Wine Conversations: Can AI Replace Somms? (I)

Our thesis: In a "field" that is so deeply rooted in the human sensory experience, AI can never really take the place of the proverbial primate.

This week, in lieu of our “standard issue” Tannic Panic! we are kicking off the “Wine Conversation” of the month with a topic that is increasingly on the minds of literally everyone in existence (we checked). AI.

With its rise in our society, both in popularity and capability, AI will undoubtedly continue to weasel its merry little way deeper and deeper into the world of wine. This month’s wine conversation is an “exploration” of just what this might look like, and begs the question – CAN AI REPLACE THE “HUMBLE” SOMM?

Our thesis: In a field that is so deeply rooted in the human sensory experience, AI can never really take the place of the proverbial primate.

---

Before we dive into our rambling ideas on the topic, let’s break down what AI does and doesn’t do well in basic terms (THE TRUTH MAY SHOCK YOU):

• What AI models (or LLMs) are good at: Analyzing and extrapolating from huge amounts of data

• What AI is bad at: Having senses and actually experiencing things, and understanding the nuances of shared or distinct tastes between people

Sensory Experience

LLMs can’t tell you what a particular wine actually tastes like. They can often tell you what specific critics who have tasted the wine have said about it, or what the producers have said about it, or they can give an averaged out generalization of notes that have been publicly released. They can also make “educated” guesses based on data it has scraped on things like varietals, vintages, producers, regions. In a way it would be like a “well-educated” tee-totaler (THE BAD KIND!) making recommendations about things they’ve never truly experienced, only this tee-totaler has never experienced the sensation of taste or smell. In other words, they are 100% reliant on the human experience of people who actually know what the experience is like.

If you wanted recommendation about food and wine pairings for Pinot Noir, AI might give generalized suggestions for pairings, like dishes that contain “mushrooms,” but it wouldn't be able to taste its way to the perfect bottle for a specific mushroom-containing dish unless it was extrapolating based on wine pairings humans had already put online. Could it get creative? Well, you could ask it to, but what ultimately would result from that would again be either repurposed information from human sources, or possibly something that actually sounds creative that it has no basis for testing or understanding on the most basic sensory level. That said, this doesn’t invalidate LLMs as a resource for general recommendations, the point is just that those recommendations are inherently far more generic and literally untested by human ‘buds (unless it’s a real opinion stolen from someone else). You might argue that this makes them useful, but it proves the underlying idea that without its own sensory experiences (give that a few more years) it requires ours to operate.

For many of us, some of the most “profound” wine tasting experiences “we’ve” had have been at restaurants where the somm concocted unconventional or unique recommendations or curated pairings that you could never arrive at by asking an AI what to pair with something.

How can an AI think of something it hasn’t been trained on or that hasn’t been discussed in public forums?

How can AI take a risk on tasting a little known bottle to determine if it just might be that magic pairing for a dish the chef just invented? A dish, we might also add, THAT AI HAS ALSO NEVER TASTED! And never will... (at least for a few more years).

When it comes down to it, these models are trained on human data say so they can’t really exist without the other, and while human data may be on the way out as these models advance, you can’t replace human data on sensory experience. Additionally, describing wines well (a topic recently covered by Tom Wark) and figuring out how to accommodate different sensory challenges like personalized pairings based on individual preferences and an ever-changing landscape of wine (both in terms of new products coming out, and old ones evolving) isn’t something that a model built on a sea of data can do without people.

Social Culture

Wine is also deeply embedded in social culture and a big part of what makes going wine tasting or enjoying bottles at a nice restaurant or with friends is interacting with each other, and hearing the “stories” of the people that make the wine. Do we want to replace this with Cyborgs?

In that sense, somms are about as at risk of losing their jobs as bartenders. People don’t want to interact with robots or machines exclusively, things like that have been attempted and they slot their way into the realm of novelty. Like entirely robotized bars or cafes.

Blind Tasting

Ah yes, on the subject of “things that aren’t impressive or useful” - AI is getting better and better at blind “tasting.” Here’s some bullets on that:

• This article talks about AI competing against human tasters in a competition. I’m sorry, but a resounding WHO CARES - what does it matter if a machine can determine what a wine is by scent? The literal only thing that makes blind tasting wines that have been curated impressive or cool is the human element. You might as well tell me that the label on the bottle blind identified the wine because that’s equally relevant. Blind tasting isn’t something done because it’s useful in real world applications, it’s more of a technical ability that can wow people and has its own level of inherent awesomeness because it pushes the limits of human capability. It’s not like writing or other art forms where the invasion of AI threatens livelihoods or diminishes the artform. At least not how we see it.

• This article by Jason Wilson (who also has a substack) talks about how an AI has been used to blind identify with 100% accuracy specific wines from vineyards and vintages, which is cool in the sense that it could potentially have applications in preventing wine counterfeiting (though it’s nascent and potentially prone to unforeseen pitfalls)

SO where does AI come in handy?

It would be unfair and “journalistically” irresponsible not to examine the value in the tool, especially in the context of its growing presence in all of our lives. So let’s talk about the areas that it can and will step in, and can be a good resource

Wine education

Though search engines are jumping on the implementation of these types of models to aid in search, one of the biggest impacts people immediately saw with the rise of Chat GPT (and pals) was that we could now type a casually worded question out and it would understand exactly what we wanted to know. And it would tell us.

You can quickly get answers and information about wine bottles, regions, vintages, production techniques, technical definitions — you name it — but as with all of the current and up date information available, it, of course, still has to be fed in. It can be a majorly useful resource, as long as all information gathered from it is fact checked and cross referenced (ideally against first hand sources of that key information). Some things are more static, like the geographical features of an area, or the scientific names for wine producing vines and French oak. Static information doesn’t actively require human contribution, so you could argue that the human element is quickly outmoded on a large volume of information once its been ingested.

But with the ever changing human elements (like who is working a vineyard, what production techniques are being used, the blend that goes into a bottle) and the inconsistent things like weather and increasingly, climate, long standing norms or averages of old data sources can’t provide reliable information without active human contribution.

How else can we use AI to our benefit?

“We all” have certain “descriptors” that we tend to use repeatedly when blind tasting wines and we try to “use” these to piece together the identity of the wine. In theory, a model that can quickly find patterns in the datapoints we produce could help train our own palates but identifying key trends.

The concept: AI-assisted blind tasting uses large language models (LLMs) trained on wine tasting notes to help tasters more accurately identify wines. By analyzing the specific descriptors—like “blackcurrant” “garrigue,” or “graphite”—that tasters use, AI can learn to associate certain words or combinations with particular wine styles, grapes, or regions. For instance, frequent mentions of “raspberry, strawberry, white pepper, and garrigue” might signal a Grenache-based Southern Rhône, while “blackcurrant, graphite, green pepper, and chocolate” could point to a Left Bank Bordeaux. Since this can be trained on a single person’s data, it could help us shed some biases that prevent us from using information we already have to our benefit.

In other words, tasters can receive feedback on which descriptors in their notes most strongly predict specific wines, helping them refine their sensory vocabulary and deduction skills. Over time, AI can reveal personal tasting patterns and biases, making the blind tasting process more systematic and insightful. This approach could not only sharpen the taster’s accuracy but also transform blind tasting into a more data-driven and “personalized” “learning experience.”

POTENTIAL “REAL WORLD APPLICATION” OF SAID CONCEPT

Have you ever watched blind tasting videos on the internet (you know, AS ONE DOES) and found that you accurately guess the wine based on the descriptors the taster generates, but then watch as the person guessing goes off the rails? Why is that?

The SuperVinoBros are “twin brothers,” Chris and Ryan Goydos, who have generated a social media platform dedicated to “wine content.” They regularly share “blind tasting” videos, where one brother attempts to identify the grape variety and region of a wine poured by the other, starting by describing the characteristics of the wine in detail (color, aroma/flavor profile, structure etc). This presents an opportunity for “the viewer” (LIKE US!) to guess the identity of the wine solely based on their tasting notes. Interestingly, we often find ourselves correctly guessing the wines based on their descriptions alone, as they wildly guess something completely wrong. Aside from illustrating our own GENIUS, this demonstrates that if “we” remove the human element from the tasting and just “looked at the data” (aroma/flavor descriptors, color description and wine profile), the Super Vino Bros might have a higher success rates in correctly identifying the wines. The fact that their success rate (as far as we know) is ~0% might mean it’s just a gimmick, but their tasting notes are good, so this is a case where an AI assistant might come in handy if they’re not FAKING IT.

---

Anyway, these ramblings have only just scratched the surface on this topic. We are very eager to see what the rest of our compatriots have to say on the topic, and, of course, YOU “dear reader.” Have “thoughts” of your own? Feel strongly about wine or AI or somms? Hate everything about us? Let us know! Leave a comment!

Until next time, HAPPY DRINKING PEOPLE.

Cheers!

Isaac & Zach