As It Happens·Q&A

What happens when you let artificial intelligence decide what's for dinner

New York Times food reporter Priya Krishna cooked an American Thanksgiving dinner made entirely from AI-generated recipes. She spoke to As It Happens guest host Helen Mann about how it all turned out.

The recipes looked promising, but NYT's Priya Krishna says the final product failed to deliver

A woman in an apron prepares a turkey in a kitchen while hunching over to look at an open laptop on the counter, her face scrunched in confusion.
New York Times food reporter Priya Krishna tries to make a Thanksgiving turkey based on an AI-generated recipe. (Timothy O'Connell/The New York Times)

Priya Krishna and her colleagues are feeling very confident that artificial intelligence won't steal their jobs.

The New York Times food journalist has cooked an entire American Thanksgiving dinner using AI-generated recipes. 

The concept was the brainchild of New York Times Cooking social media editor Becky Hughes. Krishna used an AI called GPT-3, from the company Open AI, to write the recipes, and DALL-E, Open AI's image generator, to create pictures of what the hypothetical food would look like. 

Then Krishna cooked up the recipes in the New York Times test kitchen, and asked the newspaper's food columnists to review the dishes. The results were published in the Times on Friday.

Krishna spoke to As It Happens guest host Helen Mann about how it all panned out. Here is part of their conversation.

What made you want to play around with AI to see if it could come up with a Thanksgiving menu? 

We thought, you know, artificial intelligence has gotten really smart. Do we think it can make a better Thanksgiving menu than us humans, who sort of spend year after year trying to reinvent the wheel for Thanksgiving? 

So what's the process that you had to go through on your end? How did you program or manipulate this technology to actually create these Thanksgiving recipes? 

Well, we're not programmers, so we didn't have to program anything. We literally just created an account and logged in to GPT-3, which you can do right now, and we just started typing in prompts. 

We met with some scientists from Open AI, the company that runs GPT-3, and they encouraged us to get personal. So I told GPT-3, in the white box where you type your prompts, you know: I'm Indian-American. I grew up in Texas. I like spicy flavours. I like Italian food. I like Thai food. I like these ingredients. Show me a Thanksgiving menu made for me. 

And then you hit Enter, and within seconds you've got a Thanksgiving menu that it populates with recipes. 

So it really is not just throwing out some generic idea of what a turkey should be. It's a turkey that Priya would like, in theory. 

Yes. And all of the recipes it's generating are original. It's not just copying and pasting from the internet. 

WATCH | Cooking an AI-generated Thanksgiving menu:

What … did the AI come up with in terms of this menu? 

They came up with a naan stuffing that had 32 different ingredients. That was wild. A roast turkey with a soy glaze that only called for one clove of garlic for a 12-pound bird. And then a really confusing one called pumpkin spice chaat, which was just pumpkin puree with spices and cilantro and lime juice. 

As you were cooking, what was going through your mind? 

It started really hopeful. You know, first you sauté onions in a pan. You mix ingredients together for a cake batter. You sort of feel like you're going through the correct motions that should lead you to a good end result. 

But then, ultimately, I was tasting, and most of the dishes did not turn out great. 

There's like little flags you notice as you're cooking the recipes. Like, why is there no salt? Why is there no fat in the turkey recipe?

And people who cook a lot, they get used to having a sense of at least what the quantities are, normally, in a recipe. So when you're measuring in flour for the cake or spices for the chaat or whatever, did [the measurements] make sense? 

Some of them did, some of them didn't. For example, the naan stuffing called for two full cups of dried fruit. And two cups of dried fruit for a stuffing recipe that feeds maybe four, just felt really excessive to me. 

Two pictures of cake, side by side. On the left, a white, two-layer cake with white frosting with decorative orange flourishes. On the right, a much flatter and darker two-layered cake.
On the left, an AI-generated image of what a pumpkin spice cake might look like. On the right, the actual cake prepared in the New York Times test kitchen. (Open AI, Timothy O'Connell/The New York Times)

Now, the fun thing is that you didn't just taste this yourself. You brought together four taste testers. And we're talking about some pretty big names. For those of us who love the Times cooking section, who were your tasters? 

Yeah, we brought the heavy hitters and all of our cooking columnists. Genevieve Ko, Yewande Komolafe, Melissa Clark and Eric Kim. So each of them [has] a food column. They are cooking and Thanksgiving experts. And they certainly do not hold back with their feedback. 

So take me through some of the things they tasted. Let's start with the turkey. What did everybody think of the turkey? 

I think Melissa said it was the driest turkey she'd ever tasted. 

Was that because it gave you a wrong cooking time or a weird temperature or something?

It had me cook it until it reached an internal temperature of 180 degrees [F]... Also, we weren't putting any fat on the bird — not rubbing it with butter oil before cooking it. And turkey is not the most naturally juicy meat. It sort of needs some additional fat. So I think, you know, neither of those things really set it up for success. 

What about the naan stuffing with all the fruit?

The naan stuffing was a disaster. I mean, it was just the recipe equivalent of doing too much.... There were like a quarter cup of spices in it. There was lemon zest. I mean, it was all good ingredients, but just too much of it. Like that recipe needed to be pared back 75 per cent. 

What about the [green] beans [with miso and sesame seeds]?

The green beans, people said, were good. They were like: I would eat these. 

But I think it was also after they'd tried six inedible things. So I think the bar was pretty low at that point. 

What about the [pumpkin spice] cake? 

People were very excited about the cake. It sounded promising; it looked decent. But yeah, unfortunately, it had no sugar besides some honey to sweeten it. So it just wasn't sweet at all. The base itself was more bready than cakey. So it just tasted savoury. I feel like the frosting-to-cake ratio was not right. It was just a really disappointing cake. 

It sounds like overall the whole thing was pretty disappointing. Does this make you feel more secure about your job? 

Oh yeah. We all discussed how we feel like we have job security now after trusting all these recipes. 

In going through this, did you see that there might be some potential in the future for AI generating food ideas? 

I definitely think so. I think that we would be naive to think that AI won't continue to play more of a role in our cooking. 

I mean, it already does. If you have an Amazon Alexa and you use it to convert teaspoons to tablespoons, you're using artificial intelligence. If you use Google to search recipes for Thanksgiving, you're using artificial intelligence. 

So I think it stands to reason that maybe we'd start using a system like GPT-3 to help us, you know, mine the thousands and thousands of recipes and online knowledge around cooking that exists and help us answer some questions, even if it's not just explicit recipe generation.

Interview produced by Kate Swoger. Q&A has been edited for length and clarity.

Add some “good” to your morning and evening.

A variety of newsletters you'll love, delivered straight to you.

Sign up now