⏱️ Reading time: 7 minutes
Thirty percent of French people say they have already consulted an artificial intelligence to choose a wine or a spirit. Among 18-to-25-year-olds, the figure climbs to 58%. That number, drawn from the latest SOWINE/Dynata barometer, marks a turning point: for the first time, the machine has officially invited itself into the French drinker’s purchase journey. We ask it for a pairing for tonight’s dinner, an opinion on a bottle received as a gift, an idea for breaking out of our habits. And it answers — quickly, politely, with disarming self-assurance.
One question, however, hangs in the air — one the surrounding enthusiasm carefully avoids: what, exactly, is that recommendation based on? When a sommelier advises you on a wine, they draw on years of tastings, of aromas, of memory. The AI, for its part, has never tasted anything. Can you really recommend the right wine without a palate?
THREE WAYS TO ADVISE ON A WINE WITHOUT HAVING DRUNK A DROP
Behind the “AI sommelier,” three engines
The most widespread recommendation engine doesn’t analyse the wine: it analyses the drinkers. This is collaborative filtering — Vivino’s method, borrowed from Netflix and Amazon. Its principle fits in a single sentence: people who liked the same wines as you also liked this one. The suggestion is born of a statistical correlation between millions of ratings, not of any understanding of the liquid. Effective at guessing what will please you, blind to what is actually in the glass.
A second family reasons about the wine itself — or rather about its technical sheet. Apps such as WineRing or the iAlacarte sommelier match an aromatic profile (tannic structure, acidity, residual sugar, intensity) against that of a dish (cooking method, sauce, garnish). It’s the classic logic of food-and-wine pairing, encoded in a database. Finer than collaborative filtering, it still depends on descriptors entered, upstream, by humans.
The third family, the most recent and the most spectacular, rests on large language models. Queried in natural language — “I’ve got a duck breast with cherries and three bottles in my cellar, what would you suggest?” — the model produces a reasoned, often startling answer. But all it does is reassemble, on the fly, what thousands of tasters have written on the web. It compiles; it does not experiment.
These three approaches share one blind spot. None of them tastes. They predict from proxies — co-occurrences of ratings, sheets of descriptors, scraped text. Wine, as a sensory object, remains strictly inaccessible to them. And that is precisely where the science of the palate becomes illuminating.
THE PALATE ISN’T IN THE MOUTH — IT’S IN THE BRAIN
What brain imaging reveals about the sommelier
In 2005, a team published in NeuroImage the first functional magnetic resonance imaging study devoted to sommeliers. Seven experts, seven novices, the same wine. The result: in the experts, tasting activates a particular network — the left insula and the orbitofrontal cortex, two regions dedicated to integrating taste and smell. The researchers’ conclusion is counter-intuitive: that difference reflects not finer taste buds, but a learned ability to connect sensations to memory.
The studies that followed confirm and refine this. In 2014, a Besançon team (Pazart et al.) showed that the expert reacts more immediately and more “economically” than the novice, drawing heavily on memory structures — the hippocampus, the temporal pole. Where the beginner gropes through diffuse associative areas, the professional evaluates the wine and recognises its type in a single movement. In 2016, a study of Master Sommeliers observed a reinforced memory network; in 2024, researchers at the Basque Center on Cognition measured, in sommeliers, more intact white matter in the tracts linking perception, memory and language. Training, quite literally, reshapes the brain.
Taste isn’t on the plate; it is built inside our heads.
Gabriel Lepousez, neurobiologist at the Institut Pasteur
For this specialist in sensory perception, tasting is an operation of formidable complexity: the brain must rank information arriving through the eyes, through the nose (directly and by retronasal olfaction), through the mouth and through touch, synthesise it, compare it against memory, then translate it into words. One detail illuminates all the rest: sight alone occupies 15 to 17% of the cortex, against barely 1% for smell. Faced with uncertainty between an odour and a colour, the brain almost always rules in favour of what it sees.
The human expert, fallible by design
This primacy of the brain has a flip side: it makes perception manipulable. The most famous experiment is that of Frédéric Brochet, Gil Morrot and Denis Dubourdieu, conducted in Bordeaux in 2001. Fifty-four oenology students were given a white wine to taste, then the same white wine dyed red with a neutral colourant. The verdict: the students described the second using the vocabulary of red wines — black fruit, tannins — as if the colour had rewritten their sense of smell.
A second experiment by the same researcher is just as cruel: a single Bordeaux, presented two weeks apart under the label “table wine” and then “grand cru,” drew radically opposite judgements. In other words, even the most seasoned taster remains subject to expectation, context and price. Their strength — a brain that integrates everything — is also their weakness.
SOMMELLERIE: AUGMENTED OR REPLACED?
What the machine gets right — genuinely
Against this backdrop, the AI has real strengths it would be dishonest to brush aside. First, it democratises: a sommelier’s advice is expensive and intimidating; an assistant available around the clock, free in its basic versions, lifts that barrier. Next, on everyday food-and-wine pairing, reasoning by aromatic profile is solid — it applies, without faltering, principles the average drinker doesn’t know. Finally, in a twist of irony, the AI escapes Brochet’s traps: neither colour, nor label, nor price throws off a machine that sees nothing. But that is precisely because it perceives nothing at all.
What it will not do
For that is where its competence ends. The AI will not taste the bottle opened in front of you: it cannot tell you that it’s corked, that it cooked in a car boot over the summer, that it’s going through a dumb phase. It reasons about an abstract, average, statistical wine — never about the real wine in the glass. It doesn’t read the table, doesn’t sense that a guest would rather avoid red or that the mood calls instead for something sparkling. And when it promises to choose a wine “to match your mood,” that is where it is most helpless: pairing to a dish can be encoded; pairing to a state of mind remains, for now, pure marketing.
A librarian, not a taster
Perhaps we’re simply reaching for the wrong metaphor. The AI sommelier is not a sommelier: it is a librarian of wine, prodigious at retrieving, cross-referencing and serving up information scattered across millions of bottles and reviews. On that ground — “what is a Crozes-Hermitage?”, “is this Pomerol any good?”, “which red with an osso buco?” — it renders a real and democratic service. But recommending the right wine remains, in an irreducible part, an act of embodied sensory memory that the machine cannot perform. The SOWINE barometer puts it in its own way: AI enters the picture not as a substitute, but as a new tool for guidance. The palate, meanwhile, stays between the two ears.
Would you trust an AI to choose the wine for your next dinner — or do you keep the final word? Tell us in the comments, and find our by-the-glass tastings on Instagram @wineandtech.
Sources: SOWINE/Dynata Barometer 2026 · Castriota-Scanderbeg et al., NeuroImage, 2005 · Pazart et al., Frontiers in Behavioral Neuroscience, 2014 · Banks et al., Frontiers in Human Neuroscience, 2016 · Basque Center on Cognition, Human Brain Mapping, 2024 · Morrot, Brochet & Dubourdieu, “The Color of Odors,” Brain and Language, 2001 · Gabriel Lepousez, Institut Pasteur (interviews in Le Figaro and Le Rouge & Le Blanc).
Leave a Reply