Atlas/About the project
About the project
Vinum ex Machina is the Atlas of Artificial Intelligence in Wine — a bilingual research project on the present and future of AI in viticulture, winemaking and the wine business. This page sets out the atlas’s structure and a summary of what it found.
What is on this site
Casebook
A register of audited cases of AI in viticulture, winemaking and the wine business worldwide. Every case was checked against its original source and graded for how well that source supports it.
OenoBench
A wine-knowledge benchmark for large language models: multiple-choice questions across six domains and four difficulty tiers, 16 model configurations scored against it, on both quality and cost. The scores separate what a model remembers about wine from what it can work out.
Keynote
The author’s essay on what AI in wine has delivered so far and on where the technology could go next. It argues from the work behind the case register and the wine benchmark.
Research
Where the atlas’s own papers and studies on AI in wine will be published. Proposals for collaboration and joint research are welcome.
Project summary
One rule runs through everything the atlas measured, and the keynote is where it is argued out: AI settles in wine wherever the answer arrives quickly. Irrigation, disease control and buyer behaviour are already counted more reliably by machine than by calendar and intuition; where the feedback comes once a year, with the vintage, it is still catching up. What governs the economics of a deployment is not how sophisticated the model is but how fast you can find out it was wrong.
The Casebook is where that shows. It holds 299 cases, each reopened against its original source and graded for how well that source supports it, and the industry splits along exactly that line: 40% of winemaking and laboratory cases are still research, against 15% in the wine business, where 71% have reached operation. The corpus is lopsided the same way — 177 records on the vineyard against 75 on the wine business and 47 on the cellar and the laboratory — because a vineyard produces observations by the hour, from a satellite pass to a weather station, and a cellar answers once a year, with the vintage. It reaches back to 1989, yet 146 cases started in 2023 or later: the subject is older than the web and half the corpus is younger than ChatGPT. 147 of the 299 — 49% — reached commercial operation or scale, and for 40 of those the record rests on the vendor’s marketing alone.
OenoBench asks the other half of the question — not what the industry has done with AI, but what AI knows about the industry. It puts 3,266 multiple-choice wine questions to 16 configurations of large language models, across six domains and four difficulty tiers, with every answer priced. o3 leads at 83.6%, and there is a leader but no podium: the top six sit within 2.6 percentage points of one another. What the run does separate is the two kinds of question. On those that need contextual reasoning rather than straight recall every configuration loses ground — 26.6 percentage points at the least (GPT-5), 39.6 at the most (DeepSeek-V3) — and the confidence interval excludes zero for 16 of the 16. Whether that is the same boundary the register draws, given that a cellar’s questions are almost all contextual, is what neither dataset can settle on its own; it is among the questions the research section plans to take up.
What comes next is in neither dataset yet. Nearly inevitable by 2028: risk-based spraying everywhere there is a data subscription, and language models in the back office — wine descriptions, customer replies, translations, reporting. None of those will look like “AI in wine”; they will look like routine work disappearing. Probable by 2032: an industry standard for measuring effect, the consolidation of data services around three to five platforms, and the first models that transfer between terroirs. What does not move is the boundary itself: the machine counts the measurable — litres of water, grams of copper, the probability of a customer leaving, a spot on a leaf — and what wine is actually drunk for has no correct answer and stays with the person making it.