Human or AI Voice Agent? On the Phone, It'll Get Increasingly Hard to Tell

A Q+A with the host of Shell Game, a new podcast that explores the eerie frontiers of AI technology

This image? Also obviously created by A.I.

I originally scheduled this post to go out a week ago, but then postponed it because, you know, the Election. Not that anyone is less distracted now. But for those of you who could use a reprieve from national politics…

I wanted to share this FASCINATING conversation I had about one of my favorite podcasts of 2024—one I think everyone should listen to. It’s called Shell Game, and is all about exploring an impending future in which A.I.-powered voice clones can act autonomously on humans’ behalf. I was shocked to learn that this technology already exists—you can essentially give an A.I. chatbot like ChatGPT your voice, and then hook it up to a phone line so that it can start independently calling and conversing with real humans on the other end of the line.

Shell Game is at times spooky, funny, and deeply unnerving—as you can get a sense for in the trailer:  

Today’s newsletter also marks the beginning of a pivot I’m making with The Longform Lowdown. So far, this newsletter has really only been a collection of posts in which I’ve talked about my own projects. I didn’t intend it that way; One of my original missions was to highlight great narrative work by other writers far more talented than me. There could perhaps be no better inaugural interview subject than the host and creator of Shell Game, Evan Ratliff.

Evan Ratliff. Photo by Jonah Green

Evan’s name is probably familiar to fans of longfom journalism. He was—after all­—one of the co-founders of the Longform podcast (RIP), which interviewed hundreds of legendary writers about their craft during its 12-year run. Evan also co-founded The Atavist magazine, which I once had the pleasure of writing for, authored The Mastermind, and used to work for WIRED—where he wrote one of my all-time favorite features about trying to “vanish” in 2009 under an assumed identity while WIRED offered $5000 to any reader who could find him.   

Like that article, which tested whether it was possible to evade digital tracking and surveillance, Shell Game explores powerful new technologies we likely can’t escape. It’s been on my mind ever since I listened to it, and I agree with its premise that we need to start thinking about the implications of what’s coming with autonomous A.I. agents—and what it means for us humans. To that end, Evan was kind enough to sit down for an interview and tell me more about the backstory of his podcast, and the conclusions he’s drawn from it. I hope you enjoy our conversation below, which I’ve edited for clarity and length.

Chris: Ok, I’d love to start at the beginning, before you made the podcast. How did you come up with the idea of experimenting with autonomous voice agents?

Evan Ratliff: So, I’d gotten to this place where I really didn’t like A.I....All the A.I. stuff that came out—even ChatGPT—I thought, this just makes me feel bad. I’m not wowed by it. And it’s making me worried about the world. I’d played around with it like everyone else and then quit. But I also used to be a person who loved technology. I’d worked for WIRED magazine and loved experimenting with every new thing that would come along. So then—and I don’t know what prompted it exactly—I started thinking, well, I don't want to become a left-behind person.

That’s when I tried to figure out if there was a way to mess with A.I. that would be interesting to me, and the voice stuff was kind of the start of it. I got into voice cloning at first. And then, after I figured out that I could add the chatbot to power it, there was one day where I was like, I wonder if I could connect this to my phone number. It was quite complicated to do at first. But when I got it working and made a call, that really was a moment of like, this shit is crazy.

So then you start having this A.I. bot, which has a cloned version of your voice, call real customer service agents to interact with them. And I must say, when I listened to the podcast, I found many of these early calls super awkward. I was laughing out loud, even if I felt some sympathy for the humans on the other end of the line. How did you react when you first heard these recordings?

Just the same as everyone who hears them—laughing out loud, sitting at my computer. I hadn't told anybody about it except for my wife. The two of us would sit around cackling as we listened to these recordings. A lot of times, I wouldn’t actually hear the calls as they were happening, but would see the recording and transcript afterwards. And if the call went on for a couple minutes, I was like, oh, this is going to be good. It didn’t really get alarming for me at first because I was just messing around and having the A.I. bot call customer service lines. It was only in the context of deciding to do the show that I starting thinking, how far could I push this?

And you push it pretty far. You have your A.I bot go to therapy. You use it to call your unsuspecting friends. You have it conduct interviews for you—both as an interviewer and interviewee. Suddenly it wasn’t so funny anymore for me. And I noticed that many people weren’t fooled by the A.I.—they sensed something was off. But some, including a good friend of yours, were tricked…Do you think we’ll get to a point soon where—even for people that know us well—it'll take them a lot longer to be able to discern if they’re talking to a real person not?

Yes, I think so. The improvements are already increasing at a fast pace. Like, Open A.I. has a new voice technology that they’re releasing called voice-to-voice. This is unlike my A.I. bot, which takes audio from speech and transcribes it to text, feeds that text into a chat bot, and then spits out more text that’s turned back to voice. This actually goes from voice to voice. And so it's much faster and sounds more human. They’re not quite yet connected up with voice clones, but we’re months away from that being true. And I think it will fool people. And people will use it in different ways, including to represent themselves out in the world. A year after Shell Game, I think you'll find that I could have done a totally different show in which a ton more people were fooled.

In one of your recent newsletters, you mentioned that you've sort of created a problem for yourself where, at least on phone calls, people who’ve heard Shell Game are constantly asking you if they’re speaking to the real Evan or the A.I. Evan…It occurs to me that you're perhaps the first person to ever go through that. What is that experience like?

I created this problem for myself. And I not only created it initially by putting out the show, but I did send my A.I. to interviews. So, if I came to this interview and had said, “I gotta go off camera,” you would have been right to ask: is this really you?

So yeah, I'm at the forefront of destroying trust. And the funny thing is, it's a terrible thing to do as a journalist, because the one thing you have is the trust of the people who you're speaking with. Now, I maybe have a little worry about sources asking, are you sending your A.I. to interview me? But I'm also still in the honeymoon phase where, if someone asks on a call if it’s me, that it means they know about the show—so that part feels good. But it does call into question: if I send my A.I. instead of me, is that deception? I think this will increasingly be an issue once more people want to send A.I’s to meetings or have them do all these other things on their behalf. I was even just starting to look into what the First Amendment rights of an agent are. Does anyone know that? Does my voice agent have the same First Amendment rights as me?

I’m interested to see what you find out. And since we’re on the topic of laws, I am curious: during your reporting, did you hear about any moves towards putting up guardrails or regulations around this—or is it your sense that we're totally just scrambling to keep up with the technology?

The latter.  The short answer is the latter. I mean, you see some efforts to regulate A.I. in a broad way, like that California bill that was recently vetoed. But that was to address large scale dangers and to create a “shut-off” button if A.I. got too powerful and did things we didn’t want it to do. That’s as close as anyone’s gotten to putting these regulations in. Part of why I wanted to make the show was because there’s no regulation around the sort of day-to-day ways in which A.I. infiltrates our lives, particularly with these autonomous agents. And I think the people who stand to benefit financially from this in the tech world, they're always like, go, go, go! We'll figure it out. Put this out into the world. See how the consumer reacts. But then the people who would naturally kind of be like, I don't know…should we put the brakes on this? Those people are not paying attention right now. They're just like, I don't want to hear about this. This sounds terrible. I hope this world doesn't come to pass. And that means the people who are accelerating this tech into the world, they always kind of win.

Even since you released Shell Game, there are already new generative A.I. technologies making use of voice imprints including—and this is getting real meta since we’re talking about podcasts—a Google product called NotebookLM which can produce a full-fledged podcast based on an uploaded prompt. You wrote about this is a September 30 post, and mentioned it might mean that we see more “A.I. slop.” My question for you is: when it comes to the kinds of high-quality, human-made podcasts like Shell Game, do you think it’ll get harder and harder to reach audiences and break through all the A.I. slop? Or will there be so much low-quality, A.I-made crap out there that the quality work will somehow rise above it?

I tend to think the problem that we’re not always willing to admit is we created real slop. Like, the problem is that we made it so easy to be replaceable in this environment because there's so much terrible content that already exists. I mean, the number of humans just talking to each other and regurgitating the same stuff over and over again on different podcasts is now borderline infinite.

If you're trying to produce very-well reported, resource-intensive journalism, at least in the near-term future, it doesn't necessarily change the fortunes of doing that. Other than that, in the same way that we've already put too much low quality content out there for people to consume, that will just continue to be a problem. But that's a problem we created over the last 10 years. It's not like A.I. suddenly arrived to create that problem.

Shell Game does such a good job exploring the potential pitfalls, benefits, and ethical questions around autonomous voice agents. But ultimately, for me, the idea that we might soon struggle to tell what’s real and what isn’t is really scary to me. Are there any silver linings you see in this brave new world?

I think the only optimism I could draw from it—and I tried to write this into a New York Times op-ed—is that I think it does drive you more towards wanting to hang out with people in person. It really pushed me towards that—not just because then they know it's really me and not a voice agent—but also misunderstandings don't happen in the same way if you are spending time with people. Misunderstandings already happen on social media and on text messages, because the emotion isn't conveyed the right way. So yeah, the whole project did push me back towards wanting to get offline more.

To find Evan online, here’s his website. Also please consider supporting Shell Game as a paid subscriber. One thing we talked about was his decision to release Shell Game independently, rather than take investment from large podcast companies that demand creative control. This meant he bankrolled the whole project himself—and he’s also thinking about a second season! You can find Shell Game’s substack here.