Depending on the mood, artificial intelligence is either biding its time before poisoning the water — and frankly, it’ll have to get in line for that — or it’s set to be a happily servile addition to the human race, sorting out the boring bits while we’re all out necking Martinis or wild foraging or something (adjust to taste).
Either way, a change is going to come. AI’s apparently unlimited brain means it seemingly can turn its ghostly hands to anything. Earlier this year, reports came out of ChatGPT inventing the “best cocktail in the world”. Safe to say, this concoction of (deep breath) gin, lime juice, elderflower liqueur, absinthe, Cointreau, orange bitters, honey syrup, sparkling wine and an orange twist proved that its name, the Heavenly Sipper, was a bare-faced lie. Heavenly? Hellish is the word. But it’s an interesting thought. Could AI really understand how to build drinks better than a well-trained mixologist? What about the best in the world? And so the Standard turned to Chelsie Bailey, who is the head bartender at the Savoy’s American Bar, generally held to be the most prestigious gig in the game.
The concept? Ready Steady Cook for cocktails; pick two ingredients and a glass, type it into ChatGPT and wait for instructions. One team worked exclusively to the measures and method mandated by AI, the other used actual intelligence. Representing humankind was Bailey. On team Robocop, the Standard’s David Ellis and drinks writer Rob Buckhaven, artfully combining an astonishing lack of know-how with boundless inefficiency.
So, the next time you’re in need of a drink and wondering whether to turn to a laptop or the learned ear of a bartender, the strictly scientific results are below — the calamities can be watched above, too.
Rob’s take
AI will inevitably take our jobs, it’s just a question of when. With that gloomy mantra burrowing into my brain, I was nervous when we challenged a top bartender to a cocktail-making duel against ChatGPT, one of the AI chatbots poised to swipe our livelihoods.
To kick off, cocktail one: the drinks chosen were sweet vermouth, Fernet-Branca (Ellis playing the fool, as usual). AI’s offering comprised 60ml sweet vermouth and 30ml of Fernet-Branca in a Martini glass, which tasted like a teenager’s attempt at a cocktail using bottles pilfered from their parents’ drinks cupboard. Cloying, unbalanced, resembling a Poundland Long Island Iced Tea. Bailey’s creation showed harmony, dialling back the sweet vermouth for a delicacy I thought impossible with Fernet in the mix.
Cocktail two called for an Old Fashioned-style cocktail using 15ml Green Chartreuse and 60ml Mezcal, with a sugar cube and Angostura bitters thrown in by the Dalek. No amount of stirring could dissolve the sugar dune or balance the alarming quantity of Chartreuse, delivering an overbearingly saccharine thwack of anise. Being a pro, Bailey went lighter on the Chartreuse and used simple syrup for a vivifying drink that made you want to play pétanque in a local square somewhere.
To encourage Marvin the Paranoid Android to raise his game, we specified for the last drink to “make this in the style of The American Bar at the Savoy”. Ingredients included 45ml Cynar and 15ml Maraschino, served in a highball glass. Buffering long and hard, Marvin riffed on our instructions by adding in soda water, 30ml lemon juice and 15ml simple syrup. Although the toppy measure of Cynar gave me a numb tongue, the flavour combination showed potential. Once again, Bailey comes out on top; she knows better than to add so much Cynar and dials down the Maraschino completely too.
For a laugh, and perhaps because we were three drinks on an empty stomach down at this point, I asked Bailey to prepare ChatGPT’s the Heavenly Sipper. The list of ingredients is long — and mentioned at the start of this piece — so I’ll skip to the result. An unwieldy tangle of flavours which confirmed that Bailey’s job security remains intact, and that AI’s literal take on cocktail-creation can’t compete with the nuance of human experience. Well, yet…
David’s take
Oh, I’d expected it to taste like filth — but really, I thought, this just takes the piss. Jesus. Buckhaven’s explained the rules so I won’t bore on, but for round one we’d put together sweet vermouth and Fernet-Branca. I hesitate to call AI insouciant — insouciance require a certain debonair sense of flair, and that, I think, must surely be innately human — but I can’t say ChatGPT seemed to care much for measures. In goes a great whack of Fernet, in is a slosh of vermouth. And even that’s after we’ve shaken out the Americanisms; early suggestions are in ounces, not millilitres, and the portions of alcohol are far larger than your typical Brit would expect. Calling on the wrong experts, I reckon.
As I see it, AI is just much better at Googling than I am. It can scour the internet in the fraction of a second, weigh up all the information and pluck the best possible answer from those out there. But sipping a monstrously unsubtle Hanky Panky riff, I realise that perhaps the AI is off its game this morning. Everything just feels a bit skew-whiff. Has it already had a couple itself? Out of the four of us — and with full respect to Bailey, who I acknowledge is actually a professional — I really didn’t have the computer down as the one most likely to be half-cut pre noon.
That said, I wonder if there’s a chance ChatGTP is being stubborn. Buckhaven appears faintly maniacal with power. “Make me a drink,” he demands of the laptop, narrowly avoiding slapping the screen. Though given the clumsiness of his typing, it may have actually been a hand-to-eye coordination issue. On we press.
The following drinks are little better. I begin to wonder if ChatGPT does not crawl Google at all, but instead is figuring out its methods by sight, perhaps scanning old episodes of Mad Men. One, a puddle of Green Chartreuse and Mezcal, is meant to be an Old Fashioned of sorts; only, there are no instructions other than to bosh liquids in over a sugar cube. Granted, the stirring is obvious — but not if you’ve nevermade an Old Fashioned before. The sugar cube floats ominously. I suspect AI can explain what to buy, but not how to use it. Buckhaven says our drink makes his tongue numb, which silences him, and so is a plus.
I notice Bailey is galloping furlongs ahead because hers is not a mechanical mind reciting recipes by rote; it is a creative mind, tweaking and adjusting, with a fine sense of proportion and subtlety. The ice she uses for mixing, for instance, isn’t the ice she serves drinks with; her garnishes are not wedges but slices; there are gentle additions of balancing agents. In the opposite corner, Buckhaven and I are following the AI’s instructions to the letter, and the result is pure amateur hour. And as it goes, I’m more than happy for the bot to take the blame.
Chelsie’s take
I can’t say I was nervous before taking on ChatGPT. I’d sized it up beforehand and played about with it, I think it’s pretty good when you have ingredients you don’t know what to do with — but to create an original drink or a world-class drink, it’s simply nowhere near as developed as a human bartender.
I always say cocktails are a bit like the game Six Degrees of Kevin Bacon; you can always move from one to the other with a few changes. Say you want to move from a Margarita, swap in vodka in place of the tequila and add cranberry and you have a Cosmopolitan, that sort of thing. I imagine that’s how the AI works but it can’t taste things. I’ve got almost two decades of experience, so I just know which products work well with others. Not every triple sec or vermouth is the same, after all; they all fit together differently, but AI doesn’t realise that.
The other thing that I noticed is that until it was specifically asked to add other ingredients, the AI didn’t think to try using other bits and pieces. It literally took the ingredients and suggested what it thought was the best way to balance those — and that’s not the best way to design a cocktail. Adding little things, making adjustments, that’s where the good stuff comes in. But what AI offers is just a very, very basic skeleton of a cocktail. The principles are there, but it doesn’t have the skill set to really elevate the drink.
I also think it’s the human interaction that makes going out for a drink. For decades now, there have been cocktail machines that pour out the right measures, but it’s never taken off. I imagine this could be the same. And everything the boys made was something I’d expect to have at a house party — that could have just been them, though. I’m absolutely not worried AI is coming for my job — and neither are Buckhaven or Ellis, by the looks of what they poured out.
