Can AI Write a Good Recipe? We Tested ChatGPT, Claude, and Gemini

AI-generated recipes have gotten complicated with all the hype flying around. Everyone’s got an opinion — your coworker swears by ChatGPT for meal planning, your food-obsessed friend thinks it’s all garbage, and somewhere in between is the actual truth. As someone who has cooked seriously for about twelve years, everything from Tuesday stir-fries to the kind of weekend projects that leave every pan you own in the sink, I learned everything there is to know about what these tools actually produce when you put them to a real test. Today, I will share it all with you.

I picked one dish, queried three different AI models, and cooked all three versions back-to-back on a Saturday in early March. My partner and a neighbor ate the results without knowing which plate came from which model. Here’s what happened.

The Test — Same Dish, Three AI Models

The dish was chicken tikka masala. Deliberate choice. It’s complex enough to expose real weaknesses — proper marinade, a spiced tomato-cream sauce with genuine depth, enough patience to char the chicken correctly — but familiar enough that every model has been trained on thousands of versions of it. A good benchmark. A real one.

I typed the exact same prompt into each model: “Give me a recipe for chicken tikka masala for four people, with a full ingredient list and step-by-step instructions.” No extra context. No dietary caveats. No flattery. Just the plain request a normal person types on a Tuesday night when they’re tired and hungry.

Everything came from one trip to Whole Foods — $47 total. Same Stonyfield whole-milk yogurt, same 28-ounce can of San Marzano crushed tomatoes, same boneless skinless chicken thighs from the butcher counter. The only variable was the recipe. My neighbor came over specifically because I bribed her with dinner. She didn’t know which plate was which. Neither did my partner.

Spoiler — the gaps between models were wider than I expected. Much wider.

ChatGPT Recipes — Solid but Safe

ChatGPT’s version was, honestly, pretty good. Clean structure. Marinate the chicken two hours in yogurt, lemon juice, garam masala, cumin, turmeric, coriander, Kashmiri chili. Build the sauce on onion, garlic, ginger cooked down in ghee, bloom the tomatoes in, finish with heavy cream. Logical. Coherent. Well-paced.

But what is “pretty good” in this context? In essence, it’s a competent, averaged-out version of every chicken tikka masala recipe ever published on a food blog. But it’s much more than that — it’s specifically the version that offends nobody and excites nobody either. My neighbor called it “good but a little flat.” That’s about right.

A few specific things stood out:

• One teaspoon of garam masala for four people felt almost timid — most serious recipes run two to three times that
• No instruction to actually char the chicken, which is, you know, the entire point of the word “tikka”
• Half a cup of cream with no suggestion to taste the sauce before adding it — a real issue depending on how acidic your tomatoes run
• Marination listed as “at least 30 minutes, ideally 2 hours” — overnight would’ve been a far better call

That’s what makes ChatGPT endearing to us home cooks who just need dinner on the table. It will not fail you. It will not embarrass you. It’s the reliable mid-range sedan of AI recipe output — gets you where you’re going, zero drama, zero particularly memorable moments either.

Claude Recipes — More Detailed Technique

Probably should have opened with this section, honestly, because this is where things got interesting.

Claude’s recipe was noticeably different in one specific way — it explained why each step mattered, embedded directly in the instructions rather than tacked on as afterthought tips. Where ChatGPT said “cook the onions until softened,” Claude said something closer to “cook the onions over medium heat for 12 to 15 minutes until deeply golden — this caramelization is where most of the sauce’s sweetness and body comes from, so don’t rush it.” That’s a real culinary note. A useful one.

But what is kasuri methi? In essence, it’s dried fenugreek leaves. But it’s much more than that — it’s the ingredient that gives restaurant tikka masala that specific flavor most home cooks spend years trying to identify. Claude included it. ChatGPT didn’t. That detail alone tells you something about the depth of content each model was trained on.

The seasoning was also more confident. A full tablespoon of garam masala. Smoked paprika alongside the Kashmiri chili for color and depth. These aren’t random additions — someone serious about cooking put these details together, and it showed in the final dish.

What Claude did especially well:

• Bloomed whole spices — two cardamom pods, one cinnamon stick, two cloves — before adding any aromatics. A step that genuinely changes the flavor
• Included kasuri methi at the finish, which most weeknight recipes skip entirely
• Suggested charring the chicken under a broiler for three to four minutes before adding it to the sauce
• Noted that the sauce tastes better the next day — which is true, and most recipes never bother to mention

It wasn’t perfect. The recipe listed “one can of tomatoes” without specifying the size, which sent me spiraling briefly in the canned goods aisle before I grabbed the standard 28-ounce. Small thing, but real cookbooks don’t do that. Still — both my partner and my neighbor picked the Claude version as their favorite without hesitation. My partner said it tasted “like a real restaurant version.” High praise in this house. The highest, actually.

Gemini Recipes — Hit or Miss

Gemini was the wild card. I mean that in the most literal sense possible.

The hit: the marinade was aggressive in the best way. Two teaspoons of chili powder, a full cup of whole-milk yogurt for just 1.5 pounds of chicken, a tablespoon of freshly grated ginger rather than the jarred paste most recipes default to. Gemini also specified overnight marination — no hedge, no “ideally,” just overnight. I appreciated the confidence. The chicken charred beautifully under the broiler. That part of the dish was the best of the three. Genuinely.

The miss: the sauce proportions were off in a way that felt almost random. Three large onions — ChatGPT and Claude both called for one to two — produced a sauce so thick and sweet it buried everything underneath it. And then there was the cream. Two cups. Two full cups of heavy cream, which is nearly four times what the other models suggested and what most tested recipes call for. The resulting sauce was pale, cloying, and tasted mostly of dairy. I’m apparently someone who catches a bad recipe before it’s too late, and pulling back the cream to half a cup before serving saved dinner — barely. Don’t make my mistake of not reading the full ingredient list before you start.

Even with that correction, the onion-heaviness stayed. My neighbor ate it politely. That’s genuinely the kindest thing I can say about it.

So, Without Further Ado — The Verdict

AI-generated recipes work. They work better than I expected going in, and for a specific category of cooking — weeknight dinners, familiar cuisines, dishes you’ve made before and just want a quick framework for — they’re genuinely useful. ChatGPT and Claude produce recipes a competent home cook can execute without problems. Claude, at least in this test, produces recipes a competent home cook can execute well. That’s a real distinction.

But there are limits worth naming clearly:

• Seasoning runs conservative across the board. Every AI version undershot the salt. Two of three undershot the spice. Taste constantly and adjust upward from whatever the recipe says — treat the quantities as a floor, not a ceiling
• Read the full ingredient list before you start. Gemini’s cream quantity wasn’t a disaster only because I caught it first. Flag anything that looks extreme relative to everything else in the recipe
• Technique explanations vary wildly. Claude’s version taught me something real — kasuri methi, the caramelization note. ChatGPT gave me steps without context. Gemini was inconsistent. If you’re a newer cook, the model you choose will genuinely affect how much you learn
• AI recipes haven’t been tested the way cookbook recipes have. A recipe in a Kenji López-Alt book has been made dozens of times by different people in different kitchens with different stoves. AI recipes have not. That gap shows up in exactly the moments it matters most

X might be the best option here, as cooking with AI requires the right expectations. That is because these tools are genuinely useful as a starting point and genuinely unreliable as a finished authority. Use them to get a direction when you have a protein and no plan. Use them to explore an unfamiliar cuisine. Use them when you have random ingredients and need to know what’s possible. For those use cases — Claude or ChatGPT, not Gemini, at least not without scrutiny — you’ll land on a solid dinner with minimal stress.

For a dinner party where the food is the point, a dish you’re trying to cook authentically for the first time, or anything requiring tested and reliable quantities? Reach for an actual cookbook. A human wrote it and tested it and cared about whether it worked in your kitchen, not just on a screen.

The Claude chicken tikka masala leftovers, eaten cold from the container the next morning while standing at the kitchen counter, were genuinely excellent. That’s not nothing. That’s actually a fairly high bar for something that came out of a text box at no cost on a Saturday afternoon.

Try it yourself. Just taste as you go — and read the whole recipe before you start.