Hand-wringing about AI, Part II: “Write Me an Ad Campaign”

You might want to read “Part I” before going any further, but if you just want a recap, that post is essentially about how AI could translate the world (and/or create millions of new novels), which, on one hand, could be useful in bringing unique, diverse voices to an English audience, but, on one of the many other hands, leaves readers and cultural consumers with an impossible task of trying to wade through basically an infinite amount of material to find the “right book” to read at the “right time”—a task the human brain is notoriously bad at, forcing us to rely on various heuristics and other signals to guide the decision making process.

Today, I want to look at the possibility of AI assisting publishers—and translators, and authors, and readers—in ways other than simply creating more products to sell. AI as a productive tool for marketing and, with a little luck, book discovery. In other words: Can AI help solve some of our decision fatigue both by finding you the book you should be reading and/or by helping marketing directors and publicists present books in the best possible way?

This is frequently on my mind—especially when it comes to writing book descriptions (aka, jacket copy), or to be honest, creating any of the materials our distributor (Consortium/Ingram) need in order to sell our books. They always have us enter in this information over what would be winter break for me (late-December through early-January), or right at the end of the school year (May). The info they need then, for pre-sales meetings,  is referred to as a book’s metadata, and, one the base level, is super easy and quick to produce. Things like ISBN, Title, Author, Translator, Price, Rights Territories, etc. Objective, simple, factual bits of information.

But after that, things get more squishy . . . For instance, there are BISAC codes, which are used to categorize books, such as “FICTION / Literary” or “FICTION / Romance / Paranormal / Shifters” or “FICTION / World Literature / Scotland / 20th Century.” There are thousands of these spread across fifty-three categories. And some books (think Dubravka Ugresic’s nonfiction, think Dmitry Bykov’s hybrid work VZ: Portrait Against the Background of a Nation—which is about Zelenskiy, but is neither biography not cultural criticism) can fall into a number of these categories. Although figuring out exactly which categories they fall into can be confusing. (There are thousands. And although Consortium is helpful in this, no one involved in the labeling process knows both the book itself and the full list of options.) Regardless, this is an act of marketing that we have to perform as publishers. We choose what other books this book gets grouped with, and, I can only assume, some labels, some categories, would help sell more copies of a book than others through the specificity and that label’s audience.

From there, things get more subjective and require marketing insight and consideration to be optimally effective:  bios, translator bios, keywords, and my least favorite: jacket copy. Writing jacket copy is one of my least favorite things to do, since what appeals to me about a book might not be what draws in anyone else. (I remember John O’Brien criticizing my Chapel Road copy because I originally used the term “postmodern.” Worth noting that he had no issue with this line from the copy for Hortense Is Abducted—first jacket copy of mine to go out into the world—which reads, in part, “22-year-old philosophy student whose buttocks are so beautiful their description has been banned from the printed page.” Clearly, people prefer a nice butt to postmodern shenanigans.)

One of the issues with jacket copy—especially for translations—is that, frequently, this information has to be entered into Ingram’s system (CoreSource) before I’ve had time to edit the books, or even read them in full. This is partly my own fault for announcing books too far in advance (sometimes before the translations have even arrived), but honestly, I work just about a year in advance for the bulk of my editorial projects. And, to make this concrete, I’ll have to submit descriptions for books coming out in March 2026 before the end of this year.

This copy isn’t fixed in stone, but it does go out to Amazon, Bookshop.org, Edelweiss+ (site for booksellers and reviewers), our website, etc., etc., so it’s pretty damn important. It’s also the basis the sales reps use to describe our books to buyers at bookstores, and is likely what book review editors will also see.

As with the Proust AI translation exercise detailed in Part I—I haven’t forgotten, you’ll get the answer key soon!—it’s debatable how much value is added by having slightly better copy. But there is great value in freeing up your time and mental energy for other projects. And my god how I stress out over writing copy! If only I knew what words and phrases people gravitated toward (butts?), or what’s worked in the past, or if there’s a secret formula . . .

Enter AI! Every day, I’m barraged by dozens of ads for how AI can make my work life easier and more efficient. “Remember when we wrote emails without AI?” is a scoff I hear on the regular. (To those few who receive them, could you imagine AI writing my emails? I joked about this a couple years back in this post, but seriously, if you think these posts are long and rambling and digressive and sort of dirty, you should be one of my pen pals!) But given just how much corporations are leaning into the whole “integrate AI into your workflow” thing—and I admit, there are many benefits, and many emails I’ve received should have been written by AI so the sender would’ve avoided sounding like a simpleton, or simply to fix the tone—maybe it is possible that AI could solve jacket copy . . . Not to mention the other two metadata categories that cause me fits: Key Selling Points (reasons why anyone would want this book, independent of its content) and Publicity Plans (which all sort of fall into a pattern of sameness, given how little time and few resources we have to get truly creative and go big). Save me, ChatGPT! Make my metadata better!

*

Before going to ChatGPT to explore the possibilities, I fooled around with some knock-off AI options, such as AI Book Nerd, the name of which delights me. I found it in a Reddit thread where someone wanted an AI service so they could ask questions about specific books. (To make essay writing for class easier, I assume.) This is the headline at the top of its very sparse webpage:

Search for a book that you want to ask questions to – Powered by ChatGPT

You only need the book title, and Artificial Intelligence takes care of the rest

Promising! Or at least potentially fun.

Now, I had my doubts as to whether this service would actually have any Open Letter books listed, but lo and behold, Not Even the Dead pops right up!

As you can see, what I asked AI Book Nerd (which really should have its own Instagram account to share the weirdest query results) was to summarize NETD. And here’s the nonsense it spat out:

WTF. Like, literally, What. In. The. Fuck.

What blows my mind about this is that our jacket copy is readily available! It’s eminently findable! That’s truly what I was expecting to get as a response. “The conquest of Mexico is over, and Juan de Toñanes is one of so many soldiers without glory who roam like beggars for the land they helped subdue. When he receives one last mission, to hunt down a renegade Indian who’s called the Father . . .” I mean, it’s right there. What is this “mysterious literary society in Madrid”? And “Grigor Fedorov”??  GTFO.

Now, it’s super easy to goof on AI mistakes—and easy to feed it misinformation so that it believes “belief” has 3 S’s in it—and sure, it’ll get better, but at first glance, what it seems best at is lying and gaslighting users. For instance, in prepping our recent newsletter about the release of Dubravka Ugresic’s new book, I wanted to revisit the Starred Review it received from Kirkus. So I googled “Kirkus Reviews Muzzle for Witches Dubravka Ugresic” and received this tiny, and totally inaccurate, summary from Gemini AI, Google’s in house AI service.

That “overview” is immediately above the top search result which, duh and or obviously, happens to be the Starred Review in Kirkus. So, in short, Google has an AI summary bot that can’t incorporate data from its own search engine????

Again, the value in pointing out these flaws is mostly humorous. (Or maybe a work of performative art. Tune in to Part III for more on that.) Although, before shifting gears, I do have one more funny to make. I asked ChatGPT over the summer to write “an ode to J.D. Vance in the style of Rupi Kaur” and I got this gem:

your cushions hold
the weight of my dreams
the spills
of my late-night thoughts

Amazing! And wow! Can’t figure out that Kirkus reviewed A Muzzle for Witches, but is right on point with J.D. and his couch proclivities. This is the future.

*

But what if we swerve from the idea of AI being able to digest a text, distill it, and know how to market it to actual humans for a second, and instead look to AI for helping us wade through the endless amount of product out there? We’ve fretted about “the algorithm” for years, noting how Instagram ads “just happen” to match up with the topics we were just talking about. We know that Amazon, for decades now, has used their special algorithm to recommend products and has been working on “anticipatory shipping” to send you products you didn’t know you needed until they arrive.

AI adds some wrinkles to this concept. Although it’s built via algorithms, AI veers from calculation into the realm of “intelligence,” making decisions based on factors that aren’t 100% mathematic. Totally linked ideas, yet separate.

There have been a lot of algorithmic-based book recommendation engines over the years outside of Amazon. BookLamp (defunct after being sold to Apple to be the algorithm powering the wildly popular [sic] Apple Books app), What Should I Read Next?, Small Demons, WhichBook. All of these apply different strategies for how they evaluate and compare books—for instance, Small Demons is all about proper nouns, BookLamp used “Book DNA” à la Pandora to parse books across hundreds of “genetic” elements—but all of these rely on their inputs and the human thinking behind how to categorize and group books.

For example: When BookLamp was available, I once asked it to recommend books I would like if I liked Sound and the Fury and, based on the “BookDNA” which included high marks in “Vehicular Fiction” and “Southern Family Saga” recommend to me a YA title about a teenager growing up in Mississippi. Swing and a miss!

Or, in the case of WhichBook—where you choose four categories and how extreme you want the recommendations to be in each—I asked for something “funny,” “disturbing,” “unusual,” and with “extreme sexual content” and was lead directly to Monster Porn. Regardless, although the connection between the service’s structure and the user’s input are connected by advanced algorithms, these tend to have  wildly inaccurate recommendations.

But if making fun of bad AI results is silly and not advancing any ideas, dunking on the algorithm is just as pointless. In terms of books—or any cultural product—the algorithm ignores style. The connections made between artworks have to be more easily mappable. Figuring out how to quantify Faulkner’s style is too advanced for the machine circa 2013.

By contrast, an AI recommendation bot could, in theory, make decisions for you that aren’t so beholden to this sort of quantified analysis. In the abstract, these recommendations would be more akin to our own decision-making process—current mood, environmental trends, overall vibes, inexplicable longings—and would be more focused on what you should read now for the greatest benefit, whatever that may be. (Acquiring knowledge, entertainment, linguistic puzzling, fodder for the cocktail party at the country club.) I could input the 30 books on my current “to read” pile and it would decide for me what to read, and in what order. Although maybe it would lie? Tell me “there is no new book by Kate Atkinson”?

*

So, the question is: can it do that? I mean, I asked ChatGPT to convince me why the St. Louis Cardinals would win the World Series this year and it referenced a half-dozen players who haven’t played for the team in years. Devoted as I am to actual experiential results and not just half-assed theorizing, I literally just had this conversation with ChatGPT:

CWP: tell me what book i should read next

ChatGPT: I can recommend a book based on what you’re interested in! Do you prefer fiction or non-fiction? Any particular genres or themes you enjoy, like sci-fi, fantasy, history, or personal development? Let me know what you’re in the mood for.

CWP: literary fiction in translation that is linguistically and structurally complicated

[ChatGPT recommends The Wind-Up Bird Chronicles by Haruki Murakami, 2666 and Savage Detectives by Roberto Bolaño, Petersburg by Andrei Bely, and Hopscotch by Julio Cortázar. Three of these are three of my favorite books of all time.]

CWP: I’ve read all of those. Give me a similar recommendation from a book released in 2024

ChatGPT: For a literary fiction recommendation in translation from 2024, you might enjoy The Book of All Loves by Agustín Fernández Mallo, translated from Spanish by Thomas Bunstead. Set in a dystopian future, this experimental novel blends fiction, philosophy, and reflections on artificial intelligence, while examining the nature of human relationships. Its complex narrative and style, reminiscent of Borges and Bolaño, explore the meaning of love and humanity in an era dominated by technology. It’s expected to be a rich, structurally intricate read, ideal for those who appreciate both intellectual depth and narrative experimentation​. Another upcoming release is Thirst by Marina Yuszczuk, translated from Spanish by Heather Cleary. This feminist, gothic novel interweaves two time periods and explores themes of loneliness, desire, and motherhood through the stories of a vampire in 19th-century South America and a modern-day woman in Buenos Aires​.

Both of these works promise the kind of linguistic and structural complexity you’re looking for!

[OK. Wow. I may think Mallo is overrated, but I’ve read everything available in English. And a book about AI talking about the human vs. inhuman? Sounds . . . familiar. And anything Heather Cleary translates is gold.]

CWP: based on what I’ve told you, out of these three books, which should I read next: Creation Lake by Rachel Kushner, Death at the Sign of the Rook by Kate Atkinson, or The Obscene Bird of Night by Jose Donoso?

ChatGPT: Based on your preference for linguistically and structurally complicated fiction, I would recommend reading “The Obscene Bird of Night“ by José Donoso next. This novel is an intricate masterpiece of Latin American literature, known for its surreal, fragmented narrative and dense prose. It delves deep into themes of madness, identity, and reality, with a nonlinear structure that constantly shifts between perspectives and timelines. The complexity of the novel aligns well with your interest in challenging, layered texts.

While Rachel Kushner’s Creation Lake and Kate Atkinson’s Death at the Sign of the Rook are both engaging in their own ways, neither matches the level of linguistic experimentation and structural innovation found in Donoso’s work. The Obscene Bird of Night offers a unique, immersive challenge that seems perfectly suited to your tastes!