We asked AI to generate a citrus recipe and this is what happened…
Being in the food business, you tend to enjoy cooking food, eating food, trying new flavours, and seeking advice from like-minded folk who know their way around a kitchen, or our case, orchard too. Since it went mainstream, Artificial Intelligence, or AI, has helped many of us mere mortals with random questions. In fact, when I typed these three words, ‘how do I…’ into ChatGPT I was prompted with what I must assume were the top requested questions:
• How do I bake a cake from scratch?
• How do I improve my public speaking skills?
• How do I keep my plants alive?
Un-aliving plants seems to be something I’m really really good at, so the third point did make me think, ‘is this chat bias to my history’? Possibly? But I have definitely baked a few cakes in my time, so possibly not. It got me thinking…
There must be a HEAP of people out there using AI to help them cook.
It’s a fairly reasonable notion. I mean, we all have to eat, and it can get mighty cumbersome thinking of what to make with the sad leftovers in the fridge the day before shopping day.
Baking, cooking, flavour preference and the like are all innately human experiences that require, at a bare minimum, the sense of taste. Given the breadth of recipes and data already uploaded to the internet about cooking, flavours and produce, I’m guessing that AI will have a fairly good handle on what flavours work with each other. My aim was to write a brief for a recipe.
The first question that came to mind was to ask ChatGPT to “write a recipe that is practical, kid friendly, uses citrus, quick to cook and easy to make”. But this was too broad, and I needed to nut out some specifics, like what citrus are we using? What kind of cooking (i.e. baking, frying, boiling)? Is this sweet or savoury? What is the time limit for prep and cooking?
I decided to define some parameters:
Must feed a family of 5
Must be kid friendly – no hot spices or wine added, etc
Must be able to cook in under an hour
Must be healthy
Must be savoury
Must use fresh blood oranges
The first result was Blood Orange and Honey Chicken with Veggie Rice. For me, it was a little ‘Meh’. Not a lot of creativity happening here. Kids and veggie rice? Too obvious… so I asked for another recipe.
The second recipe was Blood Orange And Herb Salmon With Roast Potatoes and Greens. Not too bad flavour-wise but I’m sick of salmon. We eat it at least twice a week, so I asked for another recipe without the salmon.
I received Blood Orange And Herb Chicken Meatballs With Couscous.
Nope. Chicken meatballs? I’m not a psychopath.
I promptly told this to ChatGPT, who obediently amended the recipe to Blood Orange Honey And Pork Meatballs With Veggie Noodles. Pork meatballs with honey and blood orange glaze? Yes! But the noodle part seemed a little like something you’d serve at a daycare centre.
“Ah Ha” I thought. That was my problem. I didn’t assign ages for the children. I updated my prompt for children aged 6, 12 and 14, and it came back with Sticky Blood Orange Pork Meatball Subs. My kids would eat this, but it lost the healthy part. To which I prompted “keep it healthy. Too much bread”
The result back? Blood Orange Pork Meatballs with Rainbow Rice Bowls. We’re back to veggie rice, aka ‘Rainbow Rice’ aka ‘Boring AF Rice’.
My patience with AI was waning. I was not sure what I was doing wrong. Here was my response:
To which AI responded with Blood Orange and Ginger Pork Lettuce Cups. Not creative, but a good idea for kids. The glaze is not actually a glaze. I’m thinking it would be the sauce to go with the mince – like san choy bao – but missing a few essential ingredients like sesame oil, salt, cornflour, or maybe some Chinese five spice at a minimum? I also got AI to generate an image of what this would look like. Can’t say I’m eating with my eyes right now.
What ChatGPT thinks we want to eat (ewwww)
Scrap this. It’s out. “Back to the drawing board”, I thought.
I asked ChatGPT to write another recipe for me.
I prompted the conversation by reiterating what I wanted ChatGPT to do.
It came up with Blood orange and honey roasted chicken legs. All cooked in one dish. Cooking time of 45 minutes, seemingly delicious and oddly familiar. Hmmm…. So I asked ChatGPT one last question:
“Find a recipe on the internet that is similar to this one.”
A recipe from ‘Not Quite Nigella’ came up that was almost identical. Can’t say the recipe images were, but hey, maybe AI can only create a proper image of what it has tasted.
Left: AI generated image of recipe. Right: image of identical recipe. (Image credit NotQuiteNigella.com)
My advice is: stick with human-written recipes that have been tried, tested and tasted. Chances are you’re getting a watered-down version with AI anyway.
Want to read the full conversation I had with ChatGPT?
Download a copy by clicking this link
Additional bits…
I decided to try out the recipe for myself. Here’s where I ended up.
I found the AI recipe a little bland so I added some Dijon mustard (1 tbs) and some fresh ginger (1 tbs) to the marinade to this step after I tasted it.
I also cooked the veggies separately as blood orange juice would have turned everything red and my kids probably wouldn’t have eaten them.
It’s also worth retaining the juices from the pan and thickening it up with some cornflour. Makes a delicious dipping sauce for those who want a gravy for their roast veggies. As I cooked the chicken in a seperate dish, I was able to work on the sauce while it was cooking off with the rest of the veggies.
