WFD vs TasteBot: recipes nobody has ever cooked

Disclosure first: we make What's For Dinner?! (WFD). TasteBot is an AI recipe generator: type a craving or a list of ingredients, and a language model writes you a recipe with macros in seconds. It launched around the end of 2025 and publishes its generated recipes on a public website, which means anyone can read them and check the numbers.

So in August 2026, we did. What follows is what we found on their own published pages, which you can verify yourself. The numbers are checkable facts; the conclusions we draw from them are our opinion.

A recipe is a tested procedure

Start with their pan seared venison medallions with bone marrow butter. The published nutrition data claims 111.9 grams of protein and 111.9 grams of fat. Identical, to one decimal place. That is not how food works; it is how text generation works. The protein figure is also more than the listed meat physically contains, and the calorie field carries the whole recipe's total in the slot that is defined as per serving.

That per serving error turns out to be systematic. Their honey cinnamon apple nachos, sliced apples with a tablespoon each of honey, walnuts and oats per person, carry 510 calories in the structured data field that is defined as per serving, while the visible page itself says 255. The machine readable version of the recipe, the one other apps and search engines consume, disagrees with the page it sits on by a factor of two.

Their butter chicken lists one and a half teaspoons of salt, once. The method then salts three separate steps, the marinade, the rice and the sauce, with no split given anywhere. The cook is left to invent the quantities, which is to say, to do the recipe development themselves.

Then the photos. Every recipe image we checked is exactly 1024 by 1024 pixels, the default output size of image generators. On the venison dish, the recipe says to drizzle the butter warm and melted; the image shows solid, faceted, wax like dollops of butter sitting unmelted on hot seared meat. The "crispy pancetta" looks like fried wonton strips. To our eye, it is a picture of a dinner that has never existed.

One more detail, from their homepage rather than their recipes. Above a counter that read 400 active users when we checked sits the line "Join thousands cooking with TasteBot."

Why this is structural

None of this reads like carelessness by one bad writer. It is exactly what you would expect when recipes are generated at scale, faster than anyone could cook them. A recipe is a tested procedure: someone made the dish, tasted it, fixed the salt, weighed what was left, and wrote down what actually worked. Generation produces something that reads like the output of that process without the process ever happening. The errors above are not typos, and in our view they are very hard to explain if someone had made these dishes before publishing them.

You find out where the fiction breaks with a pan already on the heat and hungry people at the table. Dinner is a bad place to be a beta tester.

The WFD approach

Every meal in WFD is a real, recognisable dish, and the catalogue is deliberately small enough that a human checks every one before it ships: the ingredients, the quantities, the steps, the shopping list lines they generate. When the numbers cannot be trusted, the meal does not go in. That ceiling on catalogue size is a price we pay on purpose, because a dinner app is used on the same few weeknights forever, and it earns trust one correctly cooked meal at a time.

The point of WFD is also different. TasteBot generates food ideas for one person; WFD gets a household to agree. Everyone swipes on meals on their own phone, a mutual match decides dinner, matches build the weekly plan, and the plan builds one shared shopping list. Free to download and always usable free, with an optional subscription for the full feature set. Built in the UK by a couple who cook these meals at home, for couples and families who are tired of the nightly "what do you want?"

	What's For Dinner?! (WFD)	TasteBot
Where recipes come from	Real dishes, human checked before they ship	Generated by AI on demand
Nutrition data	Not the product; no invented macros	Published figures we checked were wrong
Images	Real dishes	1024 by 1024 generated images that contradict the method
Who it decides for	The household, by matching	One user, by generating
Price	Free, optional subscription	Free download; Premium tiers from weekly up to a $149.99 lifetime

If someone asks which app to trust with what a family actually eats this week, the honest answer is the one where a person has cooked the food. Here is how WFD's matching works. And if TasteBot fixes the pages above, let us know and we will update this post.