Data, Terroir, and the Dance Between AI and Human Taste

Just last year, we explored a wave of different AI studies and experiments at the crossroads of wine and machine learning: some focused on training models using the blind‑tasting methods of sommeliers and others on algorithms capable of passing the Introductory, Certified, and Advanced Sommelier Theory Exams. Those early projects were about imitation: could a machine absorb human habits and knowledge? Not entirely surprising, this year brings quick progress on the AI front. A new experiment has built an AI designed not to mimic theory or classroom practice, but to stand “nose‑to‑nose” with human tasters, and in a very different corner of the industry, another initiative is applying machine learning to a less tangible frontier: listening to the global conversation around wine itself.

In March, a blind tasting held in France staged a quiet but telling showdown. Twelve wines were selected by a court‑appointed official and served blind to two very different groups: a panel of sommeliers, critics, and seasoned tasters on one side; on the other, an artificial intelligence called Deep Red, developed by the Lyon-based company M&Wine. The assignment was the same for both – identify the wine’s country, region, and grape variety. The human tasters relied on their senses, knowledge base, and years of experience. Deep Red relied on data: a database of 35,000 wines from recent vintages and the ability to read a wine’s chemical fingerprint.

When the results were tallied, the machine was the clear winner. Deep Red identified the country of origin for all the wines, while the human panel averaged six. On the calls for the region and grape variety – the AI scored ten out of twelve; the panel averaged three per person. AI was helped by the fact that the lineup consisted of recent (post‑2020) single-varietal wines - the conditions Deep Red knows best. Its training data covered thousands of wines, but only from 2020 onward. But, as a sommelier, it does speak to how we will say that Italian wines taste “Italian,” etc. I look forward to more coming out on this – again, it is the idea of terroir being brought down to the molecular level.

AI isn’t quite ready to sit for the Blind Tasting portion of a Court of Sommelier exam just yet, as the contest also revealed its limits. A Jura wine – a pale, fragrant Poulsard – slipped beyond the patterns Deep Red had learned. With too few references to draw on, it failed. Beyond that, the machine’s abilities stop where human experience begins: it cannot (yet) judge wines made before 2020, and it has no sense for texture, complexity, or beauty. “The machine can process and recall data,” said Philippe de Cantenac, the competition’s organizer. “But it cannot taste. Humans find pleasure in a wine, and understand its uniqueness, in a way an algorithm cannot.”

Elsewhere, AI is shaping a very different frontier: storytelling. Vitabella, a Paris-based marketing and advisory firm, has built a tool not for tasting but for listening. Known for clients as storied as Primum Familiae Vini – a circle of historic families including Antinori, Pol Roger, Torres, and Tenuta San Guido – as well as estates such as Château Cheval Blanc, Louis Roederer, and Opus One, the platform gathers the vast digital conversation around fine wine - press, critics, influencers - and turns it into insight. Their new ChatGPT-based intelligence tool lets any winery see which markets are leaning in, which stories resonate, and where their voice carries farthest, all mapped in real time. It’s free to use, a gesture toward democratizing wine marketing.

Curious, I tested it on my own local market. When I asked which wines were the strongest sellers, it gave me a list - then offered to show which bottlings led the pack. I accepted. Next, it asked if I wanted to see which restaurants had the largest wine programs, and handed me a list that included a place that had closed years ago. I noted the error; it generated a new list. Then it asked if I wanted the sommelier and wine director contacts. Of course I did; many of these people are members of my local somm community, and often dear friends. The names it gave me were familiar… but the list would have been accurate seven years ago. Once I pointed this out, the list shrank considerably, and this time, it was right. So yes, human market knowledge still wins. But as a starting point - a basic map for wineries without deep local insight - this tool is surprisingly useful.

These two uses of AI - one decoding a wine’s chemistry, the other mapping its reputation - could hardly be more different. Yet both point to the same truth: in an era, awash in data, machines can help us see patterns no single mind can hold. If the March competition made anything clear, it’s that the best future for wine will be one where memory and information work side by side. Another point, also made clear, is that AI has yet to capture the pleasure, passion, and poetry that we see in wine or the connections and network of friends that we make along our wine journey.