On AI as Cake
A non-anthropomorphic metaphor for understanding generative AI and writing.
Like millions of viewers around the world, my wife and I are fans of The Great British Baking Show. We came to it later than many: during the pandemic, we marveled at the ingenuity of their designs, we salivated (and occasionally gagged) over their unique flavor combinations, and we found comfort in the contestants’ friendships. We came to appreciate the delicate combination of art and science needed to bake puff pastry, bread, choux buns, pie, tortes, tarts, and, above all, cake.
I thought of the Bake-Off immediately when I encountered the metaphor AI as cake in a forthcoming article by Anuj Gupta, Atef Yasser, Anna Mills, and Maha Bali. The authors argue that the metaphors we use to describe AI can serve as productive means for fostering critical AI literacy. They attribute AI as cake to a tweet by Bali:
If we used cake as a metaphor for AI, I'm asking educators is: 1. When/where would it be acceptable for students to take shortcuts such as buying from a bakery, baking from a box, or buying a Twinkie instead of baking from scratch? 2. How would you encourage home made? #highered
I’m more of a cook, so baking from a box is generally my forte, whereas my wife will periodically challenge herself to a recipe from the Bake-Off. (The princesstarta is still one of her proudest accomplishments.)
For me, box baking gets the job done. Each year, my daughters and I can whip up a strawberry cake with cream cheese icing for my wife’s birthday. It’s always moist and delicious. But to Bali’s point, box baking has taught me very little about baking in itself. I suppose I could muddle through some combination of egg, flower, sugar, a fat, and baking powder to produce something close to edible, but I don’t yet have enough experience to know how to manipulate the proportions of those ingredients to produce something original. Betty Crocker it is. My wife, on the other hand, has baked a range of treats, from cookies to cherry pie to baklava. Some recipes she can produce by heart. New ones she has a more intuitive sense of how to follow.
We can apply the metaphor of AI as cake in this relatively binary way: writing with AI is like baking from the box, and writing without it is like baking from scratch. But I don’t think this was Bali’s point. For example, when my wife decided to bake baklava, she bought phyllo dough from the store, rather than making the notoriously difficult pastry completely from scratch. Similarly, we might use AI for some aspects of the writing process so we can spend more mental time and energy on others. For my part, I don’t let generative AI write prose for me because I prefer to bake my words from scratch, but I have sometimes found it to be a useful sounding board, providing some basic ingredients (keywords, topical points, the occasional outline) that I mix and elaborate in my own way.
If all this sounds like a functionalist metaphor rather than a critical one, it is. (Indeed, that’s how Gupta, Yasser, Mills, and Bali categorize it.) But I think it has critical potential. Allow me to take up an adjacent metaphor for a moment.
The way I use generative AI is akin to Peter Elbow’s idea of the writing process as cooking. In his classic Writing Without Teachers, Elbow explains that writing-as-cooking
consists of the process of one piece of material (or one process) being transformed by interacting with another: one piece of of material being seen through the lens of another, being dragged through the guts of another, being reoriented and reorganized in terms of the other, being mapped onto the other.
We can do this with AI-generated text just as we can with a quote from Dostoevsky or the political economy of Ghana. All it takes is treating AI-generated text as raw ingredients (flour, eggs, etc.) and transforming them through complex interactive processes.
We can also adapt Elbow’s cooking metaphor as a way of understanding generative AI, the technology, as cake. The raw ingredients of training data are mixed together algorithmically, interacting with novel prompts, to produce an output. It’s an imperfect metaphor—there is a lot more classification and selection in AI’s intermediate steps—but the output is the same: in most cases, a Twinkie.
Perhaps, then, it would be more useful to say that generative AI is what Socrates called “cookery” in Plato’s Gorgias. Socrates distinguishes cookery from true arts, such as medicine or justice, a distinction he ultimately uses to disparage rhetoric. I won’t defend rhetoric here.* Instead, I want to suggest that his accusations against rhetoric-as-cookery—that it is “routine,” “the habit of a bold and ready wit, which knows how to manage mankind,” “an experience in producing a sort of delight and gratification”—might be said of ChatGPT. It is built on an algorithmic routine. It has a bold and ready wit, in that it often makes assertions and holds to them, even when they are blatantly false. It can produce delight and gratification, especially when we marvel uncritically at its speed and fluency.
Getting beyond Twinkies and cookery requires one of two things: either treating AI output as ingredients rather than a product, or else adding measures of enough new and unique ingredients in the prompting stage to produce a more tailored output—in which case, perhaps we are closer to the artistry of baking after all.
AI as cake thus invites us to think critically, both about how generative AI works, and about how we are using it. I also love this metaphor because it is not inherently anthropomorphic. I don’t know about you, but I am tired of seeing robotic hands, robotic faces, and robotic bodies on every single post about AI. I think these images, metaphorical in their own right, perpetuate an illusion of AI sentience and imply the inevitability of AI ubiquity, even takeover. AI as cake offers an alternative vision: human-made, but not human; delicious, but unhealthy if consumed to excess.**
*As I sometimes-rhetorician, I have to stop myself from delving into a defense of rhetoric. If you’re curious, Aristotle is as good a place to start as any.
**Please consider contributing to Bad Ideas about AI and Writing: Toward Generative Practices for Teaching, Learning, and Communication, edited by Anna Mills, Mandy Olejnik, Miranda Rodak, Shyam Sharma, and me. Details at https://www.tinyurl.com/badideasbook.