In Fall 2025, the Take Back Your Brain Campaign held a cookie recipe contest for Carleton students. Instead of submitting tried and true recipes, however, students were asked to submit prompts for generative AI to create the recipes. All in order to answer the question: can AI bake?
Students were given a few rules for their prompts. They could mention cuisines and techniques, but no specific ingredients or flavors—that was for the AI to generate. Students could also add adjectives or other creative flourishes to steer the AI, save from telling it to generate something completely inedible. Finally, everything had to be spelled out in 150 words or less. In the end, thirteen students submitted prompts. To this pool the contest organizers added one more prompt that was itself AI-generated, rounding out the count with a wild card competitor.
Once all the prompts were in, the contest organizers ran each one through three different GenAI tools. They selected the recipe most representative of the prompt and sent each one, completely unedited, to a different volunteer baker. The bakers comprised fourteen faculty and staff members who generously donated their time and kitchens to the cause. Because they were told to follow the AI-generated recipes to the letter, the bakers also graciously endured the occasionally unintuitive or downright ludicrous instruction.
Dough (and runoff) from the “Miso-Sumac-Lime No-Bake Rolled Cookies”
The “Brown Butter Maple Pecan Bites,” just out of the oven
Once the baking phase was complete, it was time to taste the cookies and give out awards. The contest itself was held on November 20 in the Anderson Atrium. A panel of three judges—Profs. Arnab Chakladar (ENGL) and Dev Gupta (POSC) and Pres. Alison Byerly—graded the cookies on taste, texture, and originality. The cookie with the highest marks would be crowned “Best of the Best,” while the lowest marks would land the “Worst of the Worst” award. A third prize for “Most Interesting Prompt” would also be given out to the cookie produced by the most creative, most well-crafted prompt.
The event day
Before the judges arrived, though, students, faculty, and staff got a chance to socialize and try each cookie for themselves. Some of the cookies went quickly; others were still in abundance by the time the judges arrived—an early hint as to the final verdicts.
Once the judging began, attendees could watch the judges in real time as they tried the cookies and rated them. The judges in turn gave commentary and, often hilariously, their live reactions to some of the more questionable cookies.
Judges Byerly and Chakladar try the ill-fated “Miso-Sumac-Lime No-Bake Rolled Cookies”
In the end, the “Lavender Lemon Melt-Away Cookies” took home “Best of the Best,” earning the highest overall marks in taste, texture, and creativity. The “Fiery Ginger-Mint Campfire Cookies” were dubbed “Worst of the Worst” (edging out even the miso-sumac-lime cookies, the 100% AI competitor). And finally, the “Gluten-Free No-Bake Chinese Danish Croissant Cookies” were given “Most Interesting Prompt.”
If you would like to attempt to make any of the contest recipes, below is a cookbook containing the fourteen recipes, as well as some commentary on the event (which can be found in the below section as well).
Some Reflections
In 2024, Google Search’s “AI Overview” function went viral for suggesting home cooks use non-toxic glue to keep cheese from sliding off their pizza. Around the same time, it recommended that users eat one rock every day as a mineral supplement. Blunders like these are humorous and a little cringeworthy, but they can also become focal points for larger debates about the true efficacy of generative AI. Critics point to them as evidence that GenAI is far less “intelligent,” powerful, and helpful than advertised. Supporters see them as temporary stumbling blocks that distract from the real promise of AI. For its part, Google was quick to frame the mistake as just that, a one-off mishap.
Though not all the recipes were all that novel or successful, the results of this contest certainly suggest that GenAI has upped its culinary game since 2024. Most of the cookies were serviceable, and the cookies with the highest marks were completely gone by the end of the event. At the very least, it has become better at keeping bad recipes “merely bad,” and not totally inedible. Some of this is attributable to the fact that GenAI frequently pulls recipes, almost verbatim, from the internet. Experienced bakers will likely recognize parts of the AI recipes (collected above in the cookbook!) as snippets of other well-known recipes.
Of course, this improvement comes with concerns about copyright and intellectual property. AI companies are keen to insist that, with more data, AI will scale in capability indefinitely. But it’s unclear where this data will come from. If what’s past is prologue, we shouldn’t be surprised to see significant violations of intellectual property. In 2025, for example, Anthropic paid a historic $1.5 billion settlement to authors whose books the company had allegedly pirated to train AI chatbots. And in a recent IAI debate, Timothy Nguyen, an AI researcher at Google DeepMind, remarked that the democratization of knowledge brought about by AI will involve “winners and losers.”
Importantly, however, it isn’t immediately clear whether this claim of unlimited scaling is true. The field of AI research is still full of open questions, and it seems risky to stake so much on a prediction of this kind. It’s crucial, then, to try these tools out ourselves before making assumptions and commitments. It’s also important to remember that the success or failure of a tool today doesn’t guarantee its success or failure in the future. The surest way to evaluate GenAI tools and get a sense of what they can do is to put them in the hands of the experts who would actually be using them, or the community that they would be serving.