Review by Jacob Pleasants
Reviewing a podcast is a bit of a deviation from norm, but this one is distinct enough to warrant it. I just recently found my way to Shell Game, courtesy of a recommendation on Bonnie Stachowiak’s lovely podcast. It’s a serialized documentary podcast, the first season of which had 6 episodes that launched back in July 2024. The second season (also 6 episodes) has just wrapped up in December 2025. The first season chronicles the adventures of journalist Evan Ratliff as he experiments with creating an AI “agent.” Here’s the overview from the trailer:
“What would happen if you created a digital copy of yourself, powered by AI, and set it loose in the world? Evan Ratliff, longtime tech journalist, decided to find out. He combined a clone of his voice, an AI chatbot, and a phone line—many phone lines, actually—into what are called “voice agents.” Then for the six months, he sent them out… as himself.”
Season 1
In the first episode, Ratliff gives a rough overview of how he set up his voice agent. At first, it required a whole lot of stitching together of various AI products; now there are companies who will do pretty much all of that ducktaping-together work for you. He then tests out his voice agent by contacting some call center hotlines [1]. The conversations are recorded, but Ratliff isn’t eavesdropping on the calls in real time. Like us, he only hears them after the fact. In later episodes, we hear calls between the voice agent and phone scammers (some of which are also voice agents), two “Evan Ratliff” voice agents, the voice agent and (AI) therapists, and more.
During each episode, we thus get to hear quite a few conversations between Ratliff’s voice agent and a sometimes-human interlocutor. Now, if you are at all like me, that’s not exactly enticing. I’ve read and heard plenty of AI-generated text at this point, so the novelty isn’t much of a draw. Instead, I often find it to be rather tedious if not downright stultifying. But that was not my experience with the conversations in this podcast. There was something fascinating about hearing the idiosyncrasies, missteps, and bullshittery of the voice agent. Some conversations are perplexing, some are concerning, and others are absolutely hilarious. I was in stitches the first time I heard the voice agent make up a credit card number: 123456789, obviously! When it got called out for its nonsense, it dutifully apologized that it had gotten it wrong and corrected itself: it’s actually 987654321. It does the same thing when inventing account numbers, phone numbers, pretty much any number – except ZIP codes, in which case it reliably gives out 90210.
The phone conversations are the centerpieces of each episode, but the podcast does not merely present them to us as spectacle. Surrounding the encounters with the AI voice agent are Ratliff’s reflections and analyses, his attempts to work out what it all means. I found that tagging along with Ratliff’s interpretations and decisions to be provocative and worthwhile, even if he doesn’t leave you with too many satisfying conclusions at the end of it all. At the very least, he managed to weave together his explorations in an engaging and often entertaining way.
Takeaways, Insights
Aside from being entertaining, what exactly is the point of all of this? It’s easy for these kinds of journalistic exercises to devolve into a sort of “gee whiz” story, or what amounts to a product review in a different guise (I tried ____ for a week and this is what I found…). There is some of that going on here. Ratliff is documenting the performance of a particular technology at a particular point in time, the capabilities of which are likely to change. The first season showcases voice agents in early-mid 2024, but why should we care about the limitations of that “old” technology? So what if they suffered from latency problems (they take just a bit too long when responding in a conversation) or sometimes repeated back to the other speaker something they had just said? These are simply technical shortcomings, a reflection of the state of the art at that time.
As fun as it is to listen to the buffoonery of the AI agents, there is a deeper story that emerges over the course of Shell Game. Ratliff isn’t just stress testing the technology, but trying to glimpse the future that is being promised and pursued by countless tech companies. He mentions a then-recent interview between the Verge’s Nilay Patel and Zoom CEO Eric Yuan in which Yuan proposed that AI avatars could attend meetings in your stead. There are the many attempts to create AI therapists (not really “the future” at all). Or at least AI “companions” to address the “loneliness epidemic.” Or the more mundane promise that an AI agent will be able to book your next flight. Shell Game takes these proposals (somewhat) seriously to ask: What is this imagined future actually like? What is like to experience all of these AI agent interactions?
At the end of the first season, one of the things that most stuck with me was the uncertainty experienced by many of those who interacted with the voice agent. The voice agent was sufficiently in the “uncanny valley” to cue the other party that something was up. Many suspected pretty quickly that they were not talking to a human. But how do we really know? If you suspect that you’re talking to a machine, what do you do? Do you hang up? Roll with it? Try to take it off the rails? These are not hypothetical questions. But perhaps more profound is the larger experiential implication of all of this: What is it like to live with the constant uncertainty that the voice on the other end of the line might not be human? Shell Game does not present a rosy picture.
Season 2
Season 2 continues many of these thread by taking up the proposition that companies of the future will be populated by AI agent “employees.” Once again, this isn’t merely a hypothetical exercise; many startups are actively trying to do this very thing. Ratliff launches a startup with a team of AI agents and tries to figure out how to get them to do something useful. Season 2 winds up being much more about the minutiae that Ratliff encounters as he struggles to wrangle what proves to be a rather unruly bunch. His misadventures are plenty amusing, but I was left with somewhat fewer compelling insights, aside from confirming that the notion of an “AI coworker” is complete nonsense [2]. One interesting recurring theme is Ratliff’s tendency to anthropomorphize the bots, and the implications of that as the ostensible boss of the operation. Why is there such an impulse to give the bots names, distinct voices, personas? It may make it easier to interact with them, but that decision influences how we relate to them in some unanticipated ways.
In sum: I stuck with the second season and enjoyed it well enough, but it’s much less of a “must listen” than the first.
Notes
[1] Wait, he was using his voice agent to troll call center workers? Yes, and I do not approve of this choice. On reflection, Ratliff also decides that this was an unkind course of action and turns his voice agent on phone scammers instead – a much less morally problematic decision. That said, he never quite fully owns up to his treatment of those call center workers, who definitely did not consent to being part of his experiment. At the same time, though, there is journalistic merit to what he is investigating. A prime use case suggested by the companies that are developing these “AI agents” is to use them for precisely this kind of thing. Ratliff wasn’t deploying his voice agent on hapless call centers “for the lulz,” but to genuinely investigate what the agent would do. I have ethical concerns, but I can also see the journalistic case. Debating the merits of this journalistic action might actually serve as a really great case study to use in a classroom context.
[2] Ratliff does manage to get the bots to do some potentially worthwhile things, mostly in the domains where you’d expect it. Need to code up a basic website? That is something that AI is reasonably well equipped to do. Generate gobs of ideas for a company logo? Most of those ideas will be poor, but AI sure can produce a lot of them.