AI is making a wildly adventurous and sustainable plant diet possible
It looks like milk. It tastes like milk. But the ingredients include pineapple juice and cabbage concentrate—brought to you by a team of biochemists and computer scientists.
By Veronique Greenwood
The carton looks like every other carton in the cooler. The white cardboard pyramid, dewy with condensation, has the usual screw-top cap and waxy feel. It recalls school lunches, morning breakfast cereal, models bearing frothy white mustaches on billboards alongside the words, “Got milk?”
But where the other cartons list the contents you’d expect, this one has something different emblazoned on the side. It reads, in big black letters: “NotMilk.” Just in case you didn’t get that the first time, there is also a helpful diagram. Crossed out by a thick black line is the silhouette of a cow.
The liquid within looks exactly like milk. It tastes almost exactly like milk. But it was designed by biochemists and computer scientists who combed databases of plant-based ingredients using artificial intelligence, trying to predict what combinations would yield the milkiest not-milk. The final list developed by the Not Company, a Chilean startup that uses AI to create mimics of animal products, includes pineapple juice and cabbage concentrate. NotChicken nuggets draw on ingredients like fava beans and peach powder, without tasting like fava beans or peaches. There’s a NotBurger, too, and NotIceCream, available in the South American markets where the company got its start, as well as in Mexico. NotCo expanded into the US in late 2020—NotMilk was their first product available in US markets.
Looking at a carton in the dairy aisle at a Whole Foods near my home, two central questions came to mind: How have they accomplished this technological feat? And more significantly, Why?
Why produce something that tastes as close as science can make it to milk, and then loudly trumpet its falseness? The short answer can be seen all around you in the climate crisis.
First, the Why. Why produce something that tastes as close as science can make it to milk, and then loudly trumpet its falseness? The short answer can be seen all around you in the climate crisis. More than a quarter of all greenhouse gas emissions come from agriculture. The majority of agriculture’s emissions can be traced back to eating meat and fish. It is hard to conceive of anything more wasteful than meat production, when it comes to producing food. A third of all arable land on the planet goes to feed livestock. At the same time, a serving of plant-based protein requires up to 1000 times less water than an animal version to make. “I decided early that plant-based was one of the best mechanisms to create a very sustainable and at the same time delicious food landscape,” said Matias Muchnick, NotCo’s CEO.
But clearly the sentiment alone is not moving any needles. Despite a growing mountain of evidence on the environmental costs, consumption of animal-based foods is not declining; indeed it is escalating. In the next decade, according to the FAO, global meat consumption is expected to climb by 14%. In developing countries, eating animals is a sign of arrival in the global middle class; in the rich world, it’s simply an old habit.
If a habit is linked to someone’s identity, like eating meat often is, making the change is more complicated than just buying lentils instead of hamburgers, write Joop de Boer and Harry Aiking, a pair of scientists with backgrounds in psychology, biochemistry, and microbiology who have been collaborating for years on issues in environmental studies. When the researchers asked why people resist moving to plant-based diets, they found evidence from prior research that the work—the inconvenience of learning new recipes, the trial-and-error of finding new favorites—was easier for people to handle if they felt it meshed with their own identities. (There are many other factors involved, Aiking points out, but this is one.) A study by other researchers looking at why people did not eat healthier found that a surprising amount of the time, the culprit was inertia. People did not want to learn a new way of cooking; they did not want to learn new tastes, even if they knew their current ones were a problem. “We eat what we like,” says Aiking, reflecting that matters of health, as well as matters of sustainability, ought to drive change. “But it simply doesn’t happen.”
Enter, then, the simulacra.
Pat Brown, the Stanford biochemist who launched the Impossible Burger, spent some of the company’s early days on his knees in a clover field, in search of the molecules that would allow him to recreate the taste of beef. (Impossible Burgers use heme, a molecule found in blood and in plants, in aid of this goal.) A perfectly done Beyond Burger, a pea-protein-based patty, will seep red onto the burger bun just like a beef patty, thanks to a formulation that includes beets. NotCo’s products, in turn, are all nearly pitch-perfect mimics of the original, animal-based versions.
“We are not launching it if we don’t have statistical parity with what we are trying to replace,” says Muchnick, meaning that in tests, people would mistake the plant version for the animal one. “We came to the US understanding that 60% of plant-based milk consumers are not completely happy with the taste of the product they are buying, and 30% of dairy milk consumers are willing to change if they found something that tasted like milk. That was a big chunk of the market—that’s why we came with NotMilk.”
AI learned more about what molecules make milk the way it is, and what molecules in plants taste like, it started to make matches between them. . . The effects are truly something no human would have expected.
Now, for the How: NotCo’s central product isn’t really, when you get down to it, milk. It’s a set of artificial intelligence algorithms that look for connections between the qualities of some of our most beloved foods and the taste, smell, and texture of ingredients from plants. The technology, whimsically named Giuseppe, has the potential to make it faster and easier to design plant-based foods that meet people where they’re at—getting them onto a plant-based diet without making them give up tastes they love.
Usually, when a team of food scientists put together a snack cake or microwavable burrito, they make educated guesses about how to extend its shelf life, how to make it compatible with manufacturing equipment—or how to remove that slight chalky note in a chocolate coating or that too-artificial tang in a strawberry jelly. They create a prototype, they take a bite, they repeat.
It is a laborious process for even well-understood foods, much less novel ones. When it came to making convincing plant-based mimics of milk and other animal-based foods, Muchnick felt the process failed spectacularly. For one, it’s an inherently conservative process based on making small adjustments to known quantities. The NotCo team was aiming for something more disruptive.
Muchnick and his co-founders thought a form of artificial intelligence called machine learning could break the mold. AI is adept at finding patterns. Maybe, if they could give an AI tons of information about what humans like to eat, alongside tons of information about plants, it could see patterns where food scientists had failed.
So, they started to teach an AI about food.
The team created some of the first databases they gave to the AI themselves, tasting all kinds of food, good and bad, and recording their impressions. “You can imagine how many bad things we tasted,” said Muchnick. But it was in the service of their vision: If a person, or for that matter an algorithm, is going to match a food with a perfect plant-based mimic, “you need to understand the world of plants and animal-derived foods,” he said, “but most importantly, their connection with the human brain.”
Imagine taking a sip of milk. It’s sweet, mild, creamy, and sometimes, depending on the brand, even a bit grassy. Swish it around. It’s thicker than water or coffee. And after you swallow, there’s a feeling that lingers on your teeth and tongue. Look at what’s left in the glass: It’s opaque, white, with a very faint hint of blue. It’s not quite like anything else.
But, as the AI learned more about what molecules make milk the way it is, and what molecules in plants taste like, it started to make matches between them. It started to suggest mixtures of plant ingredients that might do something similar.
One day one of NotCo’s food scientists came up to Muchnick with a green liquid in a glass and asked him to try it. “It tasted exactly like milk,” recalls Muchnick. Then they used the algorithm’s suggestions for how to get the color closer to what people expect. “You cannot sell [green] milk on the mainstream market,” he says.
The ingredients don’t always seem straightforward. NotMilk contains pineapple and cabbage, which on their own are about as far from milk as you can imagine when it comes to taste. But in the right amounts and in combination the right bedfellows, the effects are truly something no human would have expected. “In the chemical interaction with everything else in the ingredient list, [cabbage] does its own thing to get your mind into the milk experience,” said Muchnick.
The milk experience—without the milk.
NotMilk is asking if environmentalists might be willing to get into bed with processed foods, if it means more people eating plants? If it’s healthier? If it’s more transparent?
In 2022, NotCo announced some big news: It had plans to set up a joint venture company with food behemoth Kraft Heinz. It’s not uncommon for small, scrappy food start-ups to get swallowed by the big guys, but the joint venture, which is operated largely under NotCo’s control, was a calculated move by Muchnick and his co-founders in pursuit of their broader goal. If people have allegiances to familiar foods, they have allegiances just as significant to brands. “You cannot underestimate the importance of familiarity of brands when it comes to penetrating the mass market,” said Muchnick. “Kraft is well-known [in] American households.”
When we walk into a grocery store, the visual cacophony of packaging and advertising can be overwhelming. It’s easiest to go with what is familiar and reliable. Muchnick speculates that if NotCo launched a plant-based cheese or a macaroni-and-not-cheese dinner, not that many people would necessarily take a chance on the product. But under the umbrella of Kraft, with that familiar blue logo and comforting orange color, a lot of people who otherwise would not be buying plant-based products might give the stuff a try.
When I first spoke with Muchnick, this cheese was still officially a hypothetical. But in November of 2022, NotCheese Kraft Singles, the first product yielded by the joint venture with Kraft Heinz, arrived in American stores. Their testbed was a set of 30 Giant supermarkets in the Cleveland area. Within that market, the results were promising, said Muchnick: “We became the Number 1 plant-based cheese in less than three months.” He thinks that it’s easy for people to pick up the Kraft packaging, in that moment drawing on the emotional legacy of the brand. And then when they realize the product is plant-based, they’re willing to give it a go. (As of July 2023, the slices were in more than 2,500 US supermarkets.)
What’s important is that consumers actually stop buying the animal-based product, though. A world in which people just eat more cheese, whatever its origin, is not moving us to a better climate situation—they need to give up what the simulacra mimic.
That, it turns out, is not as straight-forward as one might assume Early market research on ground beef alternatives has delivered some counterintuitive results. In one study, researchers built an economic model to see how the production and consumption of beef in the US, would change if the price of Impossible Burgers, Beyond Burgers, and other similar products decreased. As expected, as prices dropped, people’s appetite for plant-based meat grew. But cattle production didn’t budge. The reasons behind this seeming paradox come down to the interplay between consumption and production; or put another way, consumer appetites are not the only factors that steer the market.
The same dynamic played out in a real-world study in the UK. A big supermarket chain there ran a month-long campaign called “Veganuary.” They offered promotions, discounts, and meal ideas for meat alternatives to increase the visibility of these foods. At the end of the month, sales of plant-based meats grew by an impressive 57%. But when researchers looked closer, they found that people were adding plant-based meat to their shopping carts but not taking the chicken, pork, and beef out. In fact, trends show that chicken, pork, and beef sales remained remarkably stable and insulated before, during, and after the event.
Still, these are early days, and there is reason for optimism as markets mature, products improve, and old habits die. When NotMilk became available in Starbucks in Chile, it ended up taking over about 15% of all the orders, perhaps displacing some cows’ milk. If people in NotCo’s new markets do the same—if food designed with help from AI starts displacing animal versions—they might be onto something.
But to truly make a dent, Notco and other plant-based food mimics will also have to confront the elephant in the room: What they make is processed food. There’s no way around it. And if there’s anything the environmental movement has agreed on, it’s the evils of processed foods.
In the last century, the processed big food industry has poured salt, fat, and ingredients of murky origin into people’s homes and diets breeding not only distrust, but also the enduring sense that processed food contains a lie.
NotMilk overtly challenges that notion. They are asking if environmentalists might be willing to get into bed with processed foods, if it means more people eating plants? If it’s healthier? If it’s more transparent? These are questions that NotCo appears to be grappling with in its packaging, laced as it is with negation. By exposing the lie of processed foods—by pointing out of their own accord that what they have made is false—they seem to be banking on transparency to overcome the skepticism of the consumer. Yes, it’s not milk, they seem to be saying. But if it’s better, in the bigger picture, would you take a sip?
Veronique Greenwood is a science writer and essayist. Her work has appeared in The New York Times Magazine, Smithsonian, Discover, Aeon and other publications.
Top Image: ©Anthropocene Magazine, AI generated using DALL-E
New study finds that an optimal arrangement of solar panels on farms can cool the panels down by 10 degrees—crucial for their efficiency.
Researchers developed a clever algorithm to coordinate whole neighborhoods of air conditioners, water heaters, and heat pumps—all without comprising individual use.
Wind parks could benefit the natural world—in ways beyond the generation of zero-carbon energy.