A Baker Contemplates Her Own Obsolescence, Again

An essay on the rise of AI and its cultural backlash by Bronwen Wyatt

A little less than three years ago, I wrote an essay about artificial intelligence and baking for Cake Zine’s Wicked Cake. I’d become obsessed with those first, comparatively primitive AI image generators. I used the DALL-E mini to make pictures of cakes (I am a baker, after all), plugging in prompts like “buttercream cake topped with fresh flowers” and pouring over the results.

Once, on a lark, I entered the name of my own business, Bayou Saint Cake. The nine images I was presented were bizarre: cakes ranging from a mottled army to acid green or electric blue and topped with oozing rosettes and swirls. They seemed to slide in and out of reality depending on where you looked. I picked one and reproduced it in real life for a video, mimicking the conjured icing with dollops of meringue, a shard of tuile, and candied kumquats. The response fascinated me. This cake was so different from my normal work that people couldn’t slot it into their expectations of me. It made some wonder if I was having a mental break.

In this first piece, I thought a lot about the uncanny valley—the feelings of revulsion inspired by technology’s imperfect attempts to mimic living people. These almost-human replicas can be genuinely unsettling, like the eerie animation of The Polar Express. The early AI-generated images I marveled at skirted the uncanny valley by nature of their own weirdness. They were, to be honest, kind of shitty. I think that’s part of why I liked them so much. It was so obvious that a human could do better. But even then, I wondered what AI could mean for the baking industry:

“To me, the true wickedness of these cakes is not the idea that my profession will be replaced by the machine, but that the beguiling ease with which the machines generate content—whether it be a recipe, a cake design, or a glossy image perfect for Instagram—will blind us to the less predictable charms of the human touch. What will be lost when the rough edges of handicraft are smoothed over in this digital future?”

This future has arrived far more quickly than I expected.

As generative AI improved, the images became sleeker, more poreless. When I scroll Pinterest now, I am confronted by countless photos of fake food: plump blueberries that reflect pristine highlights, rivulets of maple syrup pooling just-so over piles of pancakes that look FaceTuned in their perfectly burnished caramelization¹. When I click on the image, I’m taken to an AI-generated recipe blog that serves me ads for my favorite indie clothing brands. The “author,” pictured as a fresh-faced redhead, describes making the pancakes for her daughter’s birthday breakfast. Over on Facebook, images of AI cakes proliferate as targeted engagement bait. The formula is nearly always the same. An centenarian woman holds an elaborate confection, her face creased with age and her eyes imploring. She holds a chocolate layer cake topped with berries, or candles, or sprinkles. She’s decorated her own birthday cake at the age of 116! Won’t you celebrate with her?

The AI tools that are available to us have proliferated as well. The most famous, Chat GPT, is powered by a large language model, or LLM, a type of text-based AI that could possibly challenge Google as our primary way to answer written questions. LLMs are advanced models within the field of natural language processing, a form of artificial intelligence that underpins technologies like automated translation and transcription. In June, The New York Times’ Pete Wells published an article on how chefs are using Chat GPT to fine-tune menu ideas and walk them through unfamiliar techniques. For my own part, I’ve used Chat GPT to organize grocery lists based on what I have in my pantry and convert recipes from grams to cups. It is adequate at the former and exceptionally good at the latter. I scoop my flour with a light hand, so I’ve told the program that my cup weighs 120 grams, and it remembers.

This explosion of AI tools has been met by a significant cultural backlash—enough to make me pause in admitting my own Chat GPT use here. In my original essay, I didn’t touch on the environmental implications of AI. I have since read the horror stories of the extraction of resources by the data centers that power the technology. On the other hand, I could publish a recipe for a wholesome breakfast cake and you might make it with factory farmed eggs splashed with greenwashed branding, a cup of organic milk that required 66 gallons of water to be brought to your kitchen², and sugar processed in a factory that is fouling the air in a facility just down the Mississippi River from where I write this now. There is an inevitable environmental cost to the growing and cooking of food, but the stakes are higher for recipe developers and professional chefs. We often experiment with agricultural products in order to create something new. You can only offload so many failed attempts on friends, neighbors, or family meal. According to a UN estimate, food waste accounts for up to 10% of greenhouse gas production³—a figure over three times higher than estimates of data center emissions in the US last year⁴.

I recently published a recipe for strawberry shortcake, featuring lush, buttery biscuits piled with fruit and cream. All together, the ingredients for this single dish required at least 370 gallons of water to be grown or milked or milled. I know this because I converted a food cost spreadsheet I developed to tell me the water consumed by each ingredient in a recipe⁵. I built the spreadsheet to understand the price of using AI compared to the rest of my work.

As I plugged in the gram amounts of the flour, butter, cream, sugar, and strawberries into my water calculator, I began to feel ill. I’d tested the strawberry shortcake recipe at least five times, carefully fine tuning it to maximise its flavor and texture. The cumulative cost was enormous.

There’s a statistic that gets thrown around a lot—that asking an LLM to write one hundred words is like pouring an entire bottle of water on the ground⁶. By a conservative estimate, I’d drained two thousand, seven hundred, and ninety-six water bottles over the course of writing one recipe—no Chat GPT required. The time I spent composing this essay on my laptop and emailing it to the editors of Cake Zine cost about as much, in watt-hours, as asking Chat GPT over two hundred questions⁷. The dissonance between the performance of abstaining from AI and the acceptance of other, far more extractive behaviors has begun to feel profoundly unsettling to me. It would probably be more powerful, both ethically and environmentally, to abstain from chocolate. According to the Water Footprint Calculator, chocolate requires 129 gallons of water per ounce—14 more gallons of water per ounce than beef.

Many anti-AI conversations are taking place on social media, via algorithms that are curated by machine learning. Every time I log on, I am served advertisements for chic CPG food brands and plump hydroponic berries because artificial intelligence “knows” what I pause to look at more closely. When we write an email and are prompted to complete a phrase, it is because a language model has analyzed human composition habits. Another helpfully weeds out spam emails from your inbox⁸. You can never use an LLM and still interact with AI every time you use a computer.

I do not think it is possible to maintain an ethical boundary between the use of generative AI and other types of machine learning, as these models are increasingly interwoven in nearly every technological tool we use. I think many of us tend to think of AI as a monolith, rather than a dizzying array of instruments. My father-in-law used generative AI for voice banking, which would eventually reproduce the sound of his voice when he lost it to ALS.

If your primary concern is AI’s carbon footprint, then you can take steps to mitigate that, such as avoiding using generative AI to make videos, which does draw a great deal of energy. You might also consider other actions that would have a far greater impact, like eating a more plant-based diet. Author Andy Masley says as much in his analysis of Chat GPT energy usage.

I’ve read that drawing these comparisons represents a diseased and hopeless nihilism. If we assume that artificial intelligence is inevitable, or argue that it is neutral, then we are complying in advance with our own destruction. To be clear, I do not think the tech oligarchs have our best interests in mind. Human food bloggers could become obsolete as they are supplanted by automatons with perfect search engine optimization. An entire creative ecosystem of recipe developers, photographers, and food stylists could be wiped out by the potential of AI generated cookbooks, which are already cropping up. Who would pay me to develop a recipe when AI could do it for free?

My biggest concern remains the prioritization of well-paid human labor. The preservation of intellectual property amid the “plagiarism machines” of generative AI feels less salient to me. I agree that plagiarism is bad, though I probably have a looser approach to copyright than most. My cake decorating style has been copied by actual humans, uncredited, many times. The truth is that I don’t really care anymore. Not to get all Derrida on you, but I don’t think we’ve ever had control over how our work is recontextualized once it is out in the world. A culture of attribution is important, but also impossible to execute perfectly when countless people can have the same idea simultaneously.

AI on its own isn’t particularly good at making interesting art, or useable recipes (yet). On a bad day, I might also say that humans have been making derivative work long before Chat GPT came along to make the process instantaneous. Is a recipe sourced from “scraped” data all that different from one that a person lazily reproduced with just a few tweaks? Before Pinterest was flooded with AI blogs, you could still find countless versions of the same pancake recipe online. Either way, in the case of recipes the legal consequences are nil—lists of ingredients can’t be copyrighted.

If our goal is that more individual people will want to move through the messy, frustrating process of genuine creative work, sans the shortcuts of AI, then we will need to have a much larger conversation about capitalism, equity, and the distribution of wealth. If we want to affirm well-crafted recipes made by human hands, then we will need to seek them out and pay a living wage for them. That last bit is something that both individuals and food media companies have historically been loath to do, even prior to the advent of generative AI. In some ways, a profusion of crappy AI recipes might drive more traffic to any of us who can rise above the algorithm with our own work. And perhaps, as our feeds are flooded by uncannily perfect AI dupes, we should also accept that a handmade cake might start to look a little shittier in comparison—the less predictable charms of the human touch.

I don’t believe an abstinence-only approach is effective in solving the ethical dilemmas of artificial intelligence, if it were even possible. Yes, I could stop using Chat GPT forever tomorrow, at only minor inconvenience to myself. It is the best tool I’ve used to convert my recipes, but not the only one. Yet this decision alone would have an almost meaningless effect on my carbon footprint, or my overall, unconscious use of AI as it is embedded in other technological tools. In this way, Chat GPT is just one of many ethically fraught tools in the kitchen. We grapple with equally serious moral implications every time we crack an egg.

—Bronwen Wyatt

---

Follow Bronwen on Instagram

1. I dropped both the image and the recipe text into Winston AI to confirm it was AI generated.

2. https://www.theguardian.com/news/datablog/2013/jan/10/how-much-water-food-production-waste

3. https://unfccc.int/news/food-loss-and-waste-account-for-8-10-of-annual-global-greenhouse-gas-emissions-cost-usd-1-trillion#:~:text=In%202022%2C%20according%20to%20UNEP,chains%2C%20and%20actors%20into%20account.

4. https://www.technologyreview.com/2024/12/13/1108719/ais-emissions-are-about-to-skyrocket-even-further/

5. I used statistics from the Guardian and the Water Footprint Calculator.

6. From this WP article: https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/?_pml=1

7. For many of these comparisons, I’m using conservative estimates of the power usage of AI prompts based on this piece by Andy Masley. In reality it's possible the power use of Chat GPT text prompts is quite a bit less.

8. https://www.wired.com/story/how-to-use-ai-for-email-artificial-intelligence-tips-advice/