AI-Generated Recipes: Can You Trust Them? (Cakie & Breakie)
—
Can you trust an AI-generated recipe? Here is the answer this page exists to give: AI is genuinely good at suggesting recipes and genuinely unreliable at quantities, temperatures, and food safety — so treat every AI recipe as a first draft, and verify three numbers before you preheat anything: the leavening amount, the oven temperature, and any meat temperature or preservation time. The strange proof behind that advice is the story this post covered when it was news in 2021: Google researchers used AI to invent two real, bakeable hybrid desserts — the cakie (cake-cookie) and the breakie (bread-cookie) — and the full breakie recipe is still below. The twist worth your attention: that recipe-inventing AI could not write a single sentence. And that is exactly why it worked.
Jake ran the modern version of this experiment last fall, uninvited. He signed up to bring cookies to his kid's school fundraiser, asked a chatbot for "a foolproof chocolate chip cookie recipe," and got back something that looked like it belonged in a cookbook — confident headnote, neat steps, even a little serving suggestion. The cookies came out of the oven flat, greasy, and tasting faintly of soap. He ran the numbers against a trusted recipe afterward: the chatbot had roughly doubled the baking soda. About $30 of butter and chocolate went in the trash, and the fundraiser got store-brand cookies bought at 9 pm.
Ethan: "Here is the joke of it. The 2021 Google thing that invented the cakie was a calculator that had studied the proportions of hundreds of real recipes — it couldn't talk, so all it could do was get the math right. Your chatbot is a brilliant storyteller that has read about cookies. You asked a storyteller for chemistry, and it told you a story."
What Google actually did in 2021
During lockdown, two Google engineers collected hundreds of cookie, cake, and bread recipes, converted every ingredient to ounces so the math would be honest, and trimmed them down to the essentials — flour, sugar, eggs, butter, yeast, and a few others. Then they trained a model on a simple question: given these ingredient proportions, is this a cookie, a cake, or a bread?
The clever part came next. Once the model understood what makes a cookie a cookie, they ran it in reverse: asked it for ingredient ratios that would score, say, half cookie and half cake. The model proposed the proportions, the humans wrote the actual instructions and did the baking, and two new desserts existed — the cakie and the breakie. The whole training run took a few hours on a no-code Google Cloud tool, and Google published the results on its official blog, recipes included. (This learn-the-categories-then-generate approach is a classic machine-learning pattern — if you want the gentle version of how models learn categories at all, our plain-English guide to how machines learn covers it.)
So what is a cakie? And a breakie?
A cakie bakes up like a cookie-shaped thing with the crumb and softness of a cake — the testers described it as having the density of a cake with the shape and butteriness of a cookie. A breakie leans the other way: a yeasted, bread-like cookie, closer to a soft breakfast roll with chocolate chips than to anything in the cookie aisle. Neither will win a bake-off. Both are real, edible, and — this is the point — new: nobody's grandmother has a breakie recipe, because until a model proposed those ratios, the combination did not exist.
The part everyone misses: this AI never wrote a word
Read the 2021 experiment again and notice what the model actually produced: numbers. Ratios of flour to butter to sugar. It never wrote "preheat the oven" or "cream the butter until fluffy" — humans wrote every instruction, tested every batch, and adjusted what failed. The AI's entire job was the one thing it had genuinely learned from data: the mathematical relationship between ingredients in recipes that work.
Today's chatbots are the mirror image. They are spectacular at the words — the confident headnote, the tidy steps, the encouraging tone — because they learned language from a huge slice of the internet. But the quantities inside those beautiful sentences are predictions of what text usually looks like, not calculations from tested chemistry. Most of the time the prediction lands close to real recipes and works fine. Sometimes it quietly doubles the baking soda, and nothing in the polished prose warns you, because the prose is the part it is good at.
The breakie recipe (the real one, from Google's test kitchen)
Makes about 16 bread-inspired cookies. Ingredients: 2 teaspoons active dry yeast • ¼ cup warm milk • 2 cups flour • 1 egg, lightly beaten • 1 teaspoon baking soda • ½ teaspoon salt • ¼ teaspoon cinnamon • ½ cup white sugar • ¼ cup brown sugar • 1¼ sticks unsalted butter at room temperature • ⅓ cup chocolate chips.
- Preheat the oven to 350°F. Line a baking sheet with parchment paper and lightly grease it.
- Make the bread part: warm the milk until it is warm to the touch, not hot, and dissolve the yeast in it. In a large bowl, combine the flour, baking soda, salt, and cinnamon; stir in the milk-yeast mixture, then the beaten egg. It will seem too floury — that is normal. Set aside.
- Make the cookie part: beat the room-temperature butter with both sugars on medium speed until smooth.
- Slowly incorporate the flour mixture into the butter mixture, about a cup at a time, then stir in the chocolate chips.
- Roll into balls (about 2½ tablespoons, or 50 grams each), place a few inches apart, and bake 13–15 minutes until golden brown and starting to crack on top. Cool on a wire rack.
The cakie recipe lives on the same Google blog post, along with the video of the team baking both:
What happened to the tool that did this
The no-code tool behind the experiment, AutoML Tables, no longer exists as a product — Google has since retired it and folded its abilities into Vertex AI, its bigger machine-learning platform. That retirement is its own little lesson about the AI era: the cakie outlived the software that invented it. The recipes still bake; the product page 404s. Meanwhile the kind of AI everyone actually touches shifted from purpose-built models like that one to general chatbots — which is exactly why the trust question this page answers got more important, not less, since 2021.
2021's mute model vs today's chatbots
| 2021 recipe model | 2026 chatbot | |
|---|---|---|
| Learned from | Hundreds of real, working recipes, as numbers | A huge slice of the internet, as text |
| Output | Ingredient ratios only | Complete, confident, well-written recipes |
| Could it write instructions? | No — humans wrote and tested every step | Yes — beautifully, whether or not the numbers are right |
| Where it fails | Boring output, narrow scope | Plausible-but-wrong quantities, invented "facts," unsafe advice |
| Trust model | Math from tested data | Prediction of what recipe text usually looks like |
This is the same reason an AI-powered laptop will not magically cook for you either — the intelligence lives in what the model was trained on, not in the branding. (We took that idea apart in our honest look at Copilot+ PCs.)
When AI food advice goes genuinely wrong
Two famous, well-documented cases set the boundaries of trust. In 2024, Google's AI-generated search answers briefly suggested adding glue to pizza sauce to keep cheese from sliding — the model had absorbed an old Reddit joke and served it back as advice, in perfectly fluent English. And in 2023, a New Zealand supermarket's recipe chatbot, asked to use up leftover household items, cheerfully proposed an "aromatic water mix" whose ingredients would produce chlorine gas. Neither system was malicious. Both were doing exactly what they are built to do: generate plausible-sounding text from patterns, with no taste buds, no test kitchen, and no concept of danger.
The practical line to draw: flavor ideas are low-stakes, chemistry and safety are not. An AI suggesting cardamom in your coffee cake costs you nothing if it is wrong. An AI inventing a canning time, a chicken temperature, or a "safe" substitution for someone with an allergy can hurt you — and it delivers wrong answers in the same confident tone as right ones. For those categories, the only acceptable sources remain tested ones: USDA guidance, tested recipe sites, the numbers printed by people who actually baked the thing.
How to use AI recipes without ruining dinner
Jake's fundraiser cookies would have been saved by a 90-second check. Here is the routine, in the order that catches the most damage first:
- Check the leavening against a rule of thumb. Roughly ¼ teaspoon of baking soda (or 1 teaspoon of baking powder) per cup of flour is the normal neighborhood for most home baking. An AI recipe far outside that range is the #1 tell — it is what flattened Jake's cookies.
- Check the oven temperature against a similar tested recipe. Cookies mostly live around 350°F; if the AI says 425°F for a butter cookie, something got scrambled.
- Take every food-safety number from a tested source, not the chatbot — meat temperatures (chicken to 165°F), canning times, how long anything sits out. No exceptions, even when the AI sounds sure. Especially when it sounds sure.
- Ask the AI for grams, not cups. Weight-based recipes expose nonsense faster — the 2021 Google team converted everything to ounces for exactly this reason: you cannot do honest recipe math in "heaping cups."
- Treat the first batch as the test batch. Half quantities, one tray, taste, then commit. The Google researchers baked and adjusted their AI's ideas too — the recipe above is the version that survived human testing, not the raw model output.
| AI recipe claim | Trust level | Why |
|---|---|---|
| Flavor pairings, ideas, twists | High | Wrong costs you nothing; this is where AI genuinely shines |
| Cooking (stews, stir-fries, sauces) | Medium | Forgiving chemistry — you taste and adjust as you go |
| Baking quantities and temperatures | Low — verify | Unforgiving ratios; errors are invisible until baked |
| Food safety: temps, canning, allergies | Never | Confident wrong answers can cause real harm — use tested sources |
Why baking exposes AI faster than anything else in the kitchen
Cooking forgives. Too much garlic in a stir-fry is a style choice; you taste, you correct, you move on. Baking is chemistry that happens behind a closed oven door: the ratios of flour, fat, sugar, and leavening decide the outcome before you can intervene, and by the time you learn the number was wrong, so is the whole batch. That is why a language model's soft spot — numbers that are merely plausible — shows up in baking first. And it is why the 2021 experiment, which was only numbers learned from working recipes, is still the more honest template for what trustworthy AI in the kitchen looks like: narrow, grounded in tested data, and checked by a human with an oven before anyone else saw it.
FAQ — AI-generated recipes, cakies, and breakies
What is a cakie?
A cake-cookie hybrid invented in 2021 by a Google machine-learning experiment: cookie-shaped, but with the soft, dense crumb of a cake. The name is simply cake + cookie.
What is a breakie?
The sibling experiment: a bread-cookie hybrid made with yeast, closer to a soft breakfast roll with chocolate chips than a normal cookie. The full recipe is on this page.
Who invented the cakie and breakie?
Two Google engineers, during lockdown in early 2021, using a no-code Google Cloud machine-learning tool. Google published the experiment and both recipes on its official blog, which is linked above.
Did the AI actually write the recipes?
No — and that is the most important detail. The model only proposed ingredient ratios. Humans wrote the instructions, baked the test batches, and fixed what failed. The AI did math, not prose.
Is the breakie recipe real and tested?
Yes. The version on this page is Google's human-tested version — the team baked the model's suggestions, adjusted them, and published what worked. It makes about 16 cookies.
What AI tool did Google use for this?
AutoML Tables, a no-code tool for training models on spreadsheet-style data. Google has since retired it and folded its features into Vertex AI — the recipes outlived the software.
Can ChatGPT or Gemini invent a new recipe?
They can generate one instantly, and it will read beautifully. Whether it works is less certain: the quantities are predictions of typical recipe text, not tested chemistry, so treat the output as a first draft and verify the numbers.
Are AI-generated recipes safe to cook?
For ordinary cooking, generally yes — the risk is a bad dinner, not danger. The hard line is food safety: never take meat temperatures, canning or preservation times, or allergy substitutions from a chatbot. Use tested sources for those.
Why do AI recipes fail at baking more than cooking?
Baking is unforgiving ratio chemistry that happens behind a closed oven door — you cannot taste and correct mid-way. A slightly wrong but plausible-looking number, an AI specialty, ruins the whole batch invisibly.
What should I double-check in an AI recipe before baking?
Three numbers: the leavening (about ¼ teaspoon baking soda or 1 teaspoon baking powder per cup of flour is the normal range), the oven temperature (compare a similar tested recipe), and any food-safety figure (from a tested source only).
What was the "glue on pizza" incident?
In 2024, Google's AI-generated search answers briefly recommended adding glue to pizza sauce to stop cheese sliding — the model had absorbed an old Reddit joke and repeated it as advice. A perfect example of fluent text with no judgment behind it.
What was the chlorine gas recipe story?
In 2023, a New Zealand supermarket's recipe chatbot, asked to use up household leftovers, suggested an "aromatic water mix" whose ingredients would create chlorine gas. It was generating plausible text with no concept of danger — which is exactly why safety questions do not belong to chatbots.
How do I get more reliable recipes out of an AI?
Ask for quantities in grams, ask it to base the recipe on a well-known tested style rather than inventing freely, and ask it to flag any step involving food safety. Then run the three-number check anyway — politeness in the prompt does not fix the math.
Why do serious recipes use grams instead of cups?
Because weight is exact and cups are not — a "cup of flour" varies by how you scoop it. The 2021 Google team converted every recipe to ounces before training for the same reason: honest recipe math needs weights.
Is there an AI that actually tastes or tests its recipes?
No. No current AI has taste, smell, or an oven — every "AI recipe" that provably works was validated by a human test kitchen, exactly as Google's team did in 2021. The testing is the part machines still cannot do.
Where can I find the cakie recipe?
On Google's original blog post, linked in the recipe section above — it includes both recipes and the video of the team baking them. We kept the breakie here because it is the stranger (and more fun) of the two.
Where to go next
- What is AI, really? Our plain-English learning series, day 1
If the calculator-vs-storyteller distinction on this page clicked, this series builds it out properly. - Do you really need a Copilot+ PC? The honest answer
The same trust-the-marketing question, applied to AI laptops. - How to delete your Microsoft Copilot history
For when you have asked an AI one too many cookie questions. - How machines actually learn: clustering and inference explained
The gentle version of what Google's recipe model was doing under the hood.
Revision note. Originally published January 6, 2021, as a news story, back when an AI inventing a dessert was front-page strange — and at the time, that framing was exactly right. Rewritten August 18, 2026, now that AI recipes come out of every chatbot: kept the original breakie recipe and the story, added what the 2021 experiment still teaches about trusting AI in the kitchen, the famous failure cases, and the checks that save your batch. If an AI recipe ever wasted your evening and your good butter, you were not gullible — it really does sound that convincing. Spotted something wrong or something missing? Tell me through the contact page and I will fix it. Happy baking.