I have written a run of pieces this summer about what you can do with AI around wine, and there is a reader in my head who finishes part four thinking the thing tastes. It doesn’t, and it isn’t going to. Not because the models aren’t good enough yet, but because four of the most important things about a glass of wine are not in language at all.
This is the counterweight to the rest of the run.
Cork taint
A corked bottle smells of wet newspaper and damp cellar, and the annoying part is not that smell. The annoying part is that a light contamination smells of nothing: the wine simply smells of less. The fruit is gone, the wine is mute, and if you don’t know that can happen, you conclude the wine is dull.
A model can reproduce that description, because I just wrote it down. What it cannot do is judge your bottle. And this is the mistake most beginners make: they think a disappointing bottle is disappointing, when something is broken. You only learn it by having two glasses side by side and smelling the difference.
Reduction
Some wines smell of struck matches, boiled cabbage or wet flint when you open them. That is reduction, and it is usually not a fault. Swirl for a few minutes, pour it into a decanter, and it blows off.
I mention it because it is the clearest example of something advance advice cannot handle. The right move depends on what happens in the glass in the first three minutes. A model that cannot smell your bottle can only tell you the odds. You are the one who hears the difference between “this will go” and “this is staying”.
Temperature
The same bottle at 14 degrees and at 18 degrees is not the same wine. Warmer means more alcohol in the nose, less acidity on the palate, and an impression of sweetness that isn’t there. This is the cheapest improvement in the whole subject: most red wine in Dutch living rooms is drunk too warm and most white too cold.
A model can name the temperature, and mine does, because I put that rule in my profile. But it has no idea how warm your room is, how long the bottle has been standing on the table, or that your glasses come out of the cupboard next to the oven. The rule is easy. The execution is where the whole gain sits.
The case that differs
The most underrated one. Wine from the same producer, the same year, the same label can vary between batches. Bottle-to-bottle variation, a different bottling line, a pallet that sat in the sun in transit, and with larger brands even deliberately different blends for different markets.
Everything a model knows about a wine is about the wine as a name. What sits on your shelf is a specific case with a specific history of three months of lorry and warehouse. A good merchant knows this batch tastes different from the last one. That knowledge is written down nowhere, so it ends up nowhere.
What it can do
I have not suddenly turned against the tooling I built myself. A model is good at narrowing a choice, at patiently explaining what a term means, at writing down what you taste, and at keeping track of what you drank. That is a library and a notebook, and those two are worth plenty.
The point is the division of labour. Everything about wine that is language can be delegated. Everything about wine that is sensory stays yours, and the odd part is that the second category grows as you outsource the first. Stop having to remember which region sits where and you have attention left for what is happening in the glass.
What to do now
Buy two bottles of the same wine and pour them side by side at different temperatures. It costs ten euros and it is the fastest lesson in this whole piece.
And the next time a bottle disappoints: smell it again, and ask yourself whether it is disappointing or whether something is missing. That distinction cannot be prompted.