An Austin Chef Tried Developing Recipes With AI and Formed a Love/Hate Relationship With It

AI-generated art and information are on the tip of everybody’s tongue. A song that simulated Drake and the Weeknd’s vocals went viral last month and so have some strange commercials generated by the technology. AI is a divisive subject and conversations surrounding its use have generally revolved around issues of plagiarism and misinformation as major tech companies make AI features more widely available. However, it could influence life in the kitchen as well. In April, Austin chef Donald Mullikin of the North Austin food truck Bunboy decided to take the technology for a spin, experimenting with ChatGPT to create a yakitori dish accompanied by a trio of sauces. Mullikin walked away from the experiment with mixed results.

Initially, Mullikin was using the AI chatbot to help calculate food costs and expenses at his bao-centric food truck located at Oskar Blues Brewery. He says ChatGPT is a great tool for complex or multistep equations. “I’m terrible at math, so I use language to describe what the ingredients cost me in total, what weights a portion is, how many I’m expecting to sell on a weekly basis, and factor in things like overhead and operational costs,” he tells Eater Austin. “ChatGPT spits out an exact number of what I should think to charge for a dish in order to cover my expenses and hopefully make an industry standard profit margin.” Mullikin can tell the program the cost of every ingredient in a dish and the portions that are in each serving. In turn, the AI relays the cost per portion down to fractions of a penny.

ChatGPT’s basic knowledge of common food and labor cost percentages can help restaurants calculate how to turn a profit. It’s useful for chefs like Mullikin whose gifts aren’t math, and communicate easier through language instead of inputting data into other restaurant management software. Thanks to the service doing wonders in figuring out formulas, Mullikin thought it’d be a good idea to collaborate on a dish.

This did not go as well.

Mullikin wanted to put a yakitori trio on his food truck’s menu with three sauces — bacon, yuzu kosho, and miso-dijon. He knew the chicken skewers were going to get grilled on binchotan charcoal and what the three sauces were going to be. He collaborated with ChatGPT to tinker with specific ingredients, portions, and the balance of flavors.

Mullikin shared the transcript of his talk with the AI service with Eater Austin, which shows them going back and forth tinkering with the recipe. “You can disagree with it, make suggestions, and collaborate with it,” he says. “For instance, it has a tendency to be partial to red pepper powder in Asian recipes, but over time, learned my preferences towards Lao Gan Ma [a Chinese brand of chile hot sauce], kimchi juice, and other more complex spice additions. It clearly has a basic idea of balance in a dish,” he says, but they all lacked salt, which he adjusted before serving.

The limited AI-assisted yakitori trio went on Bunboy’s menu in early April and sold out after two days. While ChatGPT technically did its job and Mullikin is pretty happy with how the dish turned out, he probably won’t use it to help write recipes in the future. As far as administrative duties like scaling recipes, converting measurements, and cost analysis, it’ll probably stick around, but on the creative end, this is just a one-time experiment. “It takes a lot of the fun of failure and experimentation out of recipe [research and development]. However, as an idea generator, I can see this being a remarkable tool in the kitchen.” Mullikin likens the service to using the Flavor Bible when he first started cooking.

As for the future of ChatGPT in the broader hospitality world, Mullikin says can be asked to imitate the styles of well-known chefs, which he says can be cool, but is also a slightly scary thought. For example, Mullikin attempted to see if ChatGPT could emulate David Chang’s cooking, but the program fell short because it doesn’t have the capacity to take classic comfort dishes and make them nostalgic and playful, an element Mullikin says is essential to Chang’s style.

Technology news reporter Umar Shakir explains that using ChatGPT is very dependent on how the user is asking questions. (Umar Shakir is a news reporter for Verge, a property of Eater’s parent company, Vox Media.) “You’re conducting it,” he tells Eater. “So you need to know what to ask, how to ask it, and knowing what order to ask it, and you’re going to get different results every time.” He likens it to orchestra conductors: “The musicians have their music sheets but you still have the sway to how the music gets played.” For recipes, the AI pulls from the internet and online recipes ultimately combining “different ingredients and recipes that it sees might be similar.”

The way ChatGPT pulls information from data sets is making people rethink its very definition of plagiarism. Mullikin says AI’s use is “creatively bankrupt if this is how you came up with recipes in general,” and without human intervention and knowledge of how to create a dish, recipes would turn out bland. AI lacks in things like playfulness, nostalgia, and memories of bites had on vacations, or in grandma’s kitchen. It is basically a robot, and it can’t ever have the heart necessary to cook with love. At least not real ones anyway. Just ask the Tin Man.

Currently, the food truck is temporarily closed because Mullikin was injured.

With additional reporting by Nadia Chaudhury.