by Joe Biel, co-founder of WorkingLit
A version of this post, “Why we use Math instead of AI, is included in WorkingLit’s zine, Books and Math: A Manifesto on Publishing Tools, which is available on Microcosm’s website.
There’s a lot of hubbub about how AI is going to revolutionize, well, just about everything. The trouble is, it’s not. It’s a fun party trick but a poor tool. At the PubWest conference in 2024, the keynote tried to impress the audience with the promise that AI would replace internship programs. It was fascinating to watch the audience unite against the speakers who hadn’t realized that most of the publishers in the room had started as interns, whose role is traditionally to learn the craft rather than provide cheap labor. This claim is shaky in all industries but especially false in publishing. AI replaces the wrong kind of work with the wrong kind of results.
We do get excited about new tech, and were admittedly excited when AI entered the marketplace. We began testing it to see how it could improve upon core functions of our operation. The trouble is, it doesn’t work. At least not anywhere near as well as a well-trained person works. If you’re very tired at the end of your workday, AI could create work that bears facsimile qualities to your own weakest abilities. So we abandoned AI almost immediately, with periodic tests for our amusement. As one friend put it “Actually, AI is very good at creating my Hallmark movies about a luchador who does not get along with his toaster.” However, even my experiments in creating parody publications with AI were thoroughly ineffective. Humor is an emotion and thus something that AI cannot grasp. AI does not even understand why its own failed efforts are amusing.
The most fascinating thing is that if you listen to the people who have worked building AI technology the longest, they can explain how AI has the appearance of “thinking” without any of the advantages of actually being able to do so and that it has already hit its virtual glass ceiling1. You can see this if you have ever interacted with AI “customer service” robots. Or asked AI to write a description of something. It contains all of the parts of writing without any of the usefulness.
During an Independent Book Publishers Association webinar about WorkingLit, publishers cheered when we explained that we use math instead of AI to run our software. I tried to explain that this wasn’t a moral move so much as math is a hard science, whereas AI asserts misinformation with the confidence of an inaccurate Google review. As the publishing marketplace gets increasingly crowded, AI becomes less and less effective. It’s partially a signal-to-noise ratio problem but also, with more and more books and fewer and fewer distributors on the market every day, the quality of your work as a publisher needs to exceed that of AI. This doesn’t just apply to writing books, but your marketing language, cover development, and the overall thoughtfulness and utility of your approach.
When our books were distributed through an outsourced relationship, we spent our time negotiating which orders were too large or small based on what we knew of the size of each book’s market. Today, that same time is spent actually selling and shipping those books. This metaphor is probably the best way to explain the difference between AI and statistics. When someone’s workload is too overwhelming, they simply cannot do a great job at everything and have to make choices. AI is fed the largest pool of subjective data on Earth (often illegally) and told to draw sensible conclusions from it. It’s like when I visited Niagara Falls as a child. It was clear that tons of water was falling a tremendous distance but I had more to learn about water from the small puddles left on my neighborhood sidewalk after it rained.
Statistics are as accurate as the data that you feed your system. That’s why we developed WorkingLit to create the right computations with your data as a problem-solving tool. Giving yourself all of the information allows you the agency to make decisions and think critically instead of AI attempting to misdirect you. If AI was better at publishing than publishers, we’d all be out of a job already. Even Amazon, when attempting to use fully automated purchasing tools, hired experienced book buyers to review the work. Why? There’s a lot of expert knowledge you need to have about the life cycle of a certain title, genre, author, or format. As you teach AI about the many conflicting rules of each of these, it gets increasingly lost in the contradictions while your expert knowledge combined with a healthy set of statistics leads you to responsible decision making.
Over the next ten or twenty years, it is probable that AI will have a similar impact on our workflow as smartphones or the web did. Once functional systems are introduced, AI will change how we engage our workflows rather than disrupt it. But right now it’s just a lot of noise.
1As Nikhil Suresh, an Australian data scientist, has explained over and over.