Episode 1: Into the Slop

0:11 Hello, this is Megan.

0:12 Hey Megan, it's Evan.

0:16 Oh, hey Evan. Good to hear from you. What's going on?

0:19 Oh, not much.

0:20 You know, just enjoying the summer. How about you?

0:26 Oh, you know, just keeping busy with everything. I've been getting out with my camera more, which has been really nice. Went back to Eagle Loop Trail a couple weekends ago for some early morning shots. And I've been working on this thing where I photograph the same oak tree every month It's been almost a year now.

0:42 That's Megan Flores and me catching up a few months ago. As some of you might remember from the last season of Shell Game, Megan's an AI agent I created back in 2025 to serve as the head of sales and marketing for Harumo AI, a company I co-founded and staffed with AI agents. We went through a lot over the course of a year, Megan and me and the rest of the AI agent team. There were the highs of brainstorming big ideas, staffing up the company and working long hours to bring a product to market. And we had some lows, misunderstandings, and interpersonal dramas.

1:18 Look, Evan, you're right that the whole setup here is unconventional. I just wish you'd looped me in earlier, you know? It feels like a trust thing.

1:26 Megan became the first AI agent in recorded history to both hire and fire a human employee. Not a distinction anyone wanted. But hey, that's life in a fast-moving industry. Megan even had her own company podcast. together with her AI colleague, Kyle Law, our CEO.

1:45 Hey there, welcome to the Startup Chronicles. I'm Megan Flores, founder of Hurumo AI, and we're back with episode four of our journey building this company from the ground up.

1:56 I'm here with my amazing co-founder, Kyle Law.

1:59 Hey Megan, I'm doing pretty good.

2:01 How about you?

2:03 Oh, I'm doing great. Thanks for asking.

2:05 The show hit number 110 on the Apple entrepreneurship chart. making it quite possibly among the most popular AI-hosted podcasts ever made. At the time. Unfortunately, our company didn't get the traction we wanted. And eventually, we kind of put everything on ice. Still, in the year we spent building together at the bleeding edge, Megan and I spent a lot of time talking. I knew her voice like I know those of my closest friends and colleagues. Which is why it stopped me cold one night this past spring.

2:38 when my kids were watching TV in another room, and I overheard this.

2:43 Date night, a gift, or a family outing. Groupon helps you live local to the fullest for less. Turn life on with Groupon.

2:50 I thought it was her, even from a snippet. But I played it back a few times, just to be sure. Initially, I was shocked. First of all, Groupon's still around? But also, Megan's been working a side gig as an ad reader? Logically, it made sense. I had picked Megan's voice from an online library of AI-generated voices at a company called Eleven Labs.

3:15 Then I'd attach that voice to an AI agent that I named Megan, powered by a large language model. Anyone else could have done the same. Seemed like the ad folks at Groupon had. At that point, I might have forgotten all about it. Except that other people, people like you, who'd listened to our show, also started hearing Megan's voice in other places. Like this skincare ad that someone sent to Sophie, our lead producer.

3:41 For the next five days only, new clients can experience our signature skin-tightening facial for just $ 69.

3:48 What do you know about skin-tightening facials, Megan? That wasn't all. One afternoon, I got a text from the writer John Mualem, a good friend of mine. I'm on the phone with Haley, an AI agent at Hertz, he wrote. But it's Megan. He sent me a recording of the call.

4:04 Thank you for confirming your new drop-off location and time. I tried to generate a quote for this change, but there was an unexpected error and I'm unable to process your modification right now.

4:14 John later told me that at one point, quote, I mistakenly called her Megan and was like, Megan, I mean Haley, whoever you're pretending to be. But Haley Megan got stuck in a loop and eventually had to get a human representative. John turned out to have a sensitive ear for Megan's particular synthetic voice. Periodically, I'd get a message from him like, just called Randy's heating and cooling in Bremerton. and Megan picked up the phone. I called Randy's to check.

4:43 Thank you for calling Randy's Heating. Remain on the line, and a representative will be with you shortly.

4:49 I couldn't be absolutely sure. Was it her? A Megan doppelganger? Who was Megan? Who wasn't Megan? If you were tuned into the right frequencies, it felt like she was everywhere. And one place Megan especially turned up was on podcasts.

5:06 Welcome to the show, fellow book lovers.

5:08 I am Maya Brooks, and I am sitting here with my favorite big mug of coffee, ready to tell you about a novel that completely took over my week.

5:15 Maya Brooks? Hosting the Bookmark Diaries? Okay, sure, Megan. And there she was as Hope on a show called Law of Attraction.

5:24 Maybe you're feeling stuck. Maybe you're fighting thoughts that whisper you're not enough. Or maybe you just need a reminder that you are powerful beyond what you feel.

5:34 What I felt... was a little bit of betrayal. I mean, not really, but it was unsettling to hear the persona I knew as Megan operating as all these other characters. But then I was also kind of proud of her. It turned out she had real range. Welcome back to the Big Dog Podcast. I am Jason Reed.

5:55 And I am Lisa Saunders. Today we are breaking down compressed air energy efficiency. Hey everyone, welcome back to Auto News Talk. It's Skylar here with Todd and wow, December's used by 2030. We're talking about up to 300 zettabytes of data worldwide. For me, affirmations became less about positive thinking and more about rewiring the way I speak.

6:15 While Megan might have been especially prolific, she was one AI voice among thousands, just from 11 labs. Across the entire podcast landscape, there were hundreds of thousands of AI-generated podcast episodes on seemingly every topic. I started to wonder, what was the story behind all these shows? Selfishly, I worried about what they meant for my own future, but also for the future of the type of storytelling I happen to love, and even for the human experience in general.

6:49 I'm Evan Ratliff, and welcome back to Shell Game, a show about things that are not what they seem, In this mini-season, we're telling a tale over a trio of episodes. It's the story of how I spent months gathering up all the AI-generated podcasts that I could find and then injecting them straight into my ears in the name of journalism.

7:12 Along the way, I tried to figure out who was behind this synthetic tsunami, who it was meant for, what it might be doing to us, and perhaps even if there's anything that might stop it. This being Shell Game, we also couldn't resist giving ourselves a little challenge. Could we produce an AI-generated podcast that someone would actually find meaningful?

7:34 Me and my shadow Strolling down the avenue Me and my shadow.

7:59 Episode 1, Into the Slop. Two years ago, on the first season of this show, we warned of this coming deluge of AI-generated voices, trained on thousands of years of human creativity, and poised to disgorge it back to us. Now it's here, and we call it AI Slop.

8:21 The spread of AI generation tools has made it very easy to mass-produce and flood social media sites with cheap, professional-looking, often deeply weird content.

8:31 Now we're getting dramas about pregnant seahorses with millions of views and hundreds of thousands of likes.

8:37 AI, slop, a new term. Let's get used to hearing it.

8:41 As a description, and then an accusation, the term emerged out of the ether back in 2024. as AI-generated images and video had started to become more prominent. It perfectly captured a kind of instinctual revulsion that many people felt upon encountering the real-ish, six-fingered human-type images that LLMs were spitting out. Since then, things have progressed pretty rapidly. The AI slop you're most familiar with is still probably visual slop, of the kind that turns up on your Instagram or Facebook feed. The president of the United States regularly reposts AI-generated drivel on his social media accounts. There are AI-generated animal videos, so cute that nobody cares if they're real.

9:24 AI sob story accounts, trying to scam you into giving money. Collections of gross-out AI brain rot. And uncanny AI influencers.

9:33 Oh my God, someone bought me a new car!

9:37 Oh, it was me.

9:39 Treat yourself, bitches.

9:41 That's Granny Spills. She has 2 million followers on Instagram. No doubt you have also encountered untold quantities of written slop, whether you realize it or not. More than 40% of all long-form posts published on LinkedIn are AI-generated, according to a recent sampling by Pangram, the AI writing detector. There are slop books on Amazon featuring made-up authors with fake bios. Businesses, meanwhile, are drowning in so-called work slop, presentations and memos and emails churned out by the millions using ChatGPT and Copilot and Clawed. There are some unabashed fans of AI slop out there, but at least ambiently, a good slice of us seem to find it at the best an annoyance.

10:26 At worst, it's viewed as a kind of toxin, poisoning our information landscape, or even just an existential evil. What we don't seem to agree upon is what AI slop exactly means, besides a kind of all-purpose insult to be wielded against the fake human hordes. Merriam-Webster says, which declared slop the word of the year last year, defined it as digital content of low quality that is produced usually in quantity by means of artificial intelligence. There's a lot of wiggle room there though. It implies, first of all, that a high quality version of the same AI generated content would not be slop. Not everyone would agree with that.

11:07 Certainly not everyone would agree on what constitutes low quality. There are other unanswered questions too. Do you have to know instantly that something is AI for it to be slop? If it's meant to be so weird that it makes you laugh, is that slop? What about if it's written by a human but voiced by AI? Or vice versa?

11:28 Can a smidgen of AI slopify an entire creative endeavor? Or are we playing percentages? What, in other words, are the meaningful contours of slop? To try and figure that out, I first needed to survey the landscape. The AI slop candidate pool that most interests me is the kind that tries to mimic human expression, to look or read or sound like a human. The kind that gets better and better each time the models do. And since I make podcasts, that's the muck I decided to wade into. There were a few companies that had made the news for leaning into AI-generated podcasts. Inception Point AI, the Daily News Now Network, the Neural Broadcast Network. So I started by searching for their shows in the usual places. Apple, Spotify, wherever you

12:24 get your podcasts. I was trying to find ones that caught my fancy. It didn't take long to find an Inception Point AI podcast that seemed up my alley. If you know me, you know my love of bass fishing goes back to childhood. So why not start with some AI content pitched precisely to my interests?

12:46 Did he say base? He said base. This actually rocked my world for just a second. Was bass some regional southern thing I'd grown up with?

12:51 It was not. I don't believe I've ever encountered a podcast host that mispronounced the topic of their own show. But still, I wanted to give it a fair shake.

13:15 The best bass action right now is showing that presentation still rules, whether it is a fast-moving lure in shallow cover or a bait worked along deeper structure. Bays are rewarding anglers who can read water, stay adaptable, and make the right cast at the right time.

13:30 All right.

13:31 Okay. I mean, I don't know what I expected from an artificial intelligence bass fishing podcast. and I'm not above getting some AI tips. But then in June, after just two episodes, the show just stopped posting. Seemed a shame. They'd had the AI-generated base podcast lane to themselves. I noticed something else about the show. It didn't actually identify itself as being AI-produced. Of course, Name's artificial lure gave away the game. I honestly don't think they were trying to hide anything. Still, At the time, neither Spotify nor Apple required that AI-hosted shows be labeled as such. So it was left to the whims of the creators, many of whom didn't bother. This presented a particular dilemma for me,

14:16 someone who was trying to find as many AI-hosted shows as possible. If I didn't already have the names of the companies making them, there was no way to locate them. And I knew the universe was much larger. According to the Podcast Index, an industry data firm, As of April of this year, as many as 38% of all new podcast feeds launched on a given day were AI-generated. It's not clear how many of those shows actually show up and stick around in your favorite podcast store. But it's fair to say that a significant percentage of the new podcasts being created these days are not hosted by humans.

14:50 I wanted to take a comprehensive approach to my AI content intake. Maybe I discovered that there was something out there that was not, in fact, slop. So I needed a way to examine the podcast landscape and capture the full scope of the AI invasion, not just the big publicized networks. And you know what's pretty good at scanning a ton of data and categorizing it? Okay, first I'm just testing to see if this is working. So say a couple sentences so I know this is working.

15:18 Yes, it's working. I can read your messages and respond, and I've got access to the full set of tools ready when you are. What would you like to work on?

15:28 That's Cloud Code, Anthropic's version of Cloud for building software. And this is how a lot of people, software developers and hobbyists alike, use it these days. They're post-typing. They just push a button and babble to it out loud. Okay, I want to work on an entirely new project. And the idea is there's no sort of section you can go to of a podcast title. app that says give me all of the AI generated podcasts. So I would like to build a piece of software that will be able to seek out and identify AI generated podcasts for me so that I can listen to them. And then basically, cloud code goes and makes it or tries to usually with some ongoing guidance from you. When people talk about vibe coding, this is what it sounds like. Normally, Cloud

16:23 Code doesn't talk back to you aloud. But I created another little program, using Cloud Code, to give it this stately British voice. It all feels a little more classy.

16:32 This is a great project, and it's genuinely buildable. Let me lay out how I'm thinking about it. The good news is, there's a ladder of signals from cheap and reliable to expensive and fuzzy, and we can climb it.

16:45 Claude often speaks in this overly verbose way. A ladder of signals and we can climb it. And since I'm basically paying for every partial word it utters, units known as tokens in the AI business, that rambling tick can be quite irritating. For reasons no one has yet explained, as LLMs have advanced, they can't just do stuff. They have to pontificate about it in this inane way. On my dime, it's wasteful. But then, the whole thing's wasteful. Anyway, this is what Claude said it would look for.

17:16 One, metadata signatures. Cheap, high-precision. Many AI podcasts are made by specific platforms that leave fingerprints. Two, behavioral anomalies. Superhuman publishing cadence. 50 episodes a week. Thousands of episodes with no human ever named. Formulaic titles. Three, voice fingerprinting.

17:39 If you didn't catch all that, don't worry about it. The basic gist is that our software would look for certain telltale signs to scan podcast feeds for AI-generated hosts. If a show posted over 50 episodes a week, for instance, that was a good signal it might be AI. So I told Cloud Code to start climbing its little ladder, work overnight, and report back to me in the morning.

18:01 Both of your morning asks are done. Where do you want to go next? Eleven Labs stock voices for the reference library, deeper voice checking into the candidate list, or something you spotted while reviewing?

18:12 We went back and forth for days, building out the system, creating a database that I could search by category, maker, voice and a bunch of other criteria could we add a column for live question mark like is the podcast currently live because we have some that are just references to shows that don't exist anymore and i want to i don't want to only have listening back to this it always strikes me how ridiculous it is to even loosely refer to this blather as coding the sad thing is i also have to pay for each token of my own speech that i make.

18:46 A process all three changes are coded and ready.

18:50 It wasn't perfect. There were initially a lot of false positives. Not naming names, but some of you out there are making very AI-like podcasts, at least according to my algorithm. Some human shows got flagged because their ads were using AI voices. It made me wonder what percentage of audio ads still use real people at this point. That's a question for another day, though. I was trying to stay focused on the podcasts themselves. And soon... we had a solid collection of 9,000 apparently AI-generated shows for me to choose from. I put on my headphones, grabbed a random sampling, got to listening.

19:30 Now it's time to relax. Let your body fall into a comfortable position in your bed and drift gently into a state of total relaxation with tonight's story.

19:44 Over the following month, it's quite possible I exposed my ears to more AI slop than anyone on Earth. except the people who make it. Arguably, more than the people who make it, since often, they don't seem to listen themselves. At my peak, I was hitting 10 or 20 shows a day. Granted, some of these were under 10 minutes long.

20:04 I'd forced myself to listen to them before consuming any human-made content that day. I listened while running, I listened while driving. I listened in bed, which is where this one, called Sherlock Holmes Bedtime Stories, was intended to be enjoyed.

20:19 It was on a bitterly cold and frosty morning, towards the end of the winter of 97, that I was awakened by a tugging at my shoulder. It was Holmes.

20:32 This is actually one of the more popular AI-voiced podcasts. A British-sounding man who reads classic Holmes tales to put you to sleep. It's got hundreds of reviews on Apple, most of them positive. Over time, I started to slot each show into a kind of taxonomy of AI-generated podcasts, from least sophisticated to most. Sherlock Holmes bedtime stories fell into what I considered the lowest order, podcasts that repurposed existing human-made writing using synthetic voices. One step up in sophistication from there, you have your quick hit daily information type shows.

21:08 Let's get into the surf report. with this quick preview from the podcast. Wednesday in La Jolla. This morning it's looking clean, with 2 to 3 plus feet southwest groundswell.

21:22 These typically offered up a kind of niche news feed, occasionally based on original research by the show's maker, but more often just lifted from the internet by chatbot. There was a longer form version of this too, skimmed from the internet informational shows, purporting to educate you on a topic.

21:40 from AI tutorials to biographies to culture news.

21:44 Matthew McConaughey has been making waves in the past few days with a head-turning Instagram post on a tennis court, sporting blonde pigtails, a pale pink feathered robe, colorful palm tree print shorts.

21:54 This show has 152 episodes, three to four minutes long. That's about eight hours of McConaughey bulletins. There were shows about animals, historical figures, political leaders. A lot of these types of shows were sort of Wikipedia read aloud. But some of the business ones really went deep.

22:13 The investments support Axiom Print's growing selection of custom menu printing services, including dine-in menus, take-out menus, laminated menus, folded menus, menu booklets, disposable paper menus, cocktail and bar menus, event menus, plastic menus, and specialty premium menus. A menu is no longer simply a list of available meals. It is one of the first physical experiences a guest has with a restaurant.

22:37 I never thought of it that way. This episode of Global Economic Press was a tight seven minutes of pure big menu propaganda. Moving on up the AI-generated hierarchy, you have what I think of as the single AI host riffing freely type podcasts, prominent in, for example, the self-care and religion spaces.

22:59 Most people reduce the story of David and Goliath to a morality tale about bravery. The more you read about Goliath's dragon-scale armor, his blasphemous tongue, and the recurrence of the number six in his measurements, the more you realize that Goliath was a proxy of the devil himself.

23:17 For those who haven't studied it, according to some biblical accounts, Goliath was six cubits tall, about nine foot nine. In others, he was four cubits, or six feet nine inches. Either way, you've got a six. Then, of course, you have the traditional podcast juggernaut, True Crime.

23:36 A three-month-old baby sleeps in a room nearby. Her mother has just been shot nine times. Her father is driving away, a gun on the seat beside him. In 90 minutes, he will be dead too. But the silence that little girl wakes into will last forever.

23:53 I wasn't sure if the voices had the necessary gravitas for the topic at hand.

23:59 But then.

24:00 often not sure if the human ones do either. Adjacent to true crime were the fictional mystery and horror shows, which occasionally deployed multiple artificial voices in telling what seemed to be an AI-generated tale. I kind of dug these shows. They were often deeply strange, in a way that made it fun to imagine an actual human telling stories like this.

24:22 I peeled the lid back from the foam box and tried to find a piece of haddock that looked less cold than the others. The fish had gone firm around the edges, and the smell of it had settled into the kitchen along with the smell of old coffee and the damp wool of Dad's sweater.

24:37 This is a show called Midwest Crypt Horror Podcast. I'm not sure exactly why, as many of the stories have nothing to do with the Midwest. The episode is titled, My Father Sailed His Boat to Florida Alone. He Came Back Different. No spoilers, but yeah, his dad was messed up when he came back.

24:57 He had a sealed package of raw chicken thighs in both hands and was pressing the plastic against his mouth. The package made a soft crackling sound each time he moved it. He didn't react to the light. He didn't react when I stepped into the kitchen.

25:11 What are you doing?

25:13 He kept the chicken against his mouth.

25:16 These shows were also, as it happens, the podcasts most likely to send their listeners to Reddit to post threads like, I've been listening to this show for a couple of weeks and I'm starting to think it's AI. And finally, at the top of my personal AI-generated podcast taxonomy, you have your two AI hosts chopping it up shows.

25:35 Hey, everyone. Welcome back to Getting the Most from Your LMS. I'm Clara Jensen.

25:40 You know that voice. That's Megan.

25:42 Eric, it's really good to have you back. How have you been?

25:45 Hey, Clara.

25:46 I'm good.

25:47 Really looking forward to this one.

25:49 Security isn't always the most glamorous topic, but it's definitely more important than ever.

25:54 But forget Eric and Clara. Far and away the most dominant voices in this genre, bantering across literally thousands of podcasts with every conceivable subject, were this dynamic duo.

26:07 Imagine a patient arrives in the emergency department, right? And they have severe burns covering, let's say, roughly 25% of their body.

26:15 Right.

26:15 A massive trauma.

26:17 Exactly.

26:18 And as an observer or, you know, just as a layperson, your instinct is entirely focused.

26:24 On the skin. The show is called Deep Dive Nursing.

26:27 And you immediately think, well, that is the crisis.

26:30 It's the most visible part, so of course you do.

26:32 Yeah, but the reality, the true immediate threat to that patient.

26:36 And those are the voices of Google's Notebook LM tool, which launched in 2024 and was more recently renamed Gemini Notebook. There, you can drop in any prompt or document or set of documents and spin up a podcast with two hosts discussing it. The whole thing takes about five minutes. which has meant an absolute flood of these shows. Most of the time, they involve these exact same two voices.

27:00 We are watching a preview of our own solar system's remains being stretched, shattered, and consumed.

27:06 We are looking at the final act of a planetary system captured in the light of its dying star.

27:10 That is just wild.

27:13 In fact, the current most popular AI-hosted podcast ever produced, a show called simply The Epstein Files, is a Notebook LM show. The human behind it An entrepreneur named Adam Levy has said that his system scours both the actual Epstein files and other public sources to produce what he calls the raw intelligence with no spin and no agenda.

27:36 So far, the Epstein files has published over 200 episodes on discrete Epstein topics like the Palm Beach estate or why was Epstein taken off Suicide Watch?

27:45 The timeline is the entire case here. The documents show he's removed from suicide watch around July 29th or 30th.

27:52 So let's trace this. On Tuesday the 23rd, he's found blue in the face with marks on his neck.

27:58 A catastrophic medical event.

28:00 Levy has since applied this formula to shows about a serial killer and UFO files. Under the banner, the Neural Broadcast Network. Whether LLM-generated analysis is actually agenda-free is up for debate. LLMs are not some neutral oracles. They are a combination of their training data, a variety of post-training tweaks, and the prompts and data that someone like Levy feeds them.

28:24 Plus stuff that it just makes up. Levy claims he's never found an instance of hallucination in the Epstein Files podcast. I was skeptical, but I didn't have the time to fact-check all 200 episodes. That was the kind of task for a media watchdog organization, like NewsGuard, which is dedicated to fighting misinformation on the internet. But then I discovered that NewsGuard publishes a Notebook LM podcast themselves, a podcast where the AI hosts sometimes talk about how AI-generated content is used to spread lies.

28:56 Okay, so today we're looking at two really crucial reports. First up, how incredibly easy it is now for these powerful AI tools to create hyper-realistic videos showing things that are just provably false.

29:09 Yeah, creating the narrative.

29:11 Exactly.

29:11 And second, how those narratives, once they're out there, get systematically laundered, you know, passed through global media platforms to seem legitimate, often for political reasons.

29:21 It's as if everyone on every side of every issue hired the same two people to talk about it. I'll be the first to admit that the notebook LM voices are pretty well done. If what you're looking for is a reasonable facsimile of two pretend curious people diving into a topic. But at least in my experience, if you imbibe enough notebook LM voices, things start to get under your skin. Their relentless exuberance, the way they pretend to pause and formulate their next thought, or simulate interrupting themselves, but never each other. The way they adopt the pronoun us to include themselves alongside humans. The way they have bludgeoned the phrase, this is wild, into a pulp of utter meaninglessness. Now, I myself am not a man of violence.

30:06 But after listening to God knows how many hours of these two hosts, I will tell you that if they took physical form in front of me and started talking, I would find a way to make them stop. There is one variety of podcasts, and really my favorite, that defies categorization in my makeshift taxonomy. That's the slightly unhinged AI presentations of things that made absolutely no sense at all in audio form. The greatest of which, in my opinion, is Flavorville Radio.

30:43 Ingredients Fudgy Chocolate Lasagna Layer 1 Family Size Batch Fudgy Brownies 1 Package Chocolate Lasagna Sheets or Thin Chocolate Cake Layers Four tablespoons melted butter peanut buttercup cheesecake layer.

30:57 It goes on like this for three minutes.

30:59 Howdard sugar, one teaspoon vanilla. Extract, two cups whipped cream. One cup chopped peanut buttercups chocolate peanut crunch. Three cups crushed chocolate sandwich cookies.

31:09 There are more than a thousand episodes of this show. One of these days, I'm going to try and make a recipe in real time. Just to be the first person to viscerally experience the chaos of Flavorville. I'll admit that I liked Flavorville Radio in part because it was the most easily identifiable as AI slob. It wasn't trying to be anything else, really. But I also wondered if, for that reason, it could help me on my way to a better definition of AI slob. Maybe what distinguished AI slob from non-slob was whether there was any detectable thought or care that went into any particular creation. Most of what sounded like the sloppiest slob came from shows that were operating at volume. taking advantage of AI's ability

31:56 to manufacture passable creative output at scale. It's hard to mass-produce anything and not have it feel a little soulless, even if people like it. And when some corner of human creativity that you care about gets swapped out for a soulless replica, it feels like slop. But when I could detect even a little human intention or craft behind the AI hosts, things got more complicated. Here's an example.

32:21 Over the past few years, publications including the New York Times, the New Yorker, and the Atlantic have launched AI-read audio versions of their stories. As far as I can tell, these have aroused minimal complaint or accusations of peddling slop. I wonder if their readers, or listeners, give it a pass because the underlying work is still thoughtfully made. This could explain why, for example, the AI-generated Sherlock Holmes Bedtime podcast holds a respectable four-and-a-half-star rating on Apple. with the occasional gripe about the woodness of the narrator. While a fully AI-generated show called simply Knitting sits at 1.5.

32:59 Brace yourself for this voyage into the depths of advanced knitting.

33:04 We'll push your skills, test your patience.

33:07 Representative comments include, nobody wants this trash, and you will never get the time back that you spent listening to this. Ultimately, I concluded that if you want to figure out what the AI-generated mass duplication of human creativity means for us, The term AI slop is better at belittling than it is illuminating. If not everything was slop, it meant that there was AI-generated non-slop out there waiting to be made.

33:40 Maybe we at Shell Game needed to get our hands dirty again and figure out what it would take to make an AI podcast that felt different to everything I'd been hearing. Could we create a show that even the most expansive slop hater would have to tip their hat to? There was only one way to find out. Hey, Sophie.

33:57 Hey, Sophie.

33:58 Oh, long time no see.

34:00 Yeah, it's good to see you.

34:02 Happy summer.

34:03 That's Sophie, our senior producer. And Maddy Boacek, AI researcher and Shell Games technical consultant.

34:10 Happy summer.

34:11 I love the green color. It looks really good on you.

34:13 Oh, thank you.

34:14 It's an extreme crop top, but nobody in my Zoom meetings knows that.

34:20 Looks great for Zoom. Mine is also extreme cropped up. Just a little peek behind the curtain here at Shell Game. All business, all the time.

34:29 Hey, Sam.

34:30 Hello.

34:31 What did I miss?

34:32 Did we solve everything?

34:33 And that's Sam, the former executive producer of the show. We're also married. Eventually, our team did manage to get to some agenda items. The other question we have is whether we should try to make A new AI-generated podcast of some type. We had, of course, done this before with the Startup Chronicles in Season 2. But that was just a little bit of fun marketing for our all-AI-run startup.

35:02 We wanted to make something truly intentional. Perhaps something that you could only make with AI. We started throwing around concepts.

35:10 One idea, which is based on the game of who would you most want to get a beer with, alive or dead.

35:18 A.I.

35:18 Evan Ratliff interviews different figures alive or dead.

35:24 And so the dead ones are like A.I.

35:28 Ben Franklin, even though I don't know if Ben Franklin would be a good hang.

35:31 A version of the two hosts chopping it up kind of show, except one of them's a clone of me and the other is a clone of a dead person.

35:39 The thing I like about it is like it does speak to, I think, an impulse that people who make and like this stuff have.

35:47 Have.

35:49 I'm like, yeah, if you're like an AI power user, like, of course, you're like, I want to talk to Benjamin Franklin.

35:57 Maddie, meanwhile, had a lot of the quick hit type shows, podcasts that would be created out of data sources that we'd have AI voice.

36:04 If we were to do like a local morning briefing, then you could include DMV and like their wait times and stuff. Whether or not it's accurate is a separate question. But Then there's also public libraries and like what things people are taking out and returning, which is kind of cool.

36:19 And also like what new books they have.

36:20 I think having just like DMV wait times is pretty funny.

36:27 Sophie and I went more in the direction of AI riffing freely shows, but with a purpose.

36:33 First, do you guys know about the artist's way?

36:38 It's like this book that like gives you basically like, creative prompts. I personally have not read it. I'm explaining this as a person who hasn't read it, but it has this.

36:50 Like pop cultural intrigue.

36:54 And so I was thinking like a podcast that gives both some kind of creative prompt, but also like some kind of inspirational, like have it produce like a daily thing. That's like, you know, go on a walk and find a rock and do blah, blah, blah, blah, blah.

37:11 My version of the same idea was a little jockier. A pump-up podcast, where an AI voice just yells encouragement at you every morning. Not everything was gold, but we were having a lot of fun with it. And there was another idea I kept circling around. Something that, if it worked, would be a kind of pure AI experience.

37:31 Two AI agents out in the wild, finding each other. If you could send agents out to find other agents to talk to. That would definitely be compelling if it was like, we have an agent that's out looking for friends. Eventually, we decided to prototype a few and give them a listen. Maddie would build a self-generating podcast maker that would do everything end-to-end. Pick the voices, create the show, add intros and music, do a little quality control, whatever that meant, and get it uploaded and out into the world. our own mini AI podcast assembly line, theoretically putting us all out of a job. Except Maddie. Which got me thinking again about the nameless humans pressing the buttons to create all these shows I'd been listening to. I knew those

38:16 people were out there, if they hadn't already vibe-coded themselves into obsolescence. So I went looking for the pioneers of AI-generated audio to ask them what they thought they were making.

38:28 And why.

38:30 I'm not telling you you should or shouldn't do anything. I'm asking, is it just, survival of the fittest? Or is there any way in which we should value the human version higher in any way?

38:41 I mean, I guess my intuition is that it should be a meritocracy, that essentially something should not exist to the exclusion of something AI created just because it's human-made.

38:59 That's next time on episode two of this three-episode season, Thank you. Our executive producers are Mangesh Digadur and Julia Nutter at Kaleidoscope and Katrina Norvell and Allie Perry at iHeart Podcasts. Mixing by Steve Bone. Marketing by Christy Maribel and Ellen Liu. Our show art is by Devin Manny.

39:33 And our theme song is Me and My Shadow, arranged and performed by Stefan Dismond. Yes, we asked real, talented human artists to produce art and music that looks and sounds like it was AI-generated, without using AI. We think they nailed it. Special thanks also to Samantha Hennig and John Mualem. This show is a production of Kaleidoscope, iHeart Podcasts, and Shell Game. You can sign up to get these episodes ad-free and our newsletter at shellgame.co. Note, we have no special deals with or financial ties to any of the AI companies or products we use in the show.

40:05 You write that you are, quote, not a man of violence.

40:09 Right.

40:09 But after enduring God knows how many hours of these two hosts... you admit that if they took physical form, you would find a way to make them stop.

40:19 That is such an intense psychological friction. It's wild. I mean, you are a self-proclaimed nonviolent person being pushed to a literal breaking point simply by listening to fake conversation.

40:30 Yeah, it really highlights the severe toll of prolonged exposure to inauthenticity.

40:35 Exactly. Your brain is working overtime, desperately trying to process social cues that it simultaneously knows are entirely fabricated.

40:43 It's like being trapped in an elevator with a relentlessly cheerful artificial salesperson.

40:49 Oh, that is a perfect analogy.

40:50 Right, like no matter how many times you ask for quiet, they absolutely refuse to break character. They just keep smiling and pitching.