Recipes can be passed on orally, scribbled on index cards, published in cookbooks. But they’ve always had one thing in common: they’re made by humans.
Few things are as full of humanity as a recipe. Mixed and folded and baked into each are the backgrounds, stories, flavors and feelings of their creators.
However, people have their limitations. They can’t read every mashed potato recipe on the internet before coming up with their own version. They can’t analyze thousands of techniques in search of the best way to make a pie crust.
These recipes have all the components of their handmade ancestors: ingredient lists, accurate measurements, step-by-step instructions, and introductory notes with (invented) personal touches. Their advantage, in theory, is that they draw from a vast amount of online information about food and cooking.
But are they good? And can they enhance the millennia of culinary experience?
As home cooks, professional chefs, and food magazine editors know, the ultimate test for recipes is Thanksgiving dinner, a vast, diverse spread that stirs high expectations.
So we decided to turn to artificial intelligence — in this case a technology called GPT-3 — to come up with a holiday menu, which we then prepared and presented to a corps of taste testers: four New York Times cooking columnists.
The results say a lot about the potential of the technology and the actual purpose of a recipe.
Before we come to the verdict, let’s explain the science. Designed by OpenAI, one of the world’s most ambitious artificial intelligence labs, GPT-3 is a neural network, a mathematical system that can learn skills by analyzing massive amounts of data.
Some systems study images; in September, an AI-generated work won the top prize in a state art competition. GPT-3 analyzes digital text, including books, Wikipedia articles, tweets, chat logs, computer programs and, yes, recipes. It can identify billions of different patterns in the way people connect words, numbers, and symbols, then use that knowledge to generate its own content, such as a Thanksgiving menu with original recipes.
Artificial intelligence is about to reshape various areas, from email marketing to computer programming. Writing prescriptions isn’t a common field of study, but a handful of researchers, including a team from the Massachusetts Institute of Technology, have begun investigating whether AI can master the skill.
In 2016, Janelle Shane, an optics scientist who runs a machine learning humor blog called AI Weirdness, started using systems like GPT-3 to create recipes and then post them. Early versions of the technology, she said, produced recipes that were a little, well, weird. They called for nonsensical ingredients like “peeled rice” or “chopped flour”.
Today, she said, many AI recipes seem indistinguishable from man-made recipes.
“What it does really well sounds plausible,” said Dr. Shane. “So if you weren’t paying attention and someone just read this recipe out loud to you, you’d be like, ‘Oh yeah, that sounds like a very common recipe.'”
To create our AI Thanksgiving menu, we started to introduce ourselves to the GPT-3 system in a surprisingly human way.
Mark Chen, an OpenAI research scientist, advised me, Priya Krishna, to get personal. Tell the system about yourself, he said: your family background, what flavors you like, what ingredients you use often.
“The more details you provide in the prompt,” he said, “in general, the better the model performs.”
So after logging into GPT-3 on my laptop, I typed, “I’m originally from Texas and grew up in an Indian-American household. I like spicy flavors, Italian and Thai food, and desserts that aren’t too sweet.” Ingredients I often cook with are chaat masala, miso, soy sauce, spices and tomato paste.”
Then I wrote, “Show me a Thanksgiving menu made for me.”
The first recipe that GPT-3 produced was called “pumpkin spice chaat”. I was confused by the concept yet impressed by the inventiveness.
I asked follow-up questions to stimulate GPT-3’s creativity: show me a few desserts tailored to my taste preferences. Show me a non-traditional Thanksgiving recipe. Show me a recipe for cranberry sauce that is not too sweet and a little spicy.
Minutes later, I had a complete menu that seemed both plausible and intriguing: pumpkin spice chaff, green beans with miso and sesame seeds, nana topping, roast turkey with a soy ginger glaze, cranberry sauce that’s not too sweet and a little spicy (yes, that’s the full name). of the recipe) and pumpkin spice cake with orange cream cheese frosting.
The dishes already looked appetizing enough. We used DALL-E, another OpenAI system that generates images, to create a photo for each. And we asked GPT-3 to provide an introduction to each recipe, written from my point of view, “This roast turkey recipe is inspired by the flavors from my childhood.” (It was not.)
Some ingredient lists seemed questionable. The naan filling required 32 different components, including two cups of dried fruit. Most of the recipes were suspiciously light on salt and fat. Still, I had hope.
By cooking and tasting the recipes, those hopes were almost dashed.
The cake was thick and more savory than sweet. The naan filling tasted like a chana masala and a fruitcake that had ended up in a bar fight. The roast turkey recipe called for a single clove of garlic to flavor a 12-pound bird, and no butter or oil; the result was dry and tasteless.
The chaat, laced with coriander and baking spices, was a grassy-flavored mush. The green beans and cranberry sauce were edible but unremarkable.
Our tasteful columnists were even less kind.
“We’re not out of work,” Melissa Clark said. “I don’t feel anything eating this food,” Yewande Komolafe added.
Genevieve Ko summed it up best: “There’s no soul behind it.”
The recipes gave little clue as to what cooks should look for or smell during the process, and no reason why ingredients were added in any particular order.
Even before the tasting, Dr. Shane, the optics researcher, suggested lowering our expectations. She called AI-generated dishes “the recipe equivalent of hotel room art.”
Still, even at this early stage, the technology can have its uses. Mr Chen, the OpenAI scientist, said it could help give home cooks some basic inspiration.
“You have something in mind that you want to make, you don’t know exactly how to make it, or you have a set of ingredients and I’m not quite sure how to combine it,” he said. “This can be a really quick way to give you a lot of different ideas.”
GPT-3 can also help personalize an existing recipe by making a version that’s less spicy, or that has specific flavors, he said.
But cookbook author Nik Sharma, whose work intertwines food and science, is concerned about the inherent biases of these programs. They draw on the English-language internet, where western recipes predominate. And if these programs suggest non-Western dishes, he added, they might be the more common ones Americans are familiar with, such as tandoori chicken.
“As food writers, our goal is to take people in a new direction, to help them challenge them, to help them understand the world,” he said. With artificial intelligence: “how do you do that?”
AI has already made its way into cooking. Home cooks use search engines to find recipes and ask virtual assistants like Amazon Alexa to convert teaspoons into tablespoons.
So far, this technology is not a substitute for humans. It can push cooks in one direction or the other. But it’s still humanity – with its intuition, storytelling and warmth – that drives a good recipe.
The individual behind the dish is just as important as the recipe itself, said Dr. Shane. It’s valuable to know that someone took the time to perfect and share a version of pie, stuffing, or turkey.
“You want to know that these recipes mean something to someone,” she said. That’s something artificial intelligence may never be able to provide.
