AI can generate recipes that can be deadly. Food bloggers are not happy

Apple released its iOS 18 this month. The update, which came with the release of the latest iterations of the iPhone, Apple Watch and AirPods, includes expanded applications of artificial intelligence, called "Apple Intelligence." Apple isn't the only company to integrate AI into its operating system. Samsung's S24 devices and UI 6.1 update included Galaxy AI-supported elements, and Google phones will soon feature Gemini AI too.

Many companies have advertised a feature in their newer phones that lets users utilize AI as a recipe generator assistant. In the Apple Intelligence demo, a user asks Siri for a dinner party meal plan with ingredients they have, and the AI returns a list of recipes using those ingredients. While that seems convenient, most of the press up to this point about the relationship between AI and cooking has been negative.

For years, chefs on YouTube and TikTok have staged cook-offs between "real" and AI recipes — where the "real" chefs often prevail. In 2022, Tasty compared a chocolate cake recipe generated by GPT-3 with one developed by a professional food writer. While the AI recipe baked up fine, the food writer's recipe won in a blind taste test. The tasters preferred the food writer's cake because it had a more nuanced, not-too-sweet flavor profile and a denser, moister crumb compared to the AI cake, which was sweeter and drier.

AI recipes can be dangerous too. Last year, Forbes reported that one AI recipe generator produced a recipe for "aromatic water mix" when a Twitter user prompted it to make a recipe with water, bleach and ammonia. The recipe actually produced deadly chlorine gas.

With AI-generated recipes, casual cooks may risk a lousy meal or a life-threatening situation. For food bloggers and recipe developers, this technology can threaten their livelihood.

Sarah and Kaitlin Leung are sisters who make up one half of the family behind The Woks of Life, a food blog focused on sharing "recipes, kitchen exploits, and travels." They started the blog in 2013 with their parents, Bill and Judy.

Recipes for The Woks of Life begin in what Sarah refers to as the "ideation phase." "Sometimes we have a group conversation," she says. "Sometimes it's about fulfilling requests for recipes that we are asked for by readers. Sometimes it's totally new and requires a lot of research and experimentation, going out to restaurants to eat that dish, watching videos, or scouring the Chinese internet for ideas."

After an idea is conceived, the Leungs will test a recipe up to 40 times. “It took my dad like a year to come up with some of his recipes,” Sarah said. All four family members have to sign off on each recipe before it gets published. “We know that our readers are trusting us with their ingredients and time. So we try to make sure our recipe not only works but also reads well and is easy to follow,” Sarah continued.

This recipe development process is also about cultural connection and understanding for the sisters. “We had the experience of realizing that we didn’t really know how to cook Chinese food that well,” Kaitlin said. “All that is really reflected in the blog. We’re still always learning, and always trying to make sure we’re finding new techniques and ingredients.”

“The stories that surround these recipes and the connections that we make with people through these recipes — it’s so deeply human,” Sarah says. That's why the sisters are skeptical of AI-generated recipes. “The machine doesn’t eat and the machine can’t taste. So what is it?

Andrew Olson believes AI has a place in the recipe development space. He's a software engineer who develops recipes for his food blog, One Ingredient Chef, which has recipes focused on featuring one whole, unprocessed ingredient.

In 2019, Olson began experimenting with GPT-2, a rudimentary version of the ChatGPT software. “I was already thinking about how it could be used for recipe development and to help people come up with new creative ways of cooking,” Olson said.

In 2023, he released DishGen, a tool that harnesses AI for cooking-specific outputs. On the website, users can input a list of ingredients to generate a recipe that looks just like one from a cookbook. Each recipe even includes a headnote with a sense-based description of the final product and suggestions for when and where to serve each dish. Within the recipe, there are little flairs that evoke the recipe copy style. Cheese is sprinkled “generously,” textures are “harmonious” and muffins are “wholesome.” Premium versions of the software even generate images of what the recipe’s final product could look like.

Olson is aware of the negative press. “Google’s telling people to put gasoline in their pasta," he says. "So DishGen has focused a lot on safety.” If you provide ingredients that may have toxic combinations, like the components of chlorine gas, the website will not generate a recipe, instead sending a short error message.

The Leungs don't think AI recipe generators can replicate the sensory experiences and account for the same variations and special touches that human recipe developers can. "What blend of meat are you using? What seasonings are going in for the right amount of meat? How much salt is there? Is the salt affected by the addition of cheese, which is salty?" Because the AI isn't eating or tasting the food, it instead amalgamates content pulled from the internet and uses preexisting, human-tested recipes to inform the recipes it generates on its own.

“These companies are taking content created by real people, not giving credit or attribution or any kind of compensation to the people that created the content to train their AI models, and then competing directly with those people who created that content. So it’s a huge sort of existential threat,” Sarah says.

Olson sees it differently. “So much of [recipe development] is getting inspiration from other recipes you’ve seen. Like, ‘oh, that’s cool, but I could make it a different way’ or ‘I could add something else.’ I don’t see this technology as any different,” he says. “They’re getting inspiration from what’s publicly available, but they’re not plagiarizing it or reproducing it word for word.”

“I’m not totally doom and gloom,” Sarah says. “AI — I think it can be used in a brainstorming context. You could talk about storage, how long this condiment could be stored in the refrigerator or you could talk about this particular ingredient and elaborate on it.”

Olson agrees. “I think food bloggers could use [AI] to be more creative, to come up with new ideas,” he says, “but I don’t think the technology is there to the point where you can have an entirely AI-generated blog, although that would be a cool concept. Maybe someone should try it and see how it goes.”

As the Leungs prepare for AI technology to reach that point, they are making sure their blog won’t be mistaken as AI-generated by leaning into their family stories. Many casual cooks have long complained about the lengthy and sometimes irrelevant stories they have to scroll past to find a recipe in a blog post. “Weirdly enough,” Sarah notes, “I think that people are going to be looking for those markers that a person created. Like, this is a story.”

Suzanne Nuyen edited this story.

Copyright 2024 NPR