Hand-wringing about AI, Part III: “We’re Stuck in the Middle”

Back for Part III? Curious if I can land this plane? (ME TOO.) If you missed the earlier pieces, here’s Part I, and here’s Part II.

To recap: we’ve seen how AI can thrust us into a world of infinite choice by theoretically translating (or eventually writing) any book out there, which is interesting from the point of view of connecting us with literature and cultures from around the world, but also exhausting, since we have more than enough to read/watch/listen to already; and then we saw how it can—with some prompting—help publishers market and explain their books. And in both instances, the material generated by AI (the Proust samples, the Attila copy) is not dramatically better or worse than what the majority of humans can craft.

That said, I stated at the top (some 8,000 words ago) that my goal with these posts was to affirm the human and our ability to sense the uncanny valley of AI creations—and avoid them. Now, as we enter the final stretch, I’m going to try and convince you that not only can we sniff out AI, but that over the next five years, as a culture, we’re going to turn away from TikTok and short form narratives, reject the AI-pocalypse that is flooding our timelines, and instead seek out the full human experience, as presented by actual, living, breathing humans. Not sure I can convince myself of this hot take, but regardless, you’ll finally find out which Proust translation is which . . .

*

And we’re going to start with one of the best podcasts I’ve listened to all year: Shell Game. A six-part series created and produced by Evan Ratliff, which is described as such:

Shell Game is a podcast about things that are not what they seem, hosted by journalist Evan Ratliff. In Season One, that thing is Evan’s voice. By creating a voice clone and hooking it up to an AI chatbot, Evan set out to discover what happens when you try to take control of the very technology that threatens to replace you. Shell Game was named one of the the best podcasts of 2024 by New York Magazine and called “awesome” by The Verge. Over the course of six episodes, Evan’s voice agents talk to spammers and scammers, to Evan’s friends and family, to colleagues and sources, to other AIs, and even to a therapist—all to better understand what AI voice is able to do, what it can’t yet do, and what to expect from a future in which more and more of the people we encounter in the world aren’t real.

I’m not going to get into the details of how he does this (you can listen to the first episode and find out), but as mentioned in that (AI-created?) copy, he created a “voice agent”—an AI that uses recordings of your voice and backend scripts to do things for you such as call the doctor and set up an appointment, dispute a credit card charge, or, ideally, attend a Zoom meeting while you get wasted poolside.

This is, to put it bluntly, horrifying shit. We’ve had robocalls for ever and ever, but this is truly next level. For instance, in one episode he has his voice agent call a bunch of his friends. Some of them cotton on immediately, but one . . . doesn’t. And he’s totally unnerved by it for months afterward. (“Am I really talking to real Evan?” “Of course this is the real Evan.” “That’s what your AI said as well.”) It’s eerie to hear the recordings of AI Evan talking to a therapist at BetterHelp who takes Evan’s voice agent completely at face value. (Can you spot the difference by someone with mental health issues and an AI?) He even goes so far as to have his voice agent provide all the promotional interviews about the podcast to the tech media.

When I first discovered this podcast, I found it hilarious. Even though the concept is unnerving to say the least, the voice agent messes up a ton and you feel secure in human superiority. But by the end of the series, I was reading this a performance piece, à la Nathan Fielder. (Especially The Rehearsal.) Real-life Evan’s affect is pretty flat, but his AI’s is . . . odd. The pitch is a bit off. It’s a bit too excitable. It’s 99% of the way there, but you can sense the robot in the background. But putting that oddness into real life situations—the very situations that all big companies are striving for, where the only people on Zoom are AI voice agents who then summarize the content of the meeting for their real life counterparts—is funny, sure, but also illuminating as it points to something about being human that’s hard to define, but easy to recognize.

Like I said, most of his friends—along with the tech reporter interviewing AI Evan and also the founder of the voice agent service who AI Evan talks with (“I just wanted to create a system so I could have an AI call my mom.”)—recognize the ruse immediately. Then it’s up to them if they’re going to just go along with it or blow up AI Evan’s spot. And when they do, especially when they poke at AI Evan just a little bit, uncomfortably comedic things happen. Especially because the AI tends to lie and invent stuff. (See Part II, but exaggerate those little lies.) And because it gets confused—a lot. And repetitive. There’s so much, “Hey, it’s Evan. Great to hear from you. We should catch up soon” in this show.

But the three main points from Shell Game that relate to this winding series of posts are that 1) every technological advance removes something from the current state of being human (such as eliminating the richness of phone calls for two-dimensional texts; or as the phone did previously by reducing the need for face-to-face contact), 2) these things are proliferating like mad and anyone can make one and train it up in next to no time, and 3) the technology is stuck in the middle, where it’s most useful in theory, incapable of running a meeting. (My god, I only wish. I would give anything to send ChadGPT to all meetings ever.) And Evan’s fear is that we’ll remain stuck there, in the midddle, indefinitely. Anxious that AI will take over our lives and jobs, continually investing time and money into its “education,” but also encountering, over and again, these weird limitations that nevertheless mark it as inhuman. Like, for instance, the two AI translations of Proust that you most certainly have identified by now! Or, um . . .

*

This taps into a fear I’ve held for years, what I call in my head the “Interminable Middle,” or in a more dramatic, catchier fashion, the “Heat Death of Culture”: the sheer proliferation of content—be it books, short TikToks, images, music, tweets, AI voice agents that call me or talk to me in meetings—is so overwhelming that we only choose to interact with culture that we’re already familiar with, and as a result of that familiarity, everything seems . . . pretty good. Not horrible, not genius. But satisfies that “certain itch.” An itch of looking for something new that’s just like what you already like. Which, from a business point of view, we’ve always been moving in that direction, since it would eliminate risk and increase the probability of knowing that a particular cultural artifact—one that’s a close imitation of the trends and aesthetics behind a recent mega-success—would be a hit.

It’s expensive and financially risky to publish something that’s too unusual. Not books that are “weird,” but books that are sui generis, use a style, structure, language, or method that will leave most readers questioning whether this is genius, or overrated, or just gobbledygook. The whole of literary history rides this line. Books that, in the moment, seem like insane gambles given the state of reading and book buying and what’s successful; but which, from the vantage point of the present, are brilliant bets on smart innovation.

From Finnegans Wake to The Familiar. From Miss MacIntosh, My Darling to Ducks, Newburyport. There are dozens of these sorts of books out there that we could all name. What we can’t identify are the ones that fade away almost immediately. The counterexamples that, under different circumstances of time and brand and chance, could’ve been the truly “innovative” book that is talked about five years after publication.

Here are two extreme business models to choose from: 1) use AI to produce 10,000 literary works a year (including a ton of translations), by which you essentially create your own trends simply by creating a massive volume of work, teaching your AI to analyze then follow these trends in order to capitalize on them, or, 2) pay someone to analyze whatever amount of words they’re literally capable of reading, digesting that knowledge, then picking a book that could become legendary, but is more likely to flop.

If you want money, pick number 1; choose 2 if you want to have fun. And likely lose your livelihood.

*

Although the best presses have a great editorial vision and enough variations and sheer volume of titles to sustain them, the second option can strike a lot of people as a form of gatekeeping. We’ll come back to this, but for the moment, I just want to acknowledge that the more people who make decisions about what does and doesn’t get get published, about what’s presented as “excellent” or “innovative” or “unique,” the better.

Regardless, from a reader’s point of view, this all sounds exhausting. Is there a need for 10,000 new works of literature every year, just so a company can not just cover costs to stay afloat, but earn enough to produce 12,000 new books the next year? Not really. And if you want to know the state of our world, including all self-published works, ebook only publications, etc., it’s highly likely around 3 million new books are coming out every year—a majority of which are fiction. By contrast, twelve thou seems manageable.

When you reach this level of production, two things seem to happen: 1) according to the power law and Pareto principle, very very few works are astronomically popular, and almost everything else attracts next to no attention, and 2) it’s more rewarding to consume what you know is pretty good, than to spend the time looking for the uniquely genius. 

As such, the power law gets reinforced, a fewer range of works reach an ever larger amount of consumers and, because we’re chasing the “infinite middle” of culture of mass appeal amid infinite choices, everything becomes a 3.5. On GoodReads, almost all books published today are a standard deviation away from a 3.5 rating. Everything is pretty good to most people.

And if AI ran the business, churning product to be consumed by smaller and more specific audiences—but with a handful of creations every year being loved by seemingly everyone—most everything we encountered would probably be pretty good. Passable. Human, AI, who cares? If it’s in front of me, is good enough, feels comfortable, and meets enough of my priors . . . I’d read at least 20-30 pages. And I’ll forget 95% of what I read or watch within five years. Or will conflate it with other things. It’ll all seem vaguely the same to me. Whether this is due to the difficulty of finding the unique amid an overwhelming amount of “decent” art or the onset of dementia—only time will tell!

But maybe it’s just because I read too much . . . Although if you don’t? . . . How will you ever know what’s good vs. satisfying, singular vs. competent? How do you make judgment when you can’t even experience a fraction of the possibilities?

That’s the Heat Death of the Universe I hand-wring over.

 *

And, looking toward the future, this situation is only going to get worse. You think there’s a lot of information available to you now? Every technological innovation multiples the amount of available information. You have more pictures on your phone than were taken, globally, in, say, 1974. Information proliferates. Faster and faster all the time.

*

Let’s talk about Proust and his slow-lived, limited-information life. His going to bed habits as described to us across the years.

Six different translations of the opening pages of Proust are presented in Part I of this series. Four done by humans, two by AI. (Part II reveals that the 4th sample was produced by Google Translate.)

For a long time, I went to bed early. Sometimes, as soon as my candle had gone out, my eyes closed so quickly that I didn’t have time to say to myself: “I’m falling asleep.” And, half an hour later, the thought that it was time to go to sleep woke me up; I wanted to put down the volume that I believed I still had in my hands and blow out my light; I had not stopped thinking while sleeping about what I had just read, but these reflections had taken a somewhat particular turn; it seemed to me that I myself was what the work was talking about: a church, a quartet, the rivalry of Francis I and Charles V.

To me, this it too long, too clunky. Also, although this gives away the game—note the use of “volume.” How often have you said, “I need to set down this volume.” It’s a book. It’s always a book you read in bed.

Of the six samples, number 2 is the only other one that uses the word “volume”:

For a long time, I went to bed early. Sometimes, barely had I blown out my candle, my eyes would close so quickly that I did not have time to say to myself, “I’m falling asleep.” And, half an hour later, the thought that it was time to seek sleep would awaken me; I wanted to put down the volume which I believed was still in my hands and blow out my light; I had not ceased to make reflections in my sleep on what I had just read, but these reflections had taken on a somewhat particular turn; it seemed to me that I myself was what the work spoke of: a church, a quartet, the rivalry of François I and Charles V.

And that, dear reader, was created by ChatGPT.

What’s really remarkable about this experiment was that my class, almost unanimously, chose number 2 as the best of the samples. And, to be fair, I can see why. It’s arguably the smoothest and, in some almost indefinable way, the least anxious. (But definitely the most 3.5.)

Your traction with the others may vary, but let’s look at James Grieve’s (sample 5):

Time was when I always went to bed early. Sometimes, as soon as I snuffed my candle, my eyes would close before I even had time to think, “I’m falling asleep.” And half an hour later, wakened by the idea that it must be time to go to sleep, I would feel the desire to put away my book, which I thought I was still holding, and blow out the light. While I had been sleeping, my mind had gone on thinking over what I had just been reading, although these thoughts had taken an odd turn—I had the impression that I myself had turned into the subject of the book, whether it was a church, a string quartet or the rivalry between François I and Charles V.

That opening line is so divisive! “Time was when I always went to bed early.” It barrels ahead, unlike the pause in “For a long time, I went to bed early.” Straight propulsion! And then “I snuffed my candle”??? That’s so bold! (See: “my candle scarcely out.” See: “when I had put out my candle.” “I snuffed.“) And I can bet, without knowing him at all, that there was significant thought put into this more than average deviation from the typical translation. And, he probably second-guessed himself at some point, uncertain if his choice works or if people will like it.

This is a human translation. Humans are messy. They fuck up and worry and doubt and fart and live. (And swear! That’s how you know I wrote this. Swearing is human.) It’s in the anxiety that we sense the human.

You feel anxiety in old Marcel, that’s for sure. (I love you, Marcel! I just need some space. To better appreciate you.)

Is that what we sense when we encounter AI? The smooth, over-confident, unquestioning nature of what it produces? But how many people stop to notice that?

*

Just so you have the full list, the translators are: Lydia Davis (1), ChatGPT (2), Original Moncrieff (3), Google Translate (4), James Grieve (5), and the Moncrieff revised by Kilmartin (6).

*

*

*

As AI improves and becomes increasingly capable of mimicking human creativity— translating our words, capturing our voices, and even calling my ex-wife so that I don’t have to—we risk losing the human touch that imbues our creations with authenticity.

And that’s precisely why the human experience will always matter. In a world flooded with AI-generated content, it’s the unique spark of the human that will stand out even more. Our imperfections, our quirks, our ability to perceive beauty in unexpected places—these are the qualities that define us and make our creations resonate. AI can churn out endless content—so much endless content—but it can’t replicate the singular.

Being human means more than just processing information; it’s about being messy. As we move forward, I believe we’ll see a resurgence of interest in the messy, in the uniquely human aspects of art and culture. People will seek out those authentic voices, the ones that speak to our shared humanity in ways no algorithm can. The more automated and digitized the world becomes, the more we’ll crave the real, the raw, the deeply personal.

We have an incredible opportunity to use AI not just to replace us but to amplify what makes us special. It can help us reach new audiences, break down barriers, and explore creative possibilities we never imagined. But it will always be our bullshit that breathes life into those possibilities.

So let’s embrace this new era with optimism. Let’s use these tools to tell our stories, to connect across cultures and languages, to share what it means to be alive in ways that only humans can. The future isn’t about AI taking over; it’s about making work-life shorter and our free-time longer—so that we can play.

*