AI can generate and create recipes, but food bloggers aren’t happy: NPR

Apple released iOS 18 this month. The update, which accompanies the release of the latest iPhones, Apple Watches, and AirPods, includes an expanded application of artificial intelligence called “Apple Intelligence.” Apple isn’t the only company integrating AI into its operating system: Samsung’s S24 devices and UI 6.1 update include Galaxy AI-enabled elements, and Google’s smartphones will soon have Gemini AI.

Many companies are touting new phone features that allow users to use AI as a recipe-generating assistant. In the Apple Intelligence demo, a user can ask Siri to cook a dinner party using ingredients they have on hand, and the AI ​​will respond with a list of recipes that use those ingredients. While this sounds useful, most of the coverage of AI and cooking to date has been negative.

For years, chefs on YouTube and TikTok have competed in cooking contests between “real” and AI recipes, with the “real” chefs often winning. In 2022, Tasty compared a chocolate cake recipe generated by GPT-3 to one developed by a professional food writer. The AI ​​recipe baked well, but the food writer’s recipe won in a blind tasting. Tasters preferred the food writer’s cake, which had a more delicate, less sweet flavor and a denser, moister crumb. Meanwhile, the AI ​​cake was sweeter and drier.

AI recipes can also be dangerous: Last year, Forbes reported that an AI recipe generator had created a recipe for an “aromatic mix” after a Twitter user instructed it to make a recipe using water, bleach, and ammonia, which actually released deadly chlorine gas.

With AI-generated recipes, novice cooks could risk a bad meal or a life-threatening situation. For food bloggers and recipe developers, the technology could pose a threat to their livelihood.

Sisters Sarah and Caitlin Leon are one half of the family behind The Woks of Life, a food blog focused on sharing “recipes, kitchen adventures, and travels.” They started the blog with their parents, Bill and Judy, in 2013.

The recipes in The Woks of Life begin with what Sarah calls an “ideation stage.” “Sometimes we discuss them as a group,” she says. “Sometimes we fulfill recipe requests from our readers. Sometimes they’re completely new recipes, and they require a lot of research and experimentation — going to restaurants to try the food, watching videos, or searching the Chinese internet for ideas.”

After coming up with an idea, the Leons test their recipes up to 40 times. “It took my dad a year to come up with the recipe,” Sara says. A recipe must be approved by all four family members before it’s published. “We know our readers are entrusting us with their ingredients and their time, so we make sure our recipes not only work, but are also easy to read and follow,” Sara continues.

This recipe development process is also one of cultural connection and understanding for the sisters. “We’ve had experiences where we realized we didn’t know all that well how to cook Chinese food,” Caitlin says. “All of that is reflected in our blog. We’re always learning, and we’re always trying to find new techniques and ingredients.”

“The stories behind these recipes and the connections you make with people through them are very human,” Sarah says. That’s why the sisters are skeptical of AI-generated recipes: “Machines don’t eat, they don’t taste, so what are they?”

One area where AI can play a role is recipe development, said Andrew Olson, a software engineer who develops recipes for his cooking blog, One Ingredient Chef, which features whole-ingredient recipes.

Olson began experimenting with GPT-2, a rudimentary version of the ChatGPT software, in 2019. “I was already thinking about how we could use this for recipe development, to help people come up with new and creative ways to cook,” Olson says.

In 2023, he released DishGen, a tool that leverages AI for cooking-specific output. The website allows users to input an ingredient list and generates recipes that look like they’d be found in a cookbook. Each recipe also includes a headnote with a sensory description of the end product and suggestions for when and where to serve each dish. The recipes have little touches that are reminiscent of his recipe copywriting style: the cheese is “generous” and sprinkled, the textures are “harmonious,” and the muffins are “healthy.” The premium version of the software also generates images that show what the end product of the recipe will look like.

Olson is aware of the negative press: “Google is encouraging people to put gasoline in their pasta,” he says, “so DishGen is very focused on safety.” If you enter ingredients that could form a toxic combination, like the ingredients in chlorine gas, the website won’t generate a recipe, but instead displays a short error message.

The Leons don’t expect an AI recipe generator to be able to recreate the same sensory experience as a human recipe developer, or take into account variations and special touches. “What meat blend do you use? What seasonings do you use to get the right amount of meat? How much salt? Is the salt affected by adding cheese? Because cheese is salty.” Because AI doesn’t eat or taste food, it instead synthesizes content it pulls from the internet and uses existing, human-tested recipes to inform the recipes it generates on its own.

“These companies are taking real human-created content and training AI models without giving credit, attribution or compensation of any kind to the people who created that content, and then competing directly with the people who created that content. That’s a huge existential threat,” Sala says.

Olson sees it differently: “A lot of times[recipe development]is taking inspiration from other recipes — ‘Oh, this is good, but I could do it in a different way,’ or ‘I could add something else,’ and this technique is no different,” he says. “They’re taking inspiration from something that’s publicly available, but they’re not stealing it or recreating it word for word.”

“I’m not completely pessimistic,” says Sarah. “I think AI can be used in a brainstorming context, where you can talk about storage, you can talk about how long this condiment will last in the fridge, or you can talk about this particular ingredient and elaborate on it.”

Olson agrees: “I think food bloggers can use (AI) to be more creative and come up with new ideas,” he says. “But I don’t think the technology is there yet to be able to have a completely AI-generated blog. But it would be a cool concept, and someone could try it and see what happens.”

While they wait for AI technology to reach that level, the Leungs are focusing on their family’s stories to avoid people mistaking their blog for AI-generated. Many hobbyist cooks have long been frustrated by the long, unrelated stories they have to scroll through to find recipes in blog posts. “I think, weirdly, people are going to look for markers that say it was made by a human, which is a story,” Sarah points out.

This story was edited by Suzanne Nguyen.