We Made Recipes Created by Our AI-Generated Food Blog – This is What Happened

Big Human created a startup using ChatGPT and our own auto-blogging technology, Unhuman. Over the next few months, we’re tracking its activity, our findings, and what we’ve learned. This is Part 2 in the series. Read Part 1 here. 

Anyone who has ever found cooking inspiration on the internet — whether from food blogs or digital publications — is familiar with the typical format of recipes. For starters, there’s normally a long narrative at the beginning of the article (an SEO must-have), so you have to scroll all the way down to get to the ingredients and instructions. There’s also usually a comment section, which can be a helpful, and occasionally entertaining, read. If you want a good example, take a look at the New York Times’ recipe for Mississippi Roast. 

Meal Missions, our recipe site created using Unhuman, generally follows the same format as a food blog. There’s a brief synopsis followed by the rest of the recipe. Unfortunately, there are no user comments (it is a brand-new site, after all) that provide feedback and additional context. At first glance, nothing seems to be amiss — aside from the slightly off-putting AI-generated images. 

But we decided to dig deeper. With the site now a few weeks old, we knew it was time to actually whip up some of these recipes ourselves. For better or for worse. 

Each recipe was given an overall rating based on three things: 

1. Ingredients. Did the recipe mention everything needed? 

2. Directions. Were the steps clear and easy to follow?

3. Taste. Would we eat it again?

Without further ado, here’s what happened when we put AI-generated recipes from Meal Missions to the test. 

Recipe 1: Scrumptious Cast Iron Skillet French Churros

User rating: 2/10

It was clear right off the bat that this one wasn’t going to fare well. The introduction to the recipe mentioned “[forming] the churros” in the cast iron skillet, yet there was no mention of piping them (like you’d see in a typical churros recipe, or how they’re presented in the recipe’s photos). That said, we made the executive decision to cut the recipe in half to avoid food waste. 

Pros: They didn’t look pretty, but the churros were edible, especially with an extra coat of cinnamon sugar. 

Cons: The ingredients list didn’t mention oil — so right before we started making the churros, there was a quick run to the nearby grocery store to grab a bottle of vegetable oil. 

Final thoughts: We’re not quite sure what a “French churro” is or what it’s supposed to look like. No plans to repeat this recipe. 

Recipe 2: Authentic Greek Salad with Feta and Olives

User rating: 5/10

The Greek salad put us in more familiar territory. Chopping up the vegetables (lettuce, tomatoes, cucumber, green bell pepper, red onions, and olives) was relatively straightforward — but it was the dressing and its lack of flavor that brought the score down. 

Pros: It felt like a very basic Greek salad. Feta cheese makes everything better. 

Cons: The dressing was too bland, and there wasn’t enough to sufficiently cover the entire salad. 

Final thoughts: “Authentic” is a tough word to describe a recipe created by ChatGPT. 

Recipe 3: Scrumptious Japanese Chicken Soup

User rating: 7/10

The best of the bunch, this Japanese chicken soup was a healthy, albeit a bit boring, dinner for two. Unfortunately, we had to sub button mushrooms for shitake mushrooms (due to grocery store availability), but overall, it was a decent meal. 

Pros: It was well balanced with protein and vegetables. 

Cons: We thought the miso paste should’ve gone in at the very end of the process (instead of the beginning), as we’d experienced when making other dishes.  

Final thoughts: Would we make it again? Maybe. But we’d definitely add some noodles next time.