On Food and Generative AI, Part 1
How might Generative AI impact the food world? This is the first of a three part series exploring the many dimensions of this topic.
This series of essays focuses on the food industry and Generative AI, a subset of the broader category of AI that is designed to create new text, image, code, audio, or video content. By now, you’ve undoubtedly at least heard of apps like ChatGPT, Bard, Dall-E, Midjourney, and many others that have all received much attention in the media. The majority of media coverage has pontificated on how Generative AI might disrupt many industries from software engineering to art and everything in between, but less focus has been paid on how it might impact the food industry.
I will say up front that I’m undecided on how much Generative AI will actually impact the food industry and I have more questions than answers at this moment. I’m deathly allergic to anyone who blindly says a new technology will “definitely change the world,” without any real evidence, but am also careful to not completely dismiss something because no one has thought of the best use for it yet. Just because we don’t have all the answers doesn’t mean we shouldn’t be asking the questions.
It’s honestly too soon to tell with food and Generative AI and the pace of change in AI is incredibly fast, even by the already fast standards of the tech industry. These essays may seem painfully dated in a year but we still need to have the conversation based on what we know right now. Regardless, I’m pretty confident that Generative AI will have some profound impact on our daily lives, and since it’s impossible to separate food from life, the potential impact of Generative AI on food deserves some deeper thinking.
Having worked in innovation and food for nearly my entire career since the early 2000s, I’ve seen a few technologies emerge and truly change the world forever, and even more that have claimed to be the next big thing only to fizzle out months or years later like Juicero, Google Wave, The Segway, Google Glass, and maybe even NFTs.
But Generative AI feels more fundamentally game-changing, like the smartphone, where from day one it was intuitively obvious how it could bring a lot of value to our lives. Generative AI a force multiplier for things we already do everyday—looking things up, writing things, making images—not an innovation that begs us to adapt our behavior to leverage its strengths. It’s not a technology solution in search of a problem and for the most part, it already works shockingly well.
This series is divided into three parts. In this first essay below I’ll be thinking about how Generative AI might change the way we interact with food information through the Internet. It’s become so mundane to simply Google a recipe, share a food pic on Instagram, or look up a menu for a restaurant that we can’t imagine what our food lives were like before the Internet, especially if you were born after the late 90s.
In part 2 next week, I’ll cover the process of how Generative AIs “learn” and how that process can have a profound impact on how they work. I’ll also explore the ways in which Generative AI might augment or replace human skills in the food industry and what it means to work in food in a world where Generative AI exists. Part 3 the week after that will highlight the potential risks and ethical challenges that Generative AI forces us to consider as food industry professionals and how we might stay safe. There will undoubtedly be more to think about after these pieces are published, so expect to see much more going forward about Generative AI in this publication.
I’ll wrap up this intro with a quote from Buckminster Fuller that I think about a lot. It encourages us to be active participants and builders of the future, not passive spectators. If there was any subject in food where this quote definitely applies, it’s this one.
“We are called to be architects of the future, not its victims.”
R. Buckminster Fuller
A New Way to Interact With Food on the Internet
The Internet forever changed our relationship with food. 25 years ago we didn’t have an endless library of recipes, books, blogs, photos, and videos about food at our fingertips. We didn’t have digital social networks to share our food conquests and connect with likeminded eaters. There was no food porn. No online food delivery. No food influencers. Our worldview on food expanded exponentially and the Internet made us more knowledgeable, worldly, discerning, and neurotic about food all at once.
Today we’re a the early stages of yet another revolution driven by generative AI that may again change our relationship with food, and the world in general, for the next 25 years and beyond.
Google searches have become an ingrained part of our lives. You type something like “best paella recipe” into Google and wade through the first page of links until you find something promising. You might have to click on a bunch of links until you settle on a recipe.
Today’s search experience isn’t going away anytime soon, but is already being augmented by generative AI chatbots like ChatGPT and Google’s Bard. These chatbots operate fundamentally different than search engines in a few big ways. For starters, they present you with one answer at a time instead of giving you tons of links where your answer may be hiding. It’s more conversational than traditional search and if you don’t get the answer you want on the first pass, you ask follow up questions until you do, as if you were talking to a human.
Chatbot conversations can be maddening or liberating, depending on your style, but it certainly feels like a thing that is only going to grow in the coming years. There is even a new kind of job called a Prompt Engineer who is sort of a chatbot whisperer and is able to write instructions to Generative AI bots that result in high quality responses (and can make up to $335,000 in annual salary).
ChatGPT was designed to understand more conversational language like, “give me a great seafood paella recipe” instead of keyword salads like “best paella spain authentic valencia seafood” that we might type into Google. With a traditional page of search results, you’re looking for a pre-existing webpage that best suits you needs, and there might not be a webpage that perfectly fits what you’re looking for.
One of the biggest ways generative AI will change our relationship with information is the fact that these chatbots can synthesize and remix information on the fly. Google search will show you everything that’s been pre-written about a topic on the Internet, but it’s not going to rewrite a webpage to better answer your question. Rather than retrieving pre-written information, a chatbot is rephrasing what it has “learned” about a subject after “reading” enormous amounts of Internet content.
The difference between how Google and ChatGPT work is like the difference between asking someone to read you an article verbatim (Google) versus asking them to interpret that article and tell you in their own words what they learned (ChatGPT). So instead of simply having rote knowledge about how to make mayonnaise, for example, ChatGPT can apply what it knows about what mayonnaise is to create an original recipe for spicy mayo using whatever spicy ingredient you have on hand, or any other novel mayo variation that may not have a pre-existing online. What may have required wading through dozens of webpages to find the right kind of spicy mayo you want to make can now be done with a single question to an AI chatbot.
This ability to create something new is what makes generative AI so powerful. This can help many professions close the gap between idea and execution. Need to create code for an app idea? Need to draft a contract but can’t afford a lawyer? Need to write copy for a website? Generative AI can do all of those things. And while I don’t think generative AI is going to completely replace human coders, lawyers and copywriters, each of those industries and many others are currently wrestling with what it means when an AI can do a big portion of their work with increasingly high quality. Generative AI seeks to help you solve problems, not just find information.
Do Chatbots Know How to Cook?
Yes and no. They can kind of write recipes but obviously can’t physically manipulate food yet without the help of some kind of robotics system. I’ll cover more about robotics and AI in the kitchen in part 2, but for now will focus on recipes.
If you want to write new recipes for things that are similar to recipes already well-documented on the Internet, like chicken noodle soup or beef stew, ChatGPT can do that pretty reliably. It can also help you make ingredient swaps, adaptations for dietary restrictions or preferences, and scale ingredient amounts up or down to adjust the serving quantity.
I asked ChatGPT, “can you recommend something Italian for dinner tonight?” and it gave me a classic Spaghetti Carbonara recipe that seemed fine. When I asked it to re-write the recipe to be Keto-friendly (disclaimer: I’m not on Keto), it replaced the noodles with spaghetti squash. It was able to adapt a brownie recipe after I mentioned that I was out of butter and it rattled off 6 fat substitution options that included coconut oil, applesauce, and mashed bananas with recommended substitution amounts for each. It was even able to give me recipes for less practical, esoteric food techniques like spherification and could adapt the recipe for making a chocolate sphere into one for an orange juice sphere, an edit that’s more complicated than simply swapping chocolate for juice.
I even tried to throw ChatGPT a curveball and asked it “how would I make a traditional gumbo in the style of Sichuan hot pot?” It gave me a standard gumbo recipe with the addition of Sichuan peppercorns, dried red chili peppers, doubanjiang and Chinese five-spice powder to the broth. It wasn’t a terribly thoughtful way to make a fusion dish but was a valiant effort considering a Google search for “Sichuan hot pot gumbo” only gave me gumbo recipes or hot pot recipes, but not a combination of the two. Again, this is an esoteric query but it illustrates the potential of ChatGPT to think outside of the box if you so desire.
While these recipes may look promising on the screen, they don’t always work out in the kitchen. New York Times food writer Priya Krishna and tech reporter Cade Metz asked ChatGPT to design a Thanksgiving menu based on some of Ms. Krishna’s biographical details. She told ChatGPT that she was Indian American, grew up in Texas, liked spicy food, and desserts that weren’t too sweet. It returned a menu and recipes for “pumpkin spice chaat, green beans with miso and sesame seeds, naan stuffing, roasted turkey with a soy-ginger glaze, cranberry sauce that’s not too sweet and a little spiced (yes, that’s the full recipe name) and pumpkin spice cake with orange cream cheese frosting.”
While these foods, like the ones ChatGPT gave me, sound reasonable and even interesting, the taste did not deliver. “The roast turkey recipe called for a single garlic clove to season a 12-pound bird, and no butter or oil; the result was dry and flavorless,” and “the chaat, laced with cilantro and baking spices, was a grassy-flavored mush” said Ms. Krishna. After tasting everything, Ms. Krishna’s food writer colleague Yewande Komolafe said, “I don’t feel anything eating this food,” while food writer Genevieve Ko dismissed it as having “no soul behind it.”
We Are At The Beginning of Generative Food AI
I don’t think ChatGPT or any of the other prominent AI chatbots have yet been able to receive a non-standard prompt, like “give me a recipe for Sichuan hot pot gumbo,” and give you a great recipe. But I also don’t think many humans can successfully write a recipe for something that off-the-beaten-path on the first try either. It’s not a normal kind of prompt for most cooks, but as people are all trying more to explore their creativity in the kitchen, Sichuan gumbo-type prompts might not be that weird anymore.
Of course recipes from ChatGPT don’t feel like they have a soul behind them, because computers are just electrified rocks that we tricked into thinking. Engineers have yet to figure out how to program “soul” into these AIs, and until they can create a full digital representation of human food emotions, we’re probably not going to get a truly soulful recipe from a chatbot.
This is not to say that chatbots aren’t helpful in the kitchen. They can certainly help cooks who typically follow recipes exactly as written to make adjustments for ingredient substitutions or other preferences that might not be obvious to some cooks. They can also be a creative springboard for people with more cooking confidence who use recipes as a loose framework rather than a literal blueprint. Experienced cooks can use AI chatbots as a brainstorming partner to spark their imagination then rely on their own human cooking skills to ensure things come out tasty.
ChatGPT was built to be a generalized chatbot that is knowledgeable in a vast range of subjects. It’s a jack of all trades, master of none that might one day become master of everything. The same chatbot that gave me that clumsy Sichuan gumbo fusion recipe can also pass the bar exam near the 90th percentile, write an article about virtually topic in the style of almost any well-known author, and generate working computer code. Yes, it still “hallucinates” occasionally, giving gibberish answers to simple questions and is surprisingly unreliable for a computer when asked to solve straightforward math problems.
But we have to remember that ChatGPT was only released publicly in November 2022 and we are only at the beginning of the generative AI development curve. Being too critical of it right now because it can’t perfectly respond to a culinary curveball is like getting angry at your toddler because she can’t roast a pork shoulder yet.
I’m reluctant to write-off ChatGPT and other modern AI chatbots as being useless for cooks too soon. For all its flaws, it does a pretty good job manipulating recipes that are already well documented on the Internet, which covers a lot of culinary ground. It’s much harder to build a generalized intelligence program that can handle any question you throw at it than it is to build a specialized one.
Google’s DeepMind created AIs that were specifically designed to quickly learn and master games like Go and Chess and beat the strongest human players alive today. The presence of these superhuman AIs in Chess has even made human players better. Those AIs don’t do anything else besides play those games, but specializing their focus enabled them to become historically remarkable players. If a specialized effort to train an AI on cooking resulted in success anywhere near the level of DeepMind’s Go and Chess playing AIs, we could have a truly revolutionary culinary tool that could make better cooks out of all of us (researchers at MIT are working on this).
I am patient and cautiously optimistic for the future of generative food AI, but careful to not get drunk on the generative AI Kool-Aid. In the meantime, the more we test the limits and experiment with these chatbots in the food domain, the more we can identify and shore up their weaknesses and find things that it can do to augment the human food experience. One only becomes a good cook after many attempts, successes, and failures in the kitchen and the same goes for a cook that happens to be an AI.
Stay tuned for part 2, coming the week of May 1, on how AI’s learn and how they may or may not augment or replace humans in the food industry. Click here to subscribe.
Footnotes
3 Recent posts from my Substack
3 Highlights from my current Generative AI reading list
This Changes Everything by Ezra Klein - The New York Times
The “Manhattan Project” Theory of Generative AI by Gideon Lichfield - WIRED
There is No A.I. by Jaron Lanier - The New Yorker
My email is mike@thefuturemarket.com for questions, comments, consulting, or speaking inquiries.