Alder Yarrow’s article on AI and wine is spot on in terms of the current state of AI and wine, but gloomily off-target regarding its consumer-facing future.
It is absolutely true that ‘most, if not all, of the AI-driven tools for wine that consumers can get their hand on are utter crap’. That’s because the current crop is going about it in exactly the wrong way. They fundamentally misunderstand how people experience and think about wine. And they seem to pursue AI for its own sake, rather than treating it as the incredibly useful complement to wine professionals that it must be.
The whole notion that the promised land for AI in wine lies in some singular, comprehensive database of all conceivably measurable aspects of millions of wines is understandable but wrong. The conceit reveals a grievous misunderstanding of wine, one that I find all too common among people who’ve never worked in a wine store.
As founder of Vagabond and someone who’s spent years on the shop floor, believe me, the trick to trustworthy wine recommendations is not ‘I like raspberries, therefore I’ll like all wines with notes of raspberries’. After all, you might like wines that don’t have hints of raspberry, and you might dislike wines that do. Things like mood, occasion and season can also influence your prevailing desire for anything raspberry-related. This all-too-common focus on superficial flavours and smells is a red herring, and a counterproductive one at that.
There are certain characteristics, however, which get close enough to the core of why we like the wines we do to be usefully predictive. For instance, one dimension of red wine that I’ve found to be a reliable guide over the years is the ‘fruity v earthy’ spectrum – think Pinot Noir from California’s Lodi v Burgundy’s Givry, very generally speaking. Most people will actually have a definite preference here. That their preference will usually encompass a range rather than a single point on a scale also needs to be considered. After all, many of us who like fruitier reds will consider a wine on the earthier side if it has some other appealing traits.
It is necessarily true that much of why we like this or that wine is subjective. I might perceive a particular wine as sweet even though its residual sugar is low. And this would seem to fly in the face of any computable solution. But there are ways to literally objectify the subjective. Public polling is exactly this. I’d also caution against the industry’s current fetish for false precision. Fewer but better-quality variables that embrace subjectivity will yield better wine recommendations.
There will always be people who don’t fit the existing profiles. And for anyone who truly loves wine, this is to be celebrated! But it also points to the most promising future for AI in wine – specifically, automation and machine learning as complements to the good work vintners do every day, not replacements for them. A good part of what we do is actually a repeated routine and therefore automatable. Leveraging pattern-recognition and hypothesis-generation from machine learning frees up vintner time for sanity-checking and discretion. This is where the magic happens, and AI is decades away from mastering this, if it’s even attainable.
Obviously, there’s a lot more to this topic than I’ve described above – the critical role of ongoing consumer feedback being but one. But ultimately, there is a logic to people’s wine preferences. And where there’s logic in a system, technology can be usefully leveraged. We just need to keep the data-set simple – not ‘big data’ – embrace the subjective and leverage the inimitable value of wine professionals. This is the key to creating a potentially very useful future for AI in wine.
We’re only just out of the gate on this exciting journey. Let’s not let well-meaning but misguided negativity spoil the party before it has even started.
