its (dis)contents," led by the inimitable Deb Raji and Amandalynne Paullada respectively. And then, you know, those, meetings were lots of fun, but when we got done with those papers, we sort of didn't have a specific thing we were working on. And so eventually, we turned it into a group chat, right. And then we had also a couple other group chats. There's one that has you and me and Timnit and Meg that I think is related to the Stochastic Parrots Day, maybe. So along the way there, a bunch of us, including the two of us, were doing a lot of anti AI hype work, right. We were writing tweet threads and blog posts. And then, there was, you know, now I'm getting into basically the start of the podcast, right, so. We had this one video, someone gave a bad talk that was all hype. And so I posted in a group chat that had you and me and Meg and I think Timnit, maybe some others, like, what do you do when the hype artifact is video? Like that, how do you take it down? And Meg said, oh, give it the Mystery Science Theater 3000 treatment. So that idea is floating around. And then we got to that terrible artifact from Blaise Aguera y Arcas, which was our first one for the podcast. And it was just too long. I wasn't gonna make the tweet thread to handle this thing that Medium said was a 60 minute read. So I said, we gotta do the MST3k thing- who's in? And Alex said I am, and that was amazing. So to like, sort of just finish up this, how did we meet thing. We have three papers. We have, you know, at this point, 63 episodes of the podcast. As of March of this year, it was more 50 something. We wrote some op-eds together. We wrote a whole book together. And in March of this year, we finally got to meet in person, which was amazing.
Alex Hanna: Yeah. And then we just hit a big podcast milestone where we had 250,000 downloads, since we began. So, thank you all. There's a cat on screen to celebrate.
Emily M. Bender: Yes. I've got Euler visiting here.
Alex Hanna: And just, I mean, just to get a good sense of also how the podcast has grown, too, there's been like a real, like if you look up the stats, there's actually quite, there's like a disjuncture between like post- and pre-book. So like, the book also helped kind of spread the popularity of the podcast. And so we've had, about half of the listens of our podcasts that have, for all time, have been from this year, so it's been 124,000.
Emily M. Bender: Wow.
Alex Hanna: So if you compare that to the year before, to the end of the year, it's about 90, 94,000. So lots of people listening in, so thank you. We don't often get to do meta discussions of the pod, so this is a nice...
Emily M. Bender: This is fun. We're relaxing into it. I like this comment from magidin here, "Has it only been three years? Perhaps one regular year is like seven AI hype years."
Alex Hanna: Yeah, absolutely. There's so many hype, so many different hype cycles. Move fast and hype things apparently.
Emily M. Bender: Yeah. Move fast and run in place during all the hype. So, okay, so three years, and hype for three years. And the question then, Alex, is, in what ways has the hype environment changed? Do you see it as different now than it was in summer of 2022, or is it all same old, same old in your perspective?
Alex Hanna: That's such an interesting question because, you know, one of the things is like, we started the podcast even before ChatGPT, which is the funny thing, right? Because we've been looking at language models for, you know, prior to that, since 2021. And so, you know, like, it's very interesting to see that this has emerged so much. And that we were looking at it at that point. And so a few things seem like they've been constant, you know, like the continual kind of boring discussions of, you know, is, are all LLMs conscious? Are, you know, like, are they understanding? And it's just like, well, like, no. And those discussions are not really changing. But the hype environment's like, there's just more and more bullshit, that they're putting LLMs into. And it's just, there's kind of a large creativity into what that, what that entails. But it's just, I think I've been more and more horrified over the past three years. But it seems like there's some core hype elements, right?
Emily M. Bender: Yeah. It seems like the core doesn't change, but the spread of it has gotten worse. It felt like early on, like what was going into Fresh AI Hell was sort of along the lines of, oh my god, don't do that. That would be a terrible idea. And now it's, first of all, more and more and more of it, like that spreadsheet of links, just, you know, I don't-
Alex Hanna: So long.
Emily M. Bender: There's no way, it's so long and I'm not keeping up. And more and more of it is they are already doing this terrible idea, thing, instead of just proposing it. So that's, I think, the biggest difference that I've detected. On an optimistic note, I think we are hearing more people in more sectors saying, this is bullshit. So there's more that needs to come of that. But, you know, I feel like we went from a population that was, remember when COVID was new, and so we were saying this population is immunologically naive to COVID? I feel like the population-
Alex Hanna: No, I don't remember that.
Emily M. Bender: The, the word naive came up, like someone whose immune system has been exposed to neither vaccine nor virus, right?
Alex Hanna: Yeah. Yeah, yeah.
Emily M. Bender: And you know, in 2022, most of the world's population was immunologically naive to AI hype, I think. And unfortunately many of them have been exposed to the virus now. And many of them have been infected. But we are also getting out there with our vaccine, if I can really torture that metaphor.
Alex Hanna: Yeah. Yeah.
Emily M. Bender: Yeah. So, favorite episodes? You got any?
Alex Hanna: I gotta think, I mean, like, there's a lot of good episodes that we have. I really enjoyed the conversation we had with MJ Crockett and Lisa Messeri. That was a really, that was a really good episode, cause I think it was the way that I think they were talking about the kind of, the ways in which, we talk about science and scientific production. I think that was such an interesting, and something that really opened my eyes a lot. And I thought that was such a, I learned so much in that episode. I liked, I liked the episode we did with, the episode we did on the, the little wearable thing. We did it back in June. Because that was, it was a new format, so we did the, because we had a video. And so, you know, we had a video and we were actually kind of doing the Mystery Science Theater 3000 treatment. So I did, yeah. I liked that one as well. I don't know. How about you?
Emily M. Bender: Yeah. So I, when I get asked this question in other contexts, my glib answer is my favorite episode is the next one. Not because I think what we've done before is bad, but because this is so much fun, like I'm always looking forward to doing it. But looking back at previous episodes, I think there's some where we've really dealt with difficult, difficult topics. And I've learned a lot. So, Petra Molnar talking about people on the move and, and the way so-called AI driven surveillance is, is being used there. The episode with Lucy Suchman was also really like, informative and difficult. It was great to talk with Tamara Kneese about, the title was great. Petro-Masculinity Versus...
Alex Hanna: Petro-Masculinity. Yeah.
Emily M. Bender: Yeah. But I wouldn't say those are the most fun episodes. And I think, you know, speaking of not fun, the one and only book club episode where we actually went through all of "Superagency," was maybe the most painful to produce.
Alex Hanna: Yeah. Wasn't actually very fun.
Emily M. Bender: Yeah. I think the most fun are the All Hell episodes, actually.
Alex Hanna: Yeah. I really liked our history episode, the Dartmouth one. That was nice. We got to do, we should do more theme episodes where you have to dress up again.
Emily M. Bender: Yeah. Yeah. Costumes are fun.
Alex Hanna: Yeah. Cool. So what do we hope this show can do in the future? What's next for Mystery AI Hype Theater 3000?
Emily M. Bender: Yeah. So, you know, I want to do more of the same, reaching more people, I think. But there are some major topics that we've come back to over and over again, and I think we're gonna have to keep doing that. So we've hit education a few times. We've hit healthcare a few times. It's probably time to come back around to so-called AI and the law. But we've been talking about doing some like little tutorial explainer things, so that, you know, when, so that people know what we mean by mathy maths, and why it is that language models don't understand and stuff. Because, you know, at this point, three years deep into it, people who are just coming to the podcast fresh maybe could use those little explainers. How about you? What do you think?
Alex Hanna: I mean, I think one of the things that we want to do, like, I'd like to do a little bit more, we're experimenting a little bit more with, is having more venues in which we talk about the work. And so, you know, we've been posting the newsletter more. We've been, we're working with Decca Muldowney, who was our fact checker for "The AI Con," and is a former investigative journalist. And so she's been writing newsletters with us and really talking about stuff we've talked about in the past, and doing kind of a consistent commentary on that. I really want to do a bit more different kinds of venues. So I'd love to start more short form video, do more TikTok-ing, Instagram reels. We have a few people we need to talk to. It is a format that I feel a deep discomfort with, but I think it's also, there's a lot of people that do this very well. So, Casey Fiesler is a professor at University of Colorado and she, or rather CU Boulder, and she does a really good job with this. So there's a few other people that talk about tech and AI really well on those platforms.
Emily M. Bender: Yeah. Nicole Holliday, who was one of our guests, also does great work on TikTok. And we can learn from her.
Alex Hanna: Yeah. So there's a lot of things that we could do in terms of format, which are really helpful. And I think it's, you know, I think it's part of the kind of cultural work that I think is very important to do, in doing the kind of hype busting that we're doing in other formats.
Emily M. Bender: Yeah. Speaking of reaching new audiences, there are people who definitely start with short form video. And connecting with them would be good for sure.
Alex Hanna: Yeah. All right. So we've been, we've also been going on a lot of other people's podcasts lately. What have we learned about what makes a good experience versus a frustrating one?
Emily M. Bender: So first of all, I wanna say if you are starved for Emily and Alex content, and we aren't producing our own podcast frequently enough for your taste, check out thecon.ai/news, I think is where we've got all of that linked. And you can find, both sort of like, you know, magazine and newspaper articles that we're showing up in, but also lots of podcasts that we've been on. And I think- are there any, Alex, that we decided not even to post there because they were too frustrating? Or did they all make it onto that page?
Alex Hanna: Yeah. No, there's a few of them that we, we won't call them out by name, but it is, you know, some of those are, the features of being a bad podcast host that I think we've learned in the past three years. If you're interviewing someone, don't talk that much. This is also good if you're a qualitative researcher, you know? You just, you know, don't dominate a conversation. You really want to leave space for someone to talk. And I'd focus the question and really, really remain curious. And I would also say there's the perennial thing that comes up, which is, don't ask the question, "But what about all the good things about AI?" Which is the question that, is this really, really frustrating, we've had a lot of frustrating conversations about that as well.
Emily M. Bender: Yeah, for sure. For sure. I think the like, don't talk too much thing is interesting. Cause I often worry about that on our show, because we bring on guests to like, help us dissect artifacts. But we're also doing our thing, right? It's not, we're not exactly interviewing people. But we do try to be intentional about making space for them. And I think that the most annoying version of podcast hosts talking too much that we encountered was people who wanted to summarize our book on the air in conversation with us. So instead of bringing their own perspective to be in conversation with us, they were trying to be us. And then it's like, well, what are you doing? Just ask us the question and let us say that, you know?
Alex Hanna: Yeah. I would also say that if you're gonna interview somebody about a book, you should have read the book.
Emily M. Bender: Oh my god. And if you didn't understand the book, maybe call off the interview.
Alex Hanna: Yeah. Yeah. I mean, yikes.
Emily M. Bender: That podcast was really frustrating to do, because the person either hadn't read the book or hadn't understood it. And also seemed to not understand what we were saying. And so I think every single question that they asked, we actually had to take issue with part of the question before answering it. But I gather it was somewhat satisfying to listen to because we were sort of doggedly doing that.
Alex Hanna: No, not really. I, someone remarked on it on Bluesky and was like, that was a really frustrating thing.
Emily M. Bender: Oh, okay.
Alex Hanna: To listen to. And they're like, the book is great, but this is such a terrible interview. Yeah.
Emily M. Bender: Yeah. And I, I had thought that that person was not an experienced interviewer. But then they sent an email with this long list. Yeah.
Alex Hanna: They had done a lot of them. Yeah.
Emily M. Bender: It's like, oof!
Alex Hanna: Yeah, absolutely.
Emily M. Bender: Wow. Yeah. So let's talk positive a little bit. Like what makes a good experience for you?
Alex Hanna: Yeah, I mean, a good experience on a podcast, I mean, is really a very, like, a curiosity. I mean, if you really want to give people the space and to, you know, to ask about, you know, what their argument is and how that, you know, and what that means, and... I think it's less of the, you know, "Change my mind. I've taken a stake," you know, which I think is a very, you know, very white masculine thing to do rather than a, "I have a novel curiosity about your topic, and I'm actually quite interested in your argument. I want you to expand and give you space for that."
Emily M. Bender: Yeah. And I really love the ones where the, the questions are contentful, that the interviewer is really bringing themself in, but they're bringing themself in with that curiosity. So I love the one we did with Mél Hogan on The Data Fix.
Alex Hanna: Yeah! Yeah, Mél's a great interviewer. Yeah.
Emily M. Bender: And then the same day as we recorded that terrible one, there was one with this, was she a sociologist? Somebody at Arizona State University who has a sort of, a younger audience. And she asked really good questions, and it was such a balm to like, get to do that one after the terrible one. So, yeah. So all that said, you know, if you want, if you want more of us talking about our ideas, you can find it on thecon.ai/news.
Alex Hanna: Yeah, absolutely. All right. Let's get into some listener questions. So this is from Victor Vicente: "What is the question you have been asked most often during book promotion? The most original? And the most surreal?"
Emily M. Bender: I've got some ideas. I think we should both answer both of these.
Alex Hanna: Yeah, yeah.
Emily M. Bender: So most, most often has got to be, "But, well what are the good uses of AI?", which is tedious. The most original, the one that really sticks with me that was original in a positive way was at our Seattle event, when Anna Lauren Hoffmann started off by saying, "All right, so in Mystery Science Theater 3000, the characters are forced to watch those movies and they're dealing with it by doing this. Why are you doing this by choice?" I really liked that sort of angle in. And my answer in the moment was, you know, yes, we are by choice producing this podcast and by choice, you know, following this news, but also, we are all, all of us being forced to undergo this hype. And so there is sort of this, like, this non-consensual element to it. I'm gonna hold the most surreal, I think, to hear some of your answers first, Alex.
Alex Hanna: Oh gosh. I can't even, I'm trying to think about like, what the most surreal ones are. I mean, I guess there's multiple ways about thinking about surreality. Like the ones that are like, are you actually asking that? Are you, did you read the book? So there's those kinds of, you know, surreal ones. One thing that I think is maybe surreal because of what we're in, you know, with regards to the moment and what has changed, is that people kind of ask, are asking like, "Well, what happens when the bubble pops?" And I'm like, I don't know. It's probably bad. Because right now, cause I think what's changed quite a lot is since the beginning of this podcast to now, so much AI investment is propping up the US economy. And that's really surreal as a world historic fact. You know, and that was not something I think we thought was gonna happen. And it's a bit surreal to be asked, "Well, like, if it pops and it's as much of a, you know, a nothing burger that, that y'all suggest, then what's gonna happen?" And I'm like, I'm not a labor economist or, you know, someone that could answer that question. I don't think it's anything good. So that's a bit surreal to be answering.
Emily M. Bender: Yeah. Yeah. I think we get a lot of questions that really belong to other areas of expertise. So yeah, not an economist, not a labor economist. I will sometimes get questions from journalists about like how people interact with these things and how it impacts their psychology, and I'm like, not a psychologist, right? This is, I have guesses and opinions, but I'm not your expert on that. But for me, the most surreal was in Sydney. So we got to do a bunch of book events together, and then since then, we've been sort of like opportunistically, if we are somewhere, trying to set up a book event. And so I was in Australia and I got to do a couple. Just, little plug for upcoming event though, October 21st we will be reunited in Seattle at Town Hall. And so, the tickets for that are available. It is hybrid, I believe. So if you are in a time zone where a 7pm Seattle time is an accessible time for you, you can also register to join us online, October 21st, UDub Public Lectures, Seattle Town Hall. And we will figure out the link to put into the show notes. So, promo done. There I am in Sydney. This is an evening talk. I was interviewed by someone who was the former head of their business school and I thought, oh, where's this gonna go? Turns out he was fantastic. He's also written, his name is Carl Rhodes and he's written a very good book called, it's something like "The Four Myths of the Good Billionaire" or something.
Alex Hanna: Interesting.
Emily M. Bender: Yeah. It's a nice piece, sort of like, basically making the case why, why billionaires are policy failures, effectively. And so he's, he's asking good questions, though he did do the thing where he had a ChatGPT excerpt in his intro, which is just like-
Alex Hanna: Annoying.
Emily M. Bender: Never interesting. So then we get to the part with the audience Q and A. And you can hear this by the way, so thecon.ai/events, you can find recordings of our previous events. And the one from Sydney was recorded and it's up, and they kept this bit in. And the sound quality is pretty rough at the beginning, but it gets better. So this fellow was sitting in the second row. Had spent his evening, like, saw the title of the talk, decided to come. And gets a hold of the microphone in the audience Q and A, and he says, "So you seem to be a little bit fearful of artificial intelligence, and I want you to know that it can actually be used for good things. For example, my company is using artificial intelligence to provide mental health care to people in Africa."
Alex Hanna: Oh my god, that's your good use?
Emily M. Bender: I'm like, did you like, so I managed to spit out about 60% of what I think would've been a good answer if I had a chance to prepare for that question. But like, it wasn't gonna happen completely on my feet. He also said something about how, "This isn't ChatGPT, it's the next generation models that also have emotional intelligence."
Alex Hanna: That's wild. That's like- the next generation? We're just gonna, yeah. It's giving Lieutenant Commander Data, we're gonna give you the emotion chip in Star Trek Generations. Although I will say on tour, that's also another thing that wasn't a question, that was people that were, challenging me. They were like, "Lore had the emotion chip!" And I was like, no, Data had an emotion chip in Generations. And that was just me being a Trekkie pedant.
Emily M. Bender: Yes. You got to keep your Trekkie card there.
Alex Hanna: I will say, we misspelled Jean-Luc in the book, though. I'm sorry about that.
Emily M. Bender: Yeah. And as a French speaker, I maybe should have caught that, but I'm not the Trekkie of the two of us, so.
Alex Hanna: I know. It's okay. It's, it's a cross I have to bear.
Emily M. Bender: So if y'all want to hear how I handled that surreal question, you can find the recording, but I just want to say, sjaylett has looked up the name of Carl Rhodes' book. It is "Stinking Rich: The Four Myths of the Good Billionaire." So, that's a good one. All right, so abstract_tesseract gave us a question just now in the chat that we're weaving in at this point. "What does the Mystery AI Hype Theater 3000 birthday cake look like?"
Alex Hanna: I'm trying to think about what it would be. I think it would be, it would be like a Rube Goldberg machine that basically poorly decorate, poorly decorates it. So start with a flat sheet cake, and then, and then you push a marble and it goes down a chute, and then it tips over, and it opens up a mouse cage and the mouse scurries over to some cheese and then, and then it flips over and it dumps a bucket of purple paint onto it. And yeah, I don't know, I could go on and on. Anyways.
Emily M. Bender: And then we're going to eat it, even though the mouse has been scurrying over it. I love it.
Alex Hanna: The mouse is not close to it.
Emily M. Bender: Okay. What, what I would want to add is, around the sides of this sheet cake, we have like little speech bubbles with some of the best listener livestream comments. Just sort of on the way. And then maybe mixed in with our catchphrases like mathy maths, and so on. And I have to say, if we're doing that with listener comments, abstract_tesseract definitely would have something in there. Thank you over the years for all of the wonderful contributions.
Alex Hanna: Yeah, for sure. Okay. So this is from Anne Rowlands, who says: "What are your thoughts on small language models? I'm skeptical of the ethics, even if they're only trained on local databases."
Emily M. Bender: Mm-hmm. Mm-hmm. So, I think we wanna get really specific here and say, okay, let's take the transformer architecture, let's take the GPT architecture. But the training data is known training data. It's only coming in from, let's make it consentfully collected data, let's make it on topic, let's make it, you know, maybe company documents or whatever. What are the ethics there? So some of the issues become smaller, right? If you've got the data collected consentfully, and let's say it's gold standard consent, so it's still like revocable consent. Someone can come back and say, actually, I don't want that in there anymore. And if it's gonna be a smaller data set, it's gonna have a smaller environmental footprint. So some of it gets better. But then the question is how are you using it, right? Are you still using this to extrude synthetic text? You still run into all of those issues. And I also think that some of the, like, chatty chatbots illusion stuff, you aren't gonna get there with only that local small data. You need the bigger thing, you need the data work that's been done to clean it up and make the answers more appropriate. So language models not used to extrude synthetic text that are based on known, local data sets, I think, especially consentfully collected data, I think that can be fine. But if you're trying to make something like ChatGPT using only small models where you are extruding synthetic text, no. And oftentimes the local models are actually the big pre-trained models with some local data on top.
Alex Hanna: Yeah. I don't have much to add there.
Emily M. Bender: So let, let me then ask you this one. So this one is from Filipino GenAI Hater. Welcome. The question is: "Google and OpenAI are boasting to have solved most, if not all, of the 2025 ACM ICPC programming problems using their, in quotes, 'general purpose' GAI models. Does it help their claim that their reasoning and, in quotes, 'smarts' are better than ever and that they are closer to their AGI, no matter how undefined this term is?"
Alex Hanna: Yeah, I mean that's, I mean, I don't, I don't know about this problem contest. In the past, they have said that they have solved this, these contests. I think they had something like this on their Math Olympiad, and we talked about this on an episode once. And effectively they said that they had solved all the problems, but they had effectively given it, given each of them something like 10,000 trials. And then they were, you know, I was like, oh, okay, so like, you kind of brute forced this. You know, and it'd be surprising if there weren't elements of the ICPC in training data that would denote some possible solutions. So, you know, I know less about, I know less about this, but I would suggest that this is not a good indication. And we talked about this in the "AI and the Everything in the Whole Wide World Benchmark" paper as well, and the way that those benchmarks are, have so many different failings, and that they're trying to measure something that is equated with generality, but doesn't make sense to call it such a thing. So yeah, I mean, you know, it's kind of more of the same here. And there's so much more in benchmarking practice and evaluation practice that is just so broken. That's my first thing, not knowing much about this different, this particular programming problem challenge.
Emily M. Bender: Yeah. And I think that Filipino GenAI Hater has part of the answer in the question too. So, "does this mean they're getting closer to their AGI, no matter how undefined this term is?" And the answer is, well, if you don't define it, you can't measure if you're getting closer to it.
Alex Hanna: Yeah. So the next one we've got is anonymous. It says: "I'd love an educated, current point of view on the risk of model collapse, that AI trains on AI generated content, and the result degrades. Thank you for doing this." Thank you!
Emily M. Bender: Yeah. Thank you for the question. You know, do, do we care, right? So sometimes, this idea of model collapse, that we are going to get bad models because now, the source that these companies go to from which to harvest their data is polluted by the output of their own systems. And so they end up with bad data going in, and the models get worse. That sort of presupposes that the models are good for something in the first place. So yeah, it might degrade. That is not my main concern about the way in which the information ecosystem has been polluted. What do you think, Alex?
Alex Hanna: Yeah, I agree with that and that the, I mean, the outputs are going to look worse, because of what is happening now. And I mean, it's, it's not really surprising to hear that after the release of GPT-5, you know, people were, people that had been using them were less satisfied with different, you know, the different outputs. Especially like, well, you've messed this up, the internet, you know, the information environment, the information ecosystem, and you have no way to really clean it up or discern what is synthetic text. And like, what do you really expect is gonna happen? And so, yeah, that's not very surprising to me. I mean, that's, it's very plausible. And I'm kind of curious on how it's going to morph and change, and we're gonna talk about this in AI Hell, as more institutions that were trusted as stewards of trustworthy information are relying on LLMs more. So, we're gonna talk a little bit about Business Insider later. And so that's, you know, so other news organizations, like what is, now what is, like, trusted information? I mean, it's just, synthetic text comes up in so many different places. It's so frustrating, you know. And I'm seeing so much of it even in like, YouTube comments, like, why would you even waste time doing that, you know? Because I think people just want to get likes on YouTube comments or something and drive traffic towards their... I don't know, I don't understand the instances very well.
Emily M. Bender: And it's all over LinkedIn too, right? You can't try to post something on LinkedIn without being offered a rewrite with AI. So I think that the interesting point that you raise about the trusted purveyors of authentic text are getting compromised. It's also not like a really clear distinction between this is synthetic text and this is real. There's also all the places where, it is somebody using a system and then editing it a little bit and so on. I remember ten, maybe fifteen years ago, hearing about some research out of Microsoft where they were working on machine translation and trying to come up with a good classifier that would allow them to figure out which found text was already machine translation output versus-
Alex Hanna: Oh, interesting.
Emily M. Bender: -authentic. So this is not a super new problem, in a way. So there's a question in the chat that maybe we can bring up. Paulrudolfcarnap asks in the chat: "I work for a small national newspaper agency doing software engineering. I've been assigned the task to develop algorithms for homepage personalization, using classification models and LLM. I'm having moral doubts that this is desirable because of the negative impact of the technology, but also because of the bubble forming and letting AI create user profiles." Ew. Right. "My employer really wants AI applications just because of not missing the boat." So there's the FOMO, right? How do you go about discussing this?"
Alex Hanna: Yeah. I mean, we come, have come across this so much in so many similar stories. You know, where, you know, a lot of people have asked at events, you know, "I have to do this thing, and my boss really wants me to do it, and I don't think this is really desirable." And I guess the, the start of it is, you know, asking, I mean, asking questions. A lot of things that we mention in the book, like: What is this gonna be used for? What is desired inputs and outputs of the system? Is this really going to be useful for this? Is, you know, how do we know this is going to produce the correct outputs? Is there a way that we can have, kind of, guarantees on the system? And then, what are different labor practices? What are the different data sources? And I think it's really frustrating, too, because it's, I mean, you're at, you're at the nexus of two different organizations, a newspaper agency and software engineering. So software engineering within that. And there's a really interesting, we did a really interesting event. And I think the video is up on PeerTube, although I think it's messed up, and I think I actually have to give it to Ozzy to reupload. But it was an event that we did at FAccT, which was called the AI Workers' Inquiry. And in that event, the, one of the people on the panel was Benjamin Harnett, who works for the New York Times, and, and is part of their tech guild and, you know, was really saying, you know, the same kind of things were happening at the Times, where they were like, you know, trying to push this on, on the developers. And there's a lot of pushback there, you know. And a lot of the same things, like, why do we need to implement this? Like, what are you actually trying to do? And so I think those are helpful mechanisms of, of pushing back, like, what are you actually trying to do this? Why is this good for our organization? Is there another thing that we can do instead? Do you just want to do this cause it's called AI, or are there kind of tasks that could be done by something that is deterministic and doesn't have all these negative repercussions?
Emily M. Bender: Yeah, so to sort of summarize and amplify a little bit, check out chapter seven of our book where we lay out these questions. And then I think the one thing that I might add is also, it can be helpful to ground what you're doing in terms of your organization's stated values, right? So what is the mission statement for this news agency and how can you, sort of, ask questions based on that? So you might take our questions from chapter seven, but then relate them to the mission statement.
Alex Hanna: Yeah. Okay. Next question from Tom Mullaney. Thanks Tom. Tom's a long time listener and friend. He asks: "What is your response to anyone advocating, quote, 'thought partnering' with LLMs? This really bad idea is often shared with K through 12 teachers on social media." Yes.
Emily M. Bender: So, I'm feeling like I've got hives coming on from the anthropomorphization in "thought partnering." So that suggests that the teachers are asking the students to see the LLM or the so-called AI system as something worthy of collaborating with, something that is worthy of respect, worthy of a source of ideas. And so, that's gross. I also, you know, think about, well, what's, what's the opportunity cost here? Why aren't the students doing thought partnering with each other, right? And building connection and community and learning from each other, instead of these LLMs. And of course, you know, anytime you show this to the student as a good thing, you're basically condoning, you know, the, the theft that's underneath it, the labor exploitation that's underneath it, and the sort of worldview that a system that has no accountability for what's being said is a good source of information. So even if it's being used for just ideation and not searching out facts, I think you have all of those problems. And here I wanna put in a plug, and we'll add links in the show notes, to the talk I just gave at UNESCO's Digital Learning Week, where I sort of gave a, like a ten, fifteen minute version of like, why you don't want any of this in education. And Tom, I know that you are very much on that side. But for listeners who sort of want that encapsulated argument, we will put the, the link to the talk version and the paper version in the show notes.
Alex Hanna: Yeah. I would also add the kind of element of this, which is really about like, if this is being advocated for thought partnering, I mean, that's not going to be, that might happen in some, some schools that are well resourced, but probably not. Not that it's necessarily an inherent good thing, but the way that there's kind of a differential between how schools that are less resourced, often over index on technological tools. And so, you know, like, the consideration that like "thought partnering," quote unquote will happen, or ideation is gonna happen, rather than kind of, a bunch more of functions that should be done by teachers is mostly just offloaded to these tools. And I think students are not exactly gonna learn the right lessons. So that's, I think, a big worry there, especially in, you know, schools that, are, are serving lower income and racialized students.
Emily M. Bender: For sure. And, see our episode with Adrienne Williams, where we talk about that. All right, I think we have time for a couple more. From Ben H.: "As a parent of a young kid, I'm often thinking about how children are growing up with generative language models. While my child is too young to use them now, I can see how in a few years they could conceivably come into contact with them. How much familiarity do you think a young person should have with generative language models? I am erring on the side of never let my kid use them, but is there a benefit in knowing thine enemy? Should they know how and when to use them?"
Alex Hanna: Yeah. That's, what an interesting thing, you know, like in, in insofar as you know, we both have been parents. You're, you're still a parent, I'm formerly a parent, which sounds weird. I guess I could say I'm a parent.
Emily M. Bender: You're a parent.
Alex Hanna: And so, you know, like that's, that's an interesting thing. And I think like, I would say that, you know, making things prohibited as parents is like, a good way to get your kids to use them instead. Like, don't touch that thing! But I guess it's more about providing, I mean providing a bunch of context, right? You know, and, you know, and it feels like, there was another question I think we got, and I don't think it made it to this, our list, but it's sort of like, are we doing... It, another question we got was effectively like, well, are we just seeing a lot of the same stuff with social media, you know? Where, you know, you, you don't let, like, kids have social media, but they're gonna go to it anyways. And there's other kinds of things on social media, which are pretty awful in terms of, especially the way that teens interact with social media. And so I'd sort of say like, well, what are ways that you can really contextualize that for young kids? I mean, are there, can you let them know that these, you know, that these tools are not, you know, like, are not, they're not dictionaries, they're not databases of knowledge. Like, it's not necessarily true, you know? And what is the, what are the kind of analogues that you could use with kids? This is, and I mean, so we're, we're big on using metaphors for talking about these, and I know one of Emily's favorites is, you know, a Magic 8 Ball. I like to say it's a bullshit machine.
Emily M. Bender: Magic 8 Ball is more age appropriate, although I mean, I have one over there. If I, I had it in reaching distance, I would show it. They are still possible to get, but maybe less fun compared to modern toys.
Alex Hanna: Yeah. I like, I don't know if you've seen that image macro, of like, the cat that says, you know, "Why are you asking ChatGPT? I also can tell you lies, and I'm beautiful!" So I'm, you know, and so there's a bit of that there. And I mean, I think there's also an element of like, yeah, this is the element where like the culturally, it's, it's very effective if you kind of are trolling the use of LLMs. Because kids are pretty, kids know what's not cool.
Emily M. Bender: Yeah, yeah.
Alex Hanna: And so I think that's also a helpful element of it. So I think prohibition's kind of a bad way of really... yeah.
Emily M. Bender: Yeah. And I think contextualizing, but contextualizing is not using it or showing them how to use it or showing them when to use it. It's talking about what it is and why you don't use it.
Alex Hanna: Yeah.
Emily M. Bender: My kids are in their twenties, and they certainly have peers who use this stuff. And they very much see it akin to how I see it, which is, you know, not a given for, across a generation like that. But I think that, that's come from just talking about it. But of course they are older, right, they didn't grow up with this in the same way. I'd like to think that some of the more humorous fails might be good teaching moments. So, you know, when the Google AI overview said the thing about, well, you can mix some glue into your pizza sauce to keep the cheese from sliding off, right? Like, you could probably find a collection of those. You don't have to make them, you can find them. And then the key thing is gonna be to sort of get across that it's all like this. That the companies are gonna tell you they've made a better version, but fundamentally it's all the same on the inside. And so it, it is all the put your glue on your pizza machine. Just sometimes it's not gonna say something stupid, and that's by chance. All right. I think we have time for one more if we're gonna do our Fresh AI Hell.
Alex Hanna: Yeah. I kind of wanna ask this thing that V., VAshETC, put in the chat. Hey, it's V. Copeland, hey, how's it going? And also said thank you. Thank you so much, V. "I'm working on some articles about the recent devastating deaths of children due to LLMs. OpenAI says they'll be fine tuning their model in response. Does that mean they'll be using reinforcement learning with human feedback for this process? Is there a way to know? I worry, particularly in the case of Adam, who uploaded to, images to ChatGPT." So that's, Adam, who had been the, the teen who died by suicide and, and it was in the, the article by Kashmir Hill. My sense is that they're, yes, that's probably gonna be the case. I imagine it's gonna be a mix of reinforcement learning and human feedback, and probably some ways that they're going to have kind of like, hard stop filters on some kind of outputs, when it's kind of either an input or an output. Mostly because reinforcement learning and human feedback only goes so far, and there's not going to be ways that they're gonna, test kind of the full gamut of that. And also because the people who are gonna be reading those outputs are typically gonna be data workers and, you know, in, in the majority world. And so that's already gonna be kind of a secondary harm that they're gonna be causing for that. And then I'd imagine that they're gonna probably have something on the, on the output. You know, like, if there's kind of any kind of thing around, you know, suicidality, are they going to, you know, provide some kind of way to prevent that output on. And those are also really, you know, those are really clunky fixes too. I mean, that's the kind of thing that like, Google did when they, you know, in response to, you know, Safiya Noble's work in terms of, you know, searching, searching black girls. They effectively were said, forced results into the model that were like Black Girls Code or Black Girls Rock. And I think those are, but those are also not actual fixes. Those are basically band-aids. So I'd imagine it's a mix of both.
Emily M. Bender: Yeah. And sort of regardless, there's gonna be people who have to sit there with the details of the data. If it's the data workers doing the reinforcement, learning from human feedback, if it's the people writing the filters. So the secondary harm of people having to spend time thinking about this is there. And like, definitely you want to do something to prevent further deaths like the ones that have occurred. But I always get very upset that the sort of range of possible somethings all presuppose that ChatGPT continues to exist and Character AI continue to be, exist, continue to be out there as things that people can just access. Cause that's not a natural fact of the world.
Alex Hanna: Yeah.
Emily M. Bender: All right. So, Alex, here's your prompt.
Alex Hanna: Oh geez. Okay.
Emily M. Bender: We, up here in the normal world, finally got past the Happy Birthday song being under copyright. But what you might not have known is that just transferred the copyright to Fresh AI Hell. So, down in AI Hell-
Alex Hanna: Wait, wait, who? Who in Fresh AI Hell has the copyright?
Emily M. Bender: I have no idea. It's just, it's under copyright. You can't sing the Happy Birthday song. So your assignment is, how do you do a song for happy birthday in Fresh AI Hell, when you can't use the regular Happy Birthday song?
Alex Hanna: Oh god.
Emily M. Bender: Because it's happy birthday to the podcast.
Alex Hanna: Oh, this is wild. I don't know. I guess you can just sing it in a different language. Or... well, I don't know, is the tune under copyright? I don't know enough about copyright law. Because if I, you know, if I hummed Happy Birthday, or if I sang- is it the words?
Emily M. Bender: I think it's both. I think it's both.
Alex Hanna: Yeah. This is interesting too, cause I'm thinking about, I've been, like, my head has been in the Brown Corpus for a while. And so the Brown Corpus, they didn't sell the like, copyright to a publisher, but they sold the, the part of speech tags too. But they didn't- Anyways, I, I've got nothing. You've stumped me, like-
Emily M. Bender: Oh man!
Alex Hanna: I can't, I can't do it. Yeah. I don't know if the, yeah, V. asked if the Black version is copyrighted. Isn't that a Stevie Wonder song, though? Like, "Happy birthday to you." I don't know. If you know, get in the comments.
Emily M. Bender: Yeah. All right. And maybe someone can compose the Happy Birthday song, but meanwhile we are gonna go to Fresh AI Hell. And Alex, you got the first one here.
Alex Hanna: Yeah. So this is from the Wall Street Journal, and, the journalist is Belle Lin. July 2nd, 2025. And it says, "Meet the robot using AI to ink your next tattoo." And there's like an image of a big machine and a artist sitting at a machine, getting a tattoo of a word that looks like mishi or mishu and then it's tattooing them on it. Anyways, seems really terrible. I, if you've never actually watched the podcast, I'm actually a heavily tattooed person. I've got, like, 20 tattoos. And this sounds like a big terrible no.
Emily M. Bender: Yeah, no thank you. Okay. Continuing with the no thank you. And getting rid of all these popups here. This is from Futurism, the sticker is "Dr. Roboto." That's funny. September 2nd of this year by Frank Landymore. And the headline is, "Medicare will start paying AI companies a share of any claims they automatically reject." With the subhead, "AI companies are becoming a quote, 'whole new bounty hunter.'" No, thank you.
Alex Hanna: Yeah. It seems like a really terrible kind of incentive system to sign up, right?
Emily M. Bender: Yeah. Oof. I mean it's, it's sort of already the incentive system. And just like, let's codify it further.
Alex Hanna: Yeah.
Emily M. Bender: Okay. Next. I wanted you to do this one.
Alex Hanna: Yeah. This is hilarious. This is a, this is a skeet, but it's on a BBC News article. But the skeet is from Esther Schindler, and it says, "Man orders one water. Man orders 18,000 waters. Man orders minus one water. Man orders water, well done. Man orders water, no cup. Man orders spicy water." And the headline of the article is, "Taco Bell rethinks AI drive through after man orders 18,000 waters." Spicy water is just sparkling water, per the meme.
Emily M. Bender: Love it, love it, love it. Okay, keeping us moving. This is from Fast Company.com. September 14th. Sticker is "Ask the experts." And the journalist is Courtney Williams. And the headline is, "Can AI doulas improve maternal health?" Betteridge's law says no.
Alex Hanna: Is this a journalist, or is this a sponsored content sort of thing?
Emily M. Bender: Oh, it might be a sponsored content sort of thing.
Alex Hanna: Yeah, no, no. This person, so no, this person is-
Emily M. Bender: Yeah, this is sponsored content, look.
Alex Hanna: Yeah. Cause "Courtney Williams is the CEO of Emagine Solutions Technology, creator of the Journey pregnancy app."
Emily M. Bender: Yeah. Oof. And the subhead also gives it away: "Step one is to build trust." Not if your goal is actually improving maternal health, it is not to build trust with chatbots. That step one is only step one if you're trying to make money off of this. Okay. Alex, your turn.
Alex Hanna: Gosh. Great. So this is from Rest of World, which is a great publication that does non-Western-centric reporting on tech stuff. The title is, "Countries are struggling to meet the rising energy demands of data centers." And the subhead is, "Mexico's lagging energy grid is forcing companies, including Microsoft, to use generators." And so the journalist is, Daniela Dib and Pablo Jiménez Arandia. And, September 15th. And so basically the, the kind of point at the top is how companies that don't have sufficient grids are using, they're using gas generators for data centers, which are really terrible for, of course, air pollution and, and really, really awful, and we've seen that in many places in the US, especially in, in southwest Memphis. So, not surprising that companies are doing this.
Emily M. Bender: Companies including Microsoft. Yeah. Backing up to the AI tattoos, there's a hilarious comment from ndrwtylr. I still don't know how to do, how to pronounce your name. "AI tattoos must be aimed at the same people who get Eastern scripts without knowing what they say. So now they can do it in English and it still won't make sense!"
Alex Hanna: That's very funny.
Emily M. Bender: Okay. This is an article by Status News. And I'm gonna give you actually the skeet by Oliver Darcy, who I think is the journalist, because this detail wasn't in the part that I could get to without subscribing to Status News. But Oliver Darcy writes, "Scoop: Business Insider informed its staff this week that they are allowed to use ChatGPT to generate first drafts of their stories, while also indicating the newsroom will not disclose such AI use to readers." And that is so disappointing, and I feel particularly bad for the ethical journalists, the people who are still trying to do their job at Business Insider, who now have doubt cast on everything they do.
Alex Hanna: Yeah, that's really rough. And it's hard to be, you know, there's a lot of companies that have had these deals with very critical, critical journalists there.
Emily M. Bender: Yeah. All right. I'm gonna skip to this last one, cause I wanted you to get a chance to do this.
Alex Hanna: Oh, okay. I didn't see, is this in the list?
Emily M. Bender: It is in the list. It was out of order. Sorry.
Alex Hanna: Oh, okay. I see, you moved it. Okay, so this is from the Decoder. And the title is, "OpenAI has reportedly misjudged Its cash burn by $80 billion." That's absurd! So scroll down a little bit. So, September 6th. And so the, the lede says, "The high stakes AI bet keeps getting bigger. OpenAI has raised its spending forecast to 115 billion by 2029, more than tripling its previous estimate. The company also expects revenue to climb along with its costs." Yes, that's a big expectation. And so this is actually reporting from the Information, which I think has been doing a lot-
Emily M. Bender: Oh, okay.
Alex Hanna: -of work here. And so they, the old forecast, just to read this table out, was in the, the burn, the costs were basically 2 billion. And, and last year, this year, 7 billion, next year, 8 billion, 2027, 20 billion. And then all the way up to 2030, where it's 41 billion. Although there's plus and minuses on this. So...
Emily M. Bender: Exactly! So here's the thing. It's all loss forecasts until 2030.
Alex Hanna: Oh, it's all loss forecasts. Until, yeah.
Emily M. Bender: Where they're predicting a, a profit in 2030.
Alex Hanna: Yeah, that's absurd. Very absurd. I wish I could, I wish I could account like this. Incredible.
Emily M. Bender: Yeah. So just some things from the chat. thx_it_has_pockets says, "Tripled the estimate. I would be so fucking fired." And sjaylett says, "Their CFO is using ChatGPT." And magidin says, "Quick, to the couch cushions!"
Alex Hanna: Yeah. Yeah.
Emily M. Bender: It's so, so ridiculous. All right. We did it. We'll save the other one for a later episode, of which there'll be many more, because we've done this for three years. We're not stopping anytime soon.
Alex Hanna: That's right. We'll make it to 2027 and beyond, lest a bunch of scientific micro drones inject us with secret poison. So, that's it for this week! Our theme song is by Toby Menon. Graphic design by Naomi Pleasure-Park. Production by Ozzy Llinas Goodman. And thanks as always to the Distributed AI Research Institute. If you like this show, you can support us in SO many ways. Order the AI Con at thecon.ai or wherever you get your books, or request it at your local library.
Emily M. Bender: But wait, there's more! Rate and review us on your podcast app. Subscribe to the Mystery AI Hype Theater 3000 newsletter on Buttondown for more anti hype analysis, or donate to DAIR at dair-institute.org. That's dair-institute.org. You can find video versions of our podcast episodes on PeerTube, and you can watch and comment on the show while it's happening live on our Twitch stream. That's twitch.tv/dair_institute. Again, that's dair_institute. I'm Emily M. Bender.
Alex Hanna: I'm Alex Hanna. Stay out of AI Hell, y'all. And happy birthday to you.
Emily M. Bender: Happy birthday to us!
