So you think your newfound ability to prompt ChatGPT for AI-generated recipes could result in a culinary masterpiece?
Hold that thought, advises the World of Vegan, a popular wellness website focused on vegan living. The site recently undertook an intriguing experiment powered by generative AI, where they prompted ChatGPT to conjure over a hundred diverse recipes. The group prompted the AI bot to whip up new and innovative recipes for a variety of occasions ranging from date night dishes to brunch and dessert ideas. From there, the site’s chef team tested each recipe to see how they tasted.
The result? Not good.
All this prompting led to what the group described as “hilariously pitiful results.” With many of the recipes, the chef team at World of Vegan spotted ingredient formulations that “would clash right away and where the mishaps would occur.” The team also felt the recipes were largely “deceptive,” seeming ordinary at first glance but often described as “rich” and “decadent” when they were quite the contrary.
“I had a feeling ChatGPT would struggle with recipe development, since developing recipes is such a delicate mixture of fine art and science,” World of Vegan founder and chef Michelle Cehn told The Spoon. “But I was shocked by just how difficult it was to find a single spring recipe written by ChatGPT that worked with a passing grade. This is a crucial warning for both food bloggers seeking shortcuts and home cooks looking for quick recipes. You’ll save yourself a lot of trouble (and wasted time, energy, and money) by bypassing ChatGPT and opting for a trusted blogger’s highly-rated recipe instead.”
One of the biggest fails cooked up by the World of Vegan team was a vegan scalloped potato dish (pictured above), which the recipe’s chef said had an ingredient list and cooking instructions that were out of order. The resulting dish had an off-putting color, a pungent sauce, and tasted bad.
According to World of Vegan, out of the 100 or so recipes the team cooked up, only one – a cauliflower taco dish – resulted in an appetizing result.
Cehn believes the resulting 1% success rate might be due to ChatGPT’s reliance on what is essentially flawed data, namely millions of subpar recipes drawn from the Internet. With this as its foundation, things are destined to go poorly once the bot is tasked to create a unique recipe.
“A human brain can’t access all that information, so people are likely independently (and unintentionally) creating duplicate recipes online. Since ChatGPT must create a truly unique recipe, it has to get a little weird to create one that’s not plagiarized.”
While one might expect a site focused on creating recipes to be skeptical about AI filling its shoes, I don’t doubt the poor results are that far off from what others may find if they conducted a similar experiment. Good recipes often result from lots of experimentation and applied knowledge, something that you don’t get when a bot freewheels up a new dish idea out of thin air.
And while a more specialized AI trained on the compatibility of various culinary ingredients – something akin to a chatbot based on Chef Watson – might yield better results, we don’t have that, at least not yet.
Bottom line: human-powered recipe creators are still necessary…for the time being.