Last month, grocery delivery service Instacart sparked a horrified reaction over its AI-generated recipes after a report by tech publication 404 Media. Some contained ingredients that didn't exist (what was “monito sauce,” people wondered?) and stomach-churning images (a questionable “hot dog stir-fry” had sausage in it. (The inside of the sausage was depicted looking very similar to the inside of a sausage). tomato).
A growing number of apps are emerging that promise to make AI an essential tool in the kitchen. Additionally, food brands such as Heinz and Avocado From Mexico are looking to integrate this technology into their websites.
Curious chefs are tinkering with AI themselves, using systems like OpenAI's ChatGPT and Google's Gemini. Ethan Mollick, a professor at the University of Pennsylvania known for experimenting with viruses using AI systems, says writing new recipes is one of the most common ways people start using technology, along with creating fairy tales and birthday cards. He said they have become one. “While this is a demonstration of the power of AI, it is also one of the worst use cases,” he said. “That doesn't mean it won't work. But it's very likely some weird nonsense.”
That may be because AI isn't as smart as we expect, especially when it comes to food, experts say. Emily Bender, a computational linguist at the University of Washington who studies AI, says you may think you're asking a question, but you'll get an answer. But all it does is spit out a series of words similar to what you previously “read.”
Thanks to years of blogging and social media posts, the internet is full of recipe web pages and food photos that AI models can now provide relatively accurate reproductions of. But because they're designed around two-dimensional words and colors, they don't capture all the multisensory experiences that make good food great. They also don't have noses or taste buds, so they can't check their own work.
Bender noted that AI systems may often encounter chicken and beef in similar contexts, for example. “So when the language model accesses the parts of the training data that are relevant to a recipe, there is no good way to distinguish between them,” she said. “You talk about the beef being rare or well-done, but you're not talking about the chicken. But that might be revealing.”
Undercooked chicken is just one potential danger. “We also have a very high ability to put together things that could be toxic or harmful, or that could interact with people's medications, but there's no way to alert us to that or notice that.” said Margaret Mitchell, a computer scientist at Hugging Face, a prominent public company. Source AI startup. Last year, a New Zealand supermarket accused a chatbot that suggested customers use ingredients they had on hand to provide recipes for “dishes” including toxic chlorine gas, a “bleached rice surprise” and turpentine-flavored French toast. It became news. . And an AI-generated book on mushroom foraging sold on Amazon was found to contain potentially deadly advice.
Some AI-generated recipes have obvious and fatal problems, while others have more subtle problems. For example, many recipes created on the Instacart website do not list ingredients in the order they should be used. This is a technique that most well-crafted recipes employ to make the recipe easier to understand. They don't always give the reader visual clues about cooking times and various steps, which can confuse inexperienced cooks. Also, serving sizes may vary. One serving of “Cowboy Steak with Herb Butter” (a 2-inch-thick ribeye steak) will probably end up eating at least a few cowpoke.
One of the problems with these bot-generated recipes is that the people who need the most help – novice cooks – are the least likely to notice their shortcomings.
Things get even weirder when it comes to food images. AI generators tend to invent dishes that are physically impossible, such as disappearing ingredients or shapes dictated by gravity. Food delivery apps DoorDash and Grubhub have used the technology to sometimes hilarious effect, with photos of shrimp with two tails, a pizza described as a “pie” and a sweet dessert reported in the report. It is said that an image is attached. What about Instacart's cowboy steak? It looked like Frankenstein's butcher had pieced it together and the grill marks were mismatched.
As with political news, there's always a hierarchy when it comes to recipes. There are mainstream traditional publications (such as the Washington Post) as well as major professional sources such as Bon Appétit and Cook's Illustrated, whose recipes are thoroughly tested and carefully constructed. Cooks can also find vetted recipes on reputable sites like Serious Eats and Food52. Cookbook authors are often the go-to source for well-crafted products. Additionally, there are numerous food blogs and influencers out there, some of which offer high-quality recipes, while others are questionable at best.
But the internet tends to flatten all that out. Casual users may not know the difference between a reliable source and an incompetent amateur. Many cooks who pick up the first recipe that comes up in a search may quickly realize that just because a human created a recipe doesn't mean it's good. Can't AI do at least more than that?
David Eastwell, a British physicist with a side job as a stock photographer, started playing around with generative AI tools during the pandemic lockdown and quickly realized he could create photos that were just as good as the ones he'd taken himself. . Curious about what more could be done with technology, he set out to create an entire website and chose a culinary theme just for fun. “Actually, I'm a very average cook,” Eastwell says. The resulting product, Air His Fryer, gives his chef recipes for dishes like nachos and soy-marinated cod. All of these recipes were generated by Eastwell using his AI tools.
Eastwell said he is running the site as an experiment. He was interested not only in the limits of technology, but also in its possibilities. AI can create workable recipes, but it can also fail, he said. As with recipes that come up with real people, readers need to be careful. “If you search for Roasted He Potato recipes online, you might find a recipe from Jane's cooking school.''You might not know who Jane is, but it's because it's made by humans.'' “There is an implicit trust written in,” he said. “It might still be trash.”
Mollick evaluates AI-generated recipes and other creations using what he calls the “best humanly available” standards. AI probably won't be able to match the creations of trained recipe developers, but amateur cooks may find something they like or even love. What an AI trained on recipes from around the world makes. In January, when Wired magazine conducted a test in which bar patrons were given two different cocktails (one from the bartender and one from an AI) based on their preferred taste, half of the patrons were unsure which one they would choose. I couldn't tell if it was man-made or not. Some people preferred drinks made by AI mixologists.
Some users rely on AI for highly customized recipes for special meals or to find ways to combine specific ingredients they have on hand. For example, DishGen, a subscription service that offers AI-generated recipe tools for $8 a month and offers promotions. This in itself contributes to reducing food waste. Still, experts say users should be careful. DishGen's recipes come with a disclaimer stating that the company has not “tested them for accuracy or safety” and that you should use your “best judgment when preparing AI-generated dishes.”
Heather John Fogerty, a food writer and recipe developer who teaches journalism at the University of Southern California, says bots will never be able to replace humans when it comes to teaching other humans how to prepare food. Sure. For one thing, she pointed out that there are often factors that aren't taken into account with robots. Scaling up a restaurant recipe that serves 40 people to something suitable for home cooking isn't just a matter of division, she says. Differences between commercial and home equipment need to be considered, and it is necessary to understand the ingredients that cannot be reproduced with AI.
More importantly, there is something intangible in the creativity and sensory awareness that goes into a recipe. “The question is, why do we still buy cookbooks and turn to publications like the Washington Post for food content? And the answer is the human element.” said Fogarty. “Context cannot be underestimated. There is real art in recipe headings and recipe notes that cannot be generated by AI.”
Mitchell sees the AI-generated recipes as more novel than practical. Despite already having options such as tools that can classify existing human data, companies are offering features on their websites and apps that promise to make AI as much a part of people's kitchens as their ovens. She pointed out that companies may be rushing to adopt technology by developing or creating recipes and selecting the best recipes instead of creating new recipes. “Now that we have hammers, people say everything is a nail,” she says. “There are many other ways to automate recipe generation that are very likely to work. Generation AI should not be the starting point.”