Stop me if youre heard this one before: AI is coming for all of our jobs. It will vacuum the knowledge out of every programmer, lawyer, and accountant alive, then systematically replace them with AI chat bots. The next time you call the IRS for help on your taxes, the voice of Scarlett Johanssen will explain the difference between a W2 and a 1099.

Or so were told. But try telling that to the restaurant owner down the street, where help wanted signs have been a permanent fixture since 2020. AI may not show up hungover, argue with the wait staff, or ask for a raise, but it also cant carve up a side of beef, flambe bananas foster tableside, or bus your table.

Safe as the service industry may seem from AI disruption, theres one quintessential knowledge worker among the staff at any high-end restaurant who may be in a different boat: the sommelier. As a wizard of wine and keeper of all its related trivia, the role of a sommelier may seem more automatable than a waiter or a chef, but its more nuanced than it looks.

Make me a match

If its been a while since your last dinner at a Michelin-star restaurant, a sommelier is the restaurant employee, perhaps donning a dapper vest and apron, who helps you select a wine to accompany your meal. Behind the scenes, he or she may also help decide which wines an establishment carries, stock and rotate cellars, and train the rest of the staff on the finer points of pairing food and wine.

Unlike the restroom attendant who hands you a paper towel after youre done washing your hands, you probably do need a sommeliers help. In restaurants with thousands of bottles in the cellar, even a visiting wine aficionado could hardly be expected to be familiar with all of them, much less know which one pairs well with a slow-cooked leg of lamb and brussel sprouts.

Perhaps you know someone with a particularly expansive bottle collection who fancies themself a sommelier, but theres more to it than showing off to neighbors with the contents of your recycling bin. Like many old-world professions, a web of guilds and professional associations exists to anoint professionals, from Britains Court of Master Sommeliers to Frances Union de la Sommellerie Franaise. Becoming a certified sommelier requires attending tastings, passing a written exam, and typically, blind taste tests.

So is AI up for the job?

Genie in a bottle

Modern artificial intelligence can outperform 89% of humans on the SAT, pass the state bar exams required to become a lawyer, and pass the United States Medical Licensing Exam required to practice as a physician. So naturally, we had to try a sommelier exam.

We fed the Court of Master Sommeliers introductory practice exam into Claudes 3.5 Sonnet model, without the attached answer key, and it scored 94%, missing just four questions. ChatGPT scored 97% and missed just two questions, on account of being unable to process the accompanying images.

Impressive. But what about in practice? Tell ChatGPT about the leg of lamb youve been eying for dinner, and it can quickly spit out half a dozen recommended wine varietals that would accompany it well, along with cogent explanations of why.

Cabernet Sauvignon The tannins help cut through the fattiness of the lamb, while dark fruit and herbal notes complement its flavor.

Of course, those recommendations may not be particularly helpful if you dont just happen to have a 1999 Far Niente to dust off in the cellar, but you can also upload a screenshot of a menu and it will refine its recommendations to work with what you have available. Cooking at home? Just take photos of the bottles in your cabinet. Try it for yourself.

And the lamb recommendation holds up, at least according to Decanter magazine, which would agree with pairing a cabernet sauvignon, based on the advice of Master Sommelier Kathrine Larsen-Robert. Perhaps thats because, more than likely, ChatGPT read the same article.

Large-language models like ChatGPT work by learning from a massive reference body of information, remixing it, and spitting it back out. Nobody sprinkled a 2001 pinot noir on a circuit board; the LLM is simply parroting what people say about that wine online, like a middle schooler who just discovered Reddit.

While that has led to some high-profile misfires, (see: Google AI Overviews telling people to eat rocks), in the low-stakes world of recommending libations, theres more room for error than in the ER. Wine is, after all, notoriously subjective: Two human sommeliers might recommend different wines to accompany your meal, but wed probably conclude they have different opinions before we say that one is flat-out wrong.

So case closed: AI can recommend wine. But thats far from a sommeliers only job.

The human touch

Imagine splurging on dinner at a legendary restaurant that always books out three months in advance. You know, the one Frasier would go to: plush red booths, starched white tablecloths, a jazz band tucked in the corner. The maitred whisks you to your candlelit table, where a waiter hands you a leatherbound wine menu thick enough to be mistaken for a Bible. Whats goes well with the filet mignon? you ask as your waiter produces a glowing iPad, keys in your question, and clumsily reads back its suggestions: Pinot Noirs bright acidity and red fruit complement the meats delicate texture.

That might be good advice. But Frasier would not be impressed.

While knowledge is part of what a sommelier brings to the (literal) table, service is the other, arguably more important half: A broad smile, witty banter, an anecdote about the time they got lost in the back roads of Burgundy on the way to sample the wine youre about to taste, at the vineyard where it was grown.

Why, after all, do we pay $100 for a bottle of wine we could buy for $30? Were not getting ripped off, were paying for the experience, which is much more than the food on our plates or the fermented grapes in a glass.

Thats not to even mention the other factors that a human sommelier can supplement generic advice with. Your sommelier knows the way your filet mignon is seasoned, because shes tried it herself; she knows the wine is stored at the correct temperature because she stocked the cellar; she intuits from the mustard stain on your suit jacket that shed be better off suggesting a $70 bottle than a $700 bottle.

So yes, AI can replace a sommelier in the same way that big-screen TVs can replace theaters. In a technical sense, they do the same thing. But your TV, however many pixels it packs, doesnt replicate the experience of seeing Mad Max with a couple hundred other people any more than an AI-generated wine suggestion can replicate a night at La Tour d'Argent.

The most damning argument against AI replacing sommeliers may be this, though: Somebody needs to actually taste the wine. After all, the knowledge AI is trained on ultimately comes from human experience. Generations of winemakers and wine appreciators have spent literally hundreds of years tasting, documenting, and sharing their experiences in a way that AI can index. If that ended tomorrow, so would AIs knowledge of current wines.

So sommeliers, have a glass and relax. AI may come for your brain, but your tongue is safe for now.