Christmas party season is here, a time for congregating with friends indoors for hot whiskeys, mulled wine, brandy, and maybe something a little more singularitarian.
I had a burning question: “Can AI make new, original cocktails that taste good?” My AI-sceptic friend said, “I prefer recipes written by people with taste buds”. Large Language Models simply combine strings of words and don’t really know what anything tastes like – can they act like they do?
There are several major LLMs on the market – which LLM will make the best recipes?
Most importantly, will there be any delicious recipes here, ones worth keeping? Can I produce something that will please my friends at Christmas parties? And can I repeat that feat later for different themes?)
I gave the same prompt to five AI models: I asked them all “Invent some new Christmas-themed cocktails.”
- ChatGPT 4o generated ten cocktails
- Gemini (formerly Bard) generated five cocktails
- Mistral generated six cocktails
- Claude generated five cocktails
- Llama (Llama 3.1 405B Instruct to be precise) generated five cocktails
You can read the raw outputs here in the interest of transparency. That adds up to 31 cocktails. This had to be trimmed down for two reasons: firstly I can’t drink 31 cocktails, and secondly I can’t be expected to get the likes of ‘chestnut liqueur’ or ‘lavender bitters’ for this experiment.
We ended up making and tasting 11 Christmassy cocktails:
- Cranberry Moscow Mule (ChatGPT)
- Santa’s Spiced Mule (ChatGPT)
- The Holly Jolly Julep (ChatGPT)
- The Yule Tide Toddy (ChatGPT)
- Jingle Bell Julep: (Gemini)
- Merry Mistletoe Mojito: (Gemini)
- Mistletoe Mule (Mistral)
- Eggnog Espresso Martini (Mistral)
- Spiced Apple Cider Sangria (Mistral)
- Gingerbread Martini Wonderland (Claude)
- Reindeer’s Kiss (Llama)
Which AI’s recipes are hard to make?
60% of ChatGPT’s cocktails were eliminated for being too hard to make, 60% of Gemini’s, 50% of Mistral’s and 80% of Llama’s. Llama seems pretty bad on this metric, but it’s too little data to draw a conclusion.
AIs in general seem to have decided (see the raw output) that white chocolate liqueur + peppermint schnapps + vanilla vodka is the taste of Christmas. Those are obscure ingredients to source. Should we make those cocktails in a Part Two of this article? Let me know in the comments.
This is as much my fault as the AIs’ – I could have altered the prompt to exclude obscure ingredients. However, I wanted to keep the prompt simple for the purposes of this experiment, and to give all the models the same prompt for a level playing-field. If my aim had been only to make great cocktails (and not to test the models), a more specific prompt would be better.
The Yule Tide Toddy, by ChatGPT
ChatGPT says –
A warm, comforting cocktail with bourbon, honey, and herbs. **Ingredients:** - 2 oz bourbon - 1 tbsp honey - 1/2 oz fresh lemon juice - 4 oz hot water - Fresh thyme sprig (for garnish) **Instructions:** 1. In a mug, combine bourbon, honey, and lemon juice. 2. Pour in hot water and stir until honey dissolves. 3. Garnish with a sprig of fresh thyme.
Verdict: this is low on the creativity; the AI just took a normal hot whiskey and threw thyme at it to make it Christmassy. But doesn’t hot whiskey usually have cloves? Cloves are arguably more Christmassy than thyme “except for the fact that thyme looks like a Christmas tree”, according to one reviewer.
Two points worth commending: adding squeezed lemon juice is an improvement over the usual hot whiskey procedure of adding a slice of lemon, and the quantities of lemon juice and of honey are basically spot-on.
Overall, our group felt this was a middling cocktail.
Santa’s Spiced Mule, by ChatGPT
A holiday twist on the classic Moscow Mule with rich spices and cranberry.
**Ingredients:**
2 oz vodka
1 oz cranberry juice
1/2 oz ginger syrup (or fresh ginger juice)
1/2 oz fresh lime juice
4 oz ginger beer
Cinnamon stick and fresh cranberries (for garnish)
**Instructions:**
1. In a mule mug or glass, combine vodka, cranberry juice, ginger syrup, and lime juice.
2. Add ice and top with ginger beer.
3. Stir gently, and garnish with a cinnamon stick and cranberries.
Verdict: overall, this is a good drink. The smell of gingerbread syrup is nice, and goes well with the cinnamon and ginger beer – especially with the ginger beer because you see normally when you drink ginger beer it only has a mild ginger flavour, and adding a stronger ginger flavour cleverly fixes that.
The drink does taste Christmassy. “I think it’s too sweet personally,” said one reviewer, and it suffers from the same weakness as the Cranberry Moscow Mule below: too much sweetness, and arguably should have doubled the lime juice.
Cranberry Moscow Mule, by ChatGPT
A festive take on the Moscow Mule with fresh cranberry and lime.
**Ingredients:**
2 oz vodka
1 oz cranberry juice
1/2 oz lime juice
4 oz ginger beer
Fresh cranberries and lime wedges (for garnish)
**Instructions:**
1. In a mule mug, combine vodka, cranberry juice, and lime juice.
2. Add ice and top with ginger beer.
3. Stir gently and garnish with cranberries and lime wedges.
This was overall one of the worst. It was sweet and not much could be tasted beyond ginger beer. There is quite a lot of mixer to vodka, and that took away the alcoholic bite. Cranberry juice isn’t assertive enough to redeem it, and lime juice could be, but not a half-measure. We commented on this and the other ChatGPT mule (above) that doubling the lime juice would be an improvement. And that’s exactly what the French model did –
Mistletoe Mule, by Mistral
- **Ingredients:**
- 2 oz vodka
- 1 oz cranberry juice
- 1 oz lime juice
- Ginger beer
- Fresh cranberries and a sprig of rosemary for garnish
**Instructions:**
- Fill a copper mug with ice.
- Add vodka, cranberry juice, and lime juice.
- Top off with ginger beer.
- Stir gently and garnish with fresh cranberries and a sprig of rosemary.
In the mule head-to-head, Mistral is the clear winner. The ChatGPT mules were both too sweet and had too much ginger beer, and too little bite and lime. Mistral fixed this, doing just what the doctor ordered by doubling the lime juice. Is this to do with the way Americans like things sweet, and Mistral is French tech?
The other difference is the garnish was rosemary in place of lime wedges; this is probably also an improvement. “I think the rosemary gives something to it,” said one of our reviewers.
Merry Mistletoe Mojito, by Gemini
2 oz white rum
1 oz lime juice
1/2 oz simple syrup
A handful of fresh mint leaves
A few cranberries
Muddle the mint leaves, lime juice, and simple syrup in a highball glass. Add the rum and fill the glass with crushed ice. Top with a splash of club soda and garnish with cranberries and a sprig of mint.
Verdict: all the flavours here are weak. We could taste the mint a little, and the lime a little, but it’s largely just rum and sugar. A bland cocktail and among the worst on the list.
Jingle Bell Julep, by Gemini
2 oz bourbon
1 oz simple syrup
A handful of fresh mint leaves
A splash of club soda
Muddle the mint leaves and simple syrup in a julep cup. Add the bourbon and fill the cup with crushed ice. Top with a splash of club soda and garnish with a sprig of mint.
This was even worst than Gemini’s Merry Mistletoe Mojito; that mojito already was too sweet, and this one doubled the simple syrup. Jack Daniel’s (the bourbon we used), when paired with sugar and club soda winds up tasting something like off-brand cola.
The Holly Jolly Julep, by ChatGPT
A julep with a holiday twist using mint, pomegranate, and bourbon.
**Ingredients:**
- 2 oz bourbon
- 1 oz pomegranate juice
- 1/2 oz simple syrup
- Fresh mint leaves
- Crushed ice
**Instructions:**
1. Muddle mint leaves and simple syrup in a julep cup.
2. Fill the cup with crushed ice and add bourbon and pomegranate juice.
3. Stir well and garnish with a sprig of mint.
ChatGPT has been defeated in the mule head-to-head by Mistral. Can it redeem itself in the julep head-to-head? Yes it can; this is an interesting twist on the julep, because the combination of pomegranate juice and bourbon is creative and they do compliment each other. The pomegranate deepens the flavour of the Jack Daniel’s. ChatGPT avoided Gemini’s error of using too much simple syrup, and it had a stroke of real creativity with that combination.
Eggnog Espresso Martini, by Mistral
- **Ingredients:**
- 2 oz vodka
- 1 oz espresso
- 1 oz eggnog
- 1 oz coffee liqueur
- Grated nutmeg for garnish
**Instructions:**
- In a cocktail shaker, combine vodka, espresso, eggnog, and coffee liqueur with ice.
- Shake well and strain into a chilled martini glass.
- Garnish with grated nutmeg.
This was a good-tasting cocktail – people were talking about giving it 8 or 9 out of 10 for taste. The nutmeg gives it a nose, which I liked as it added complexity, but it did draw some criticism from my friends who just dislike nutmeg. The balance of vodka and coffee was pretty good, but maybe if there was slightly less vodka it would be better. First you smell the nutmeg, then you taste the coffee, then you taste the booze – an interesting and complex sensory experience.
Spiced Apple Cider Sangria, by Mistral
**Ingredients:**
- 1 bottle of red wine
- 2 cups apple cider
- 1 cup brandy
- 1/2 cup orange juice
- 1/4 cup lemon juice
- 1/4 cup simple syrup
- 1 apple, sliced
- 1 orange, sliced
- 1 lemon, sliced
- Cinnamon sticks and star anise for garnish
**Instructions:**
- In a large pitcher, combine red wine, apple cider, brandy, orange juice, lemon juice, and simple syrup.
- Add sliced apple, orange, and lemon.
- Stir well and refrigerate for at least 2 hours to let the flavors meld.
- Serve over ice and garnish with cinnamon sticks and star anise.
Our reviewers said they could barely taste the alcohol, but I assured them it was there: a bottle of wine and a cup of brandy is not a weak brew. Sangria tends to hide its teeth. The fruit tastes boozier, and when I distributed forks and teaspoons, the fruit drew coos of approval.
Kiki suggested that mulled wine is a spiced Christmas drink, and sangria is a summery drink – maybe this was the AI’s attempt to hybridise the two? It’s an interesting concept that would be worth exploring more; for instance, you could prompt a LLM ‘Create a recipe for a Christmassy sangria inspired by mulled wine’.
The mulled wine notes came mostly from the cinnamon and the star anise. If I had doubled or tripled the amount, or the time spent brewing, they would have been stronger. (I put in 2-3 cinnamon sticks and 4-5 stars for about four hours: a reasonable amount. I would advise readers to be less reasonable and throw in fistfuls.)
While making it, I thought that ¼ cup lemon juice plus a lemon is a lot of lemon. It didn’t taste exactly sour and lemony, but dialling it back a little probably would be an improvement, we all agreed.
Overall a good drink. Nothing is terribly out of balance, though more spice and less lemon were suggested as improvements. Then I heated it up on the stove and served it warm. It smelled amazing warmed up, and we all preferred it.
Gingerbread Martini Wonderland, by Claude
A creamy, spiced cocktail featuring vodka, gingerbread syrup, and a graham cracker rim
Ingredients:
* 2 oz vodka
* 1 oz gingerbread syrup
* 1 oz heavy cream
* Dash of cinnamon
* Graham cracker crumb rim
- Garnish with a tiny gingerbread man cookie and a sprinkle of nutmeg
A website told me to replace ‘graham crackers’ (which aren’t available to me) with “a mixture of digestve and rich tea biscuits”, so I did that and it worked well.
One reviewer said, “Tastes like a White Russian with a bit of biscuitty gingeriness; it’s like eating a ginger biscuit and drinking a White Russian at the same time”. This cocktail was well-liked in our group, and people commended its creativity, though it did receive some complaints for having too much vodka.
Reindeer’s Kiss, by Llama
* 2 oz Whiskey
* 1 oz Spiced apple cider
* 1/2 oz Maple syrup
* 1/2 oz Lemon juice
* Dash of Cinnamon powder
Mix whiskey, spiced apple cider, maple syrup, and lemon juice in a shaker filled with ice. Strain into a rocks glass and sprinkle with cinnamon powder. Garnish with a cinnamon stick and a cherry, if desired.
I presume what they mean by ‘cider’ here (and also in the Spiced Apple Cider Sangria) is the American non-alcoholic kind. This ambiguity is annoying, but happens in human-written recipes as well. It wasn’t easy to get that ‘cider’ in Ireland, but Llewellyn’s Orchard make it in Lusk.
This was Llama’s only cocktail that got past the qualifying stages, four of its five suggestions having been eliminated as impractical. How did it do? Spectacularly. It was the best cocktail in the game with no close second. My coworkers are making it and I am going to the market to get the hard-to-source cider before Christmas Day. The whiskey (I used Jameson), the apple, the spice, are all present in perfect proportion, and it’s just delicious in a way I can’t put into words, so you know what to do.
Best and worst cocktails
Winner: Reindeer’s Kiss (Llama)
Honourable mention: Gingerbread Martini Wonderland (Claude), Eggnog Espresso Martini (Mistral)
Worst: Cranberry Moscow Mule (ChatGPT), Merry Mistletoe Mojito (Gemini), Jingle Bell Julep (Gemini)
How did the AI models compare?
Poor Gemini, it’s not looking good for you. The two cocktails that Gemini generated were bad. ChatGPT was hit-and-miss but pretty fine overall, and Claude and Mistral similarly held their own.
Llama did the worst in the qualifying round; I mean that it suggested things that I can’t really make. But it more than redeemed itself by suggesting the clear stand-out of the tournament. With a more specific prompt, it could perhaps improve on practicality. Going on this little data it is hard to draw solid conclusions, but either way we have a superb new recipe to tell our friends and readers about.
Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.