Lasagna look too good to be true? Put down the mozzarella | Canada Voices

In late December, Sarah Kim posted a recipe for her creamy Mediterranean vegetable lasagna on her blog. In a lengthy headnote, she said she was inspired by the bounty she brought home from the farmer’s market and hearing her brother gush about what he’d eaten on a recent trip to Athens.

A photo of the finished dish shows layers of zucchini, eggplant and red and yellow peppers nestled on a bed of oozing mozzarella cheese, crowned with browned Parmesan and sprigs of fresh thyme. It’s perfect. Too perfect.

The picture was AI-generated. As was the personal story about the dish. Sarah Kim, the curly haired brunette whose headshot is at the top of the page, is not made of mortal flesh. She did not roast vegetables, stir marinara and mix ricotta “while the kids were piling puzzles on the living room floor,” nor did her husband sneak “a few cheese threads,” as she wrote. Kim is made of pixels spat out by an image generator.

In just a few years, the internet has become overrun with AI food content – from recipes to photos to instructional videos – that were created without a single egg cracked, stove element turned on or dish dirtied. This has created trust issues with users who are simply trying to figure out what to cook for dinner, and threatened the livelihood of food content creators – the ones who are actually human.

The popular food blog Budget Bytes, which launched in 2009, has grown from a one-person show to a much more sophisticated operation with five staff, nine contributors and a test kitchen in Nashville. The standards are high: Each recipe is tested three times before being published. On photo-shoot day for a lasagna, the team will bake three different trays at staggered times to get the perfect shot of a gooey cheese pull.

The site gets tens of millions of visitors each month. Traffic peaked in 2024, but last year it dropped by 8 per cent and organic search declined by 20 per cent. One of the driving factors was the impact of AI overviews – the AI-generated summaries that appear at the top of a page of search engine results. They reduce click-through rates, which kill blogger ad revenue, says Jess Rice, Budget Bytes’s recipe development manager.

About half of Google searches have AI summaries, and that’s expected to increase to more than 75 per cent by 2028, according to an October report from McKinsey & Company.

Much of the social-media traffic Budget Bytes receives comes from Pinterest, but those visits have dropped by 40 per cent from their peak in early 2023. The team attributes the decline to the surge of AI-generated content on Pinterest.

“AI slop sites will fully copy our content and, in some cases, outrank bloggers by their reach,” Rice said. Her team has also come across photos stolen from Budget Bytes and appended to AI-generated recipes posted to Facebook. They’ve reported these accounts, but it feels like playing a game of Whac-A-Mole.

In response to growing complaints, Pinterest last year introduced labels for AI content on its platform, as well as filters users can turn on to reduce the amount of AI content they’re shown.

Loblaw introducing ChatGPT integration for its grocery delivery app

Still, an analysis in February by The Globe and Mail using AI-detection software found rampant use of AI-generated images on many food posts at the top of Pinterest search results, none of which were labelled as “AI modified.”

When asked whether Pinterest might tweak its algorithm to rank human-generated content higher than that made by AI, a spokesperson for the company said no. Human creators are still a priority, they said, but quality of the content is what determines what rises to the top of the results page.

In an effort to distinguish its recipes from the slop, Budget Bytes now includes a boilerplate line on each one: “All recipes are rigorously tested in our Nashville Test Kitchen to ensure they are easy, affordable, and delicious.” The team responds to comments and questions and frequently updates old recipes to integrate audience feedback in an effort to prove humans are behind the content.

In addition to Pinterest, YouTube and Meta have also updated their algorithms in ways that have helped AI-generated content proliferate, says Ann Reardon, a dietician, pastry chef and food scientist who has posted food content to her YouTube channel, How To Cook That, since 2011.

One factor is that “newness” is rewarded by social-media algorithms. An account exclusively posting AI-generated content can, in a single day, churn out the amount of recipes it would take a human weeks, if not months, to create, Reardon says.

But more content means more views, which means more ad revenue for the platform. Same with more engagement. So it’s not uncommon that an error-ridden AI recipe with thousands of comments and shares will be pushed into a Facebook feed over one that a team in a test kitchen spent several days developing and photographing.

If online platforms don’t change their algorithms, “I think that it will push out creators who have been doing it for years,” Reardon said.

And that’s an issue for both creators and home cooks, she warns. AI is not good at math, Reardon points out. She’s come across plenty of AI-generated recipes where the ingredients are right but the quantities are off. Listing 1 tablespoon of chili powder instead of 1 teaspoon could render a dish inedible.

For Deb Perelman, the pioneering blogger behind the 20-year-old site Smitten Kitchen, her audience’s trust is everything. While she worries about the risk to her own livelihood, she’s more concerned about the impact AI recipes will have on people who are looking for an easy family meal for a harried weeknight, or planning to make a birthday cake for a loved one, or simply learning the basics of cooking.

“To take a risk with somebody’s time, money or energy is so worrisome to me,” she said. “It’s a violation.”

For a while, she dismissed the threat of AI, certain that the quality of the content would be easy to distinguish from the work of human creators. But then she asked ChatGPT to create a Smitten Kitchen-style recipe headnote.

Cookbook author Carla Lalli Music gets brutally honest about recipes

“It was so good. I spiralled out. I felt like someone had taken the floor out of my life and it took me a few days to recover,” she said.

A friend pointed out she’d given the program 20 years’ worth of posts to study the way she spoke – of course it was doing an eerily convincing imitation of her voice.

Unlike other bloggers who have moved to a paid newsletter format, Perelman – who has published three cookbooks – is reluctant to put her recipes behind a paywall. She’s optimistic she can ride out the storm, the same way she’s weathered many others in the past two decades. The proliferation of AI, she says, has only added value to her content.

“I’m hopeful that it will return interest to things that are actually made by human beings. We are crying out for actual human connection and things made by real people.”

That was the conclusion home cook Lin Bryant reached after being burned by a recipe she suspects was AI-generated.

Last fall, Bryant, a seasoned baker, did a Google search for cranberry-orange cupcakes. She landed on a recipe with an enticing photo: a yellow cupcake topped with a swirled mound of blush pink frosting flecked with red bits of cranberry and topped with sugared cranberries.

As she read the recipe, the inclusion of raw cranberries gave her pause; she knew cranberries are so tart they usually need to be cooked down with sugar. But the five-star rating and plenty of rave reviews persuaded her to charge forth.

The frosting came out a “disgusting” cross between salad and icing: dense with inedible tart chunks of raw cranberry. She chucked the cupcakes in the garbage.

Later, she returned to site and the red flags started popping up. The linked Facebook page shows a suspiciously high volume of recipe posts – far more than a blogger could produce on her own. She wondered whether those sterling reviews were fake.

“No sane person would make these comments,” she said.

Bryant can’t prove the recipe was AI-generated, but the technology’s proliferation has made her wary.

If she’s looking for a recipe, she sticks to the YouTube channels or websites of recipe developers she’s trusted for years. If the only picture she can find of a blogger is the one on the “about me” page of a site, she assumes it’s a bot and closes the browser tab. Ditto if she finds a YouTube video that only shows hands and no faces. Reviews and ratings are no longer a useful tool for determining authenticity.

“I don’t know any of these people, you know?” she said. “I don’t know how good AI is. It’s very disturbing.”

With a report from Patrick Dell

Want to test your AI detection skills? Take this test:

1Which of these images was created using AI?

A

B

a. Image A is a real photograph of tacos. Image B was generated by Microsoft’s Copilot text-to-image AI generator of tacos.

b. Image B was generated by Microsoft’s Copilot text-to-image AI generator of tacos. Image A is a real photograph of tacos.

Image A is a real photograph of tacos. Image B was generated by Microsoft’s Copilot text-to-image AI generator of tacos.

2

A



B

Image A is a real photograph of green curry with tofu and vegetables. Image B was generated by Microsoft’s Copilot text-to-image AI generator of green curry with tofu and vegetables.

3

A



B

Image A was generated by Microsoft’s Copilot text-to-image AI generator of steak, potatoes and green beans. Image B is a real photograph of steak, potatoes and asparagus.

4

A



B

Image A is a real photograph of a turkey BLT sandwich. Image B was generated by Microsoft’s Copilot text-to-image AI generator of a turkey BLT sandwich.

5

A



B

Image A was generated by Microsoft’s Copilot text-to-image AI generator of a green salad with dressing on the side. Image B is a real photograph of green salad with dressing on the side.

How well did you do?

Answer all of the questions to see your result

Congrats! You are – at least for now – able to tell the difference between real life and AI.

Good effort. Looks like AI can do a decent enough impression of a food photographer to fool you sometimes. Stay vigilant!

Yikes. You seem to be perfect prey for the AI bots (which, in your defence, are getting more and more sophisticated).