There's a line in Eric Asimov's New York Times piece on AI and wine that deserves more attention than it got. Spencer Herbst, who builds artificial intelligence capabilities at PricewaterhouseCoopers in New York, described the experience of choosing wine in a restaurant this way: "It feels like a quiz you didn't prepare for." Someone whose job is building this technology, describing a wine list the way the rest of us do. That's the whole story, and it's exactly the story that gets lost whenever someone calls an AI chatbot "the new sommelier."
TL;DR
• General AI chatbots know a lot about wine, but they don't know anything about you
• Eric Asimov's verdict on ChatGPT for wine: "its floor is high, but its ceiling is low", accurate, and the gap worth understanding
• Building real wine confidence requires knowing your palate, not just knowing wine facts
• Sommy builds your Palate Profile over time so recommendations get sharper the more you taste and rank
• Nearly a third of French adults have already asked an AI to help choose a wine, so this is a settled habit rather than a novelty
What Does "Knowing Wine" Actually Get You at a Restaurant Table?
General AI chatbots have a genuinely high floor with wine. Ask ChatGPT to evaluate a wine list and it will reason through value, style, and food pairing with real competence. Eric Asimov, the New York Times wine critic, tested it against an actual restaurant list and found the picks were sound. He also found the ceiling: it never surprised him. It thinks inside the box. It recommended wines he already knew, solid, defensible, predictable.
That's fine if your goal is to avoid a mistake. It's not enough if your goal is to discover something.
The problem isn't information. Every general AI has absorbed millions of words about wine. The problem is that knowing a great deal about Burgundy tells you nothing about whether you prefer the earthy grip of a village Chambolle or the brighter, more lifted character of a Gevrey. That's palate-specific. And palate knowledge only comes from paying attention to what you've actually tried and whether you loved it.
Are Diners Really Choosing ChatGPT Over the Sommelier?
Yes, and it is measured now rather than assumed. The SoWine/Dynata 2026 barometer, which surveyed 1,003 French adults aged 18 to 65 in December 2025, found that 30 percent have already consulted an AI to choose a wine or a spirit.
The habit shows up on both sides of the restaurant table. French daily Le Figaro reported in September 2026 that restaurateurs are watching ChatGPT displace the sommelier recommendation, and that the pattern is strongest among diners in their twenties. Matthias Alberghi, a restaurant owner who used to be a sommelier, said customers get "a relatively straightforward, speedy response" and that the practice is "very common now".
Barcelona daily La Vanguardia found the same thing in Spain. Judit Ayago, sommelier at the Barcelona restaurant Masta, said her diners reach for AI because of the "solemnity" surrounding wine and "the fear of making a mistake". She put part of the blame on her own industry for being too technical and too distant from the people it serves.
Read that next to Spencer Herbst describing a wine list as a quiz he had not prepared for. A technology executive in New York and a working sommelier in Barcelona are reporting the same feeling, which is a stronger signal than either observation on its own. The demand here is not really for wine facts. It is for a way to order without feeling exposed.
Why Does the Confidence Part Matter?
The most revealing moment in the Times piece has nothing to do with ChatGPT's technical competence. Six months before Asimov spoke to him, Herbst started photographing restaurant wine lists and uploading them to ChatGPT. Where there's no sommelier, it stops him ordering blind. Where there is one, he uses it to have something to discuss with them. Asimov, who is not sentimental about any of this, says that use "has the right idea."
That's the gap worth closing: not "can AI name a good wine?" but "can it push you toward something you didn't know you'd love?"
General AI can name a good wine from a region you've never bought from. It knows Rebula, it knows the Brda hills, it can give you a confident summary. What it can't tell you is whether you specifically are the kind of drinker who would fall for it, someone who responds to high-acid whites with texture, who's been ordering Vermentino lately, who logged a Georgian orange wine as a favorite six months ago. Sommy builds that picture through your Palate Profile and uses it to make the leap from "this is a good wine" to "this is a good wine for you."
That's a different recommendation. That's the one that gets you to the bottle you didn't know to ask for.
What Happens When ChatGPT Doesn't Know the Wine?
It fills the gap, and it does not tell you that it has. Le Figaro took ChatGPT into a wine shop in Brittany and asked it to pair a bottle with mussels. It first suggested a Quincy, too expensive an appellation for the dish in Alberghi's judgement, then settled on a 2025 Loire Chenin Blanc. The former sommelier pointed out that the 2025 carries much less minerality and tension than the 2024.
The paper then checked the sources behind the answer and found the explanation. No tasting notes existed for the 2025 vintage yet, so the model extrapolated from the one before it. "When it doesn't know, it tends to invent," was the conclusion. Alberghi made the same point about the shop test: "While it normally relies on tasting notes from professionals, it also sometimes makes things up."
This is the failure a high floor hides. A general chatbot answers a question about an obscure 2025 bottling in the same assured register it uses for a wine it has read a hundred notes on, and nothing in the reply marks the difference. You cannot tell which of the two you are holding.
Which is why a recommendation is worth less than a recommendation that reports how sure it is. Where Sommy does not have enough to go on yet, it shows you a confidence level instead of a clean answer. That is less satisfying to read and considerably more useful to act on.
What Does "Confidence Builder" Actually Mean?
Dan Petroski, a Napa Valley winemaker quoted in the Times piece, put it well: "It may not be the perfect answer, but it can be a confidence builder." He was talking about ChatGPT. The phrase is more interesting when you apply it to what a personalized wine app can do over time.
A one-time chatbot interaction builds confidence for one meal. You know more than you did when you walked in, and that's real.
What builds wine confidence over months is something else: a record of what you've tried, what landed, what surprised you, and what you'd skip next time. It's the difference between reading a recipe once and actually knowing how to cook. The more you tell Sommy about the wines you've tasted, the sharper its recommendations get, because they're not drawn from what's statistically popular. They're drawn from you.
"Your wine confidence, built over time" is what Sommy is actually for. Not a smarter chatbot. A running conversation about your palate that gets better every time you drink something.
Can an AI Teach You What You Like?
Alberghi closed his remarks to Le Figaro with the sharpest criticism of this entire category, the part of it we occupy included:
"By relying solely on ChatGPT, we risk never truly knowing what we like. To understand your tastes, you have to try to open yourself up a little to this world, and that starts with thinking for yourself."
He is right, and the point deserves sitting with rather than a quick answer. Any wine recommender that only ever hands you a bottle is optimizing for tonight. It might get you through the aisle, and it leaves you exactly as uncertain the next time. Run that for two years and you have someone who depends on the tool and still cannot say what they actually like.
So Sommy is deliberately worse on day one. It will not claim to know your palate until you have rated five wines, and below that it shows a confidence level rather than a clean answer. Every dimension it infers about your taste is one you can overrule, and your correction wins over its own reading.
You should stop needing Sommy to tell you what you like. Finding it on a 400-bottle wine list is a different problem, and that one never goes away.
Why Does Using ChatGPT for Wine Feel Awkward?
There's a social cost to pulling out a general-purpose chatbot at the table, and it never shows up in a feature comparison. Typing a wine list into ChatGPT while people wait reads as homework. That reaction is worth taking seriously, because it shapes whether a tool actually gets used in the moment it's meant for.
Opening a general-purpose chatbot at a restaurant table, photographing the wine list, and waiting for a text wall of options signals something. It says you've brought a study tool to dinner. For some people that's completely fine. For a lot of people it amplifies exactly the self-consciousness that wine anxiety already creates.
A tool designed for that moment looks different. It's fast, it knows you, and you get a recommendation that reflects your actual preferences rather than a survey of what most people enjoy. Ordering wine with confidence at a restaurant should feel like a quick check with someone who knows your taste, not a research project. It should feel like checking a message, not like consulting a reference book at dinner.
Frequently Asked Questions
Is Sommy just another AI chatbot that knows about wine?
No. General AI chatbots know wine in the abstract. Sommy knows your wine: the bottles you've tried, the styles you gravitate toward, the regions you're curious about. Its recommendations come from your Palate Profile, not from a statistical average of what most drinkers enjoy.
What's the difference between asking ChatGPT and asking Sommy for a wine recommendation?
ChatGPT gives you a well-reasoned answer based on what's generally known about wine. Sommy gives you a recommendation based on what you specifically tend to love. The first answer is good for anyone. The second is built for you.
Does Sommy work at the restaurant table?
Yes, and that's one of the things it's designed for. Photograph a wine list and get a personalized recommendation based on your Palate Profile, not a generic survey of options, just the bottles that fit how you drink.
Can Sommy push me toward wines I've never heard of?
Yes. That's the point. Once Sommy understands your palate, it can identify wines in unfamiliar regions or from grapes you've never tried that match your preferences exactly. That is the difference between a tool that knows wine and one that knows you.
Does ChatGPT make things up about wine?
Sometimes, and it gives you no signal when it does. Le Figaro tested it in a French wine shop and found it describing a 2025 Loire Chenin Blanc with characteristics belonging to the 2024, because no tasting notes existed for the newer vintage yet. The paper concluded that when the model does not know, it tends to invent. The answer arrives in the same confident tone either way, which is why a recommendation that reports its own certainty is more useful than one that does not.
Wine confidence isn't the same as wine knowledge. What people actually need at the restaurant table isn't a fact database, it's a recommendation they trust. That trust comes from being known. And being known is what gets you past the bottles you already recognise.
Start building your Palate Profile and discover wines made for your taste.
