At the UK’s AI Safety Summit in 2023, Elon Musk told British prime minister Rishi Sunak that artificial intelligence (AI) was likely to be ‘the most disruptive force in history’. And if you look around you, it isn’t hard to see signs of that disruption already taking place.
We’re eagerly talking away to AI-powered digital assistants such as Siri and Alexa, university lecturers are complaining about students using AI chatbots to write their essays, and fears are growing that a range of jobs – from contract lawyer and computer coder to accountant and even teacher – will either cease to exist or change beyond recognition. Engineering designers are using AI to create new products; architects are using it to design outlandish new buildings and medical researchers are using it to cure disease.
What is AI?
Anthony Aubert with old schoolfriend and now business partner Jean-Charles Mathieu, of Aubert & Mathieu
Artificial intelligence is an umbrella term for technologies that enable computers to behave as though they are intelligent, using complex algorithms to ‘learn’ how best to process data in order to provide the answer that the user requires. There are different types of AI, but the form that has captured the zeitgeist is known as generative AI, which, as its name suggests, is capable of generating complex outputs such as computer code, paragraphs of text, images or video after being given prompts from a user.
One specific application of generative AI is what’s known as a large language model – an AI that has been trained to learn the patterns and relationships between words and phrases using an enormous language-based dataset, and to use that knowledge to converse with the user. Google’s Gemini (known as Bard until early February 2024) and OpenAI’s ChatGPT are two of the best- known large language model-based AI platforms.
It’s true that there are broader concerns globally and much debate regarding the direction of travel of AI as a whole and its potentially serious impact on vital areas such as employment, privacy, data storage and misinformation. Indeed, ChatGPT’s maker Sam Altman is rarely seen without his infamous blue backpack – there is speculation that it holds a laptop on which are the codes necessary to shut down the ChatGPT AI servers if it should ever run out of control.
But what do such wider issues have to do with wine? AI has already infiltrated many aspects of both viticulture and winemaking. It’s already at home in vineyards, helping to predict irrigation requirements, pick fruit, prune vines and detect vine diseases. And in cellars, winemakers are able to use AI-powered systems to monitor and adjust fermentation conditions in real time.
Languedoc producer Aubert & Mathieu recently took things a step further by producing a wine under the instruction of ChatGPT, right down to the name and label. ‘We are a new brand, only five years old, and it’s in our roots to be adventurous and try new things,’ says co-owner Anthony Aubert.
‘We wanted to do something with ChatGPT, because people are talking about it. I opened an account and said, “Let’s make a wine in the south of France”, telling it that we had organic Syrah and Grenache grapes available’. The AI offered basic winemaking advice, which was adjusted by a human winemaker, as well as suggestions for the wine’s name, label design and marketing plan, including presenting it in a Burgundy-style bottle ‘because it’s more prestigious’.
The result was a wine the AI chose somewhat ominously to name ‘The end’, which Aubert suggests could be read as a tongue-in-cheek take on robots putting an end to the roles of humans, something that he feels is unlikely to happen. ‘It will never replace winemakers or marketeers, but you have to know how to use it,’ he says. The price range suggested by the AI was thought to be too high, and the wine is now available on their website aubertetmathieu.com at €29.90.
Wine for the masses
Jean-Marc and Eliane Lafage, owners of Domaine Lafage in Roussillon. Credit: MAXPPP / Alamy Stock Photo
This was an interesting if limited experiment by a small producer, but what if larger companies decide to take the advice of a chatbot that’s capable of amalgamating the views of millions of wine lovers worldwide in seconds to suggest a wine that, it thinks, will please the masses? A world awash with bland Chardonnay, Sauvignon Blanc and Cabernet Sauvignon?
‘The risk is of wines that are linear, one dimensional, without character,’ says Jean-Marc Lafage, owner and winemaker at Domaine Lafage in Roussillon. ‘The result would be an abundance of wine that all tastes the same, which reduces the appeal and turns wine into a commodity without a soul or a sense of place. Human intervention and creativity will remain the key.’
Dr Laura Catena, fourth-generation winemaker at Catena Zapata in Argentina, agrees: ‘Wine isn’t made to a recipe. Each vintage and place behaves differently every year. Wine is alive because of terroir and vintage variations and because of the live yeasts and microbes involved. I don’t think a machine would be able to handle all the variables to make the right decision, at least not yet. Maybe in the future a machine could help make decisions, but I don’t find that appealing as a wine consumer.’
AI & terroir
A recent AI-powered study provided scientific evidence for the elusive concept of terroir. As reported on decanter.com in December 2023, scientists at the University of Geneva used AI to ‘summarise’ the so-called chromatograms – the full chemical breakdowns of a substance, which in a wine’s case can comprise up to 30,000 data points – of 80 red wines from seven Bordeaux estates in 12 vintages. The AI summaries consisted of simple X and Y coordinates that, when placed on a graph, coalesced into seven ‘clouds’ of points that represented the seven estates. Crucially, the seven clouds themselves formed two groups that correspond to the right and left banks of the Garonne, implying that the terroir of the two regions imposed an indelible chemical signature on the wines.
The fact that the AI-driven analysis was able to predict with 100% reliability which château a wine came from suggests that its potential to combat wine fraud is huge. Professor Alexandre Pouget, who led the research, suggested that the technique could also be used to monitor quality during the winemaking process and as a useful tool during the blending stage.
Personal service
In fact, it’s consumers who are in the crosshairs as an increasing number of companies look for ways to use AI to influence our wine-buying habits and choices.
Currently, most wine recommendations provided by apps and online retailers are generated using crowd-sourced opinions – you liked that so, based on other people’s buying habits, you’ll probably like this. ‘AI can provide opportunities to elevate those features,’ says Michael Waring, head of design and user experience at UK online retailer Naked Wines, which has been working with recommendation algorithms for some time. ‘For example, to find patterns in data to learn more about what a customer likes and will like as time goes on, in order to improve our recommendations and match wines to people’s individual tastes.’
This is the approach taken by Preferabli, an app powered by a huge database of wine characteristics (anywhere between 500 and 800 per wine, the company says) captured by experts, including several Masters of Wine and Master Sommeliers (reportedly a dozen of them). So far, the service has more than a million wines listed, including Château d’Yquem going back to World War I.
Users enter a wine and are given a selection of others that Preferabli’s generative AI software, which is protected by eight US patents, thinks that they will like. These recommendations aren’t automated responses based on other people’s opinions, but AI-driven learning about the user’s personal tastes, which it uses to build an individual, one-to-one profile.
‘We take what you know, however much or little, as a point of departure for your discovery journey,’ explains Preferabli co-founder and CEO Pam Dillon. ‘For consumers, it’s kind of like The Wizard of Oz, where you’re black and white for most of your life and then one day you see colour.
‘In the beginning, people didn’t understand why we were doing this,’ adds Dillon. ‘They thought we were nuts. Why are you evaluating every wine and spirit in the world? Why are you building algorithms that are so deep? But we could see something coming. They didn’t understand how powerful AI could be, or that AI could truly create discovery and that there were ways of using data to bring wine and spirits alive.’ Richard Bampfield MW is among the experts who have lent their palates to Preferabli. Could such apps mean that we will stop turning to wine merchants, wine critics or sommeliers for advice on which wines to buy? ‘I don’t see it like that, because I’m not sure that’s the way all humans work,’ says Bampfield. ‘However, I think some people will be very happy if the interaction with someone is taken out of the equation. Wine can be incredibly complex, so something like Preferabli can be helpful. It gives people another option.’
Pam Dillon, Preferabli
The virtual sommelier
US-based app Tastry is ploughing a similar furrow; however, rather than relying on human palates to characterise wines, it uses chemical analysis to break them down into a cloud of more than a million data points. It then uses machine learning algorithms to ‘decode’ the wines’ ‘flavour and aroma matrices’ – the way that the chemical components interact to influence our perception of the way a wine tastes and smells.
Drawing on a database of the food and wine preferences of about 100,000 consumer palates (as well as more than 200 million AI-generated ‘virtual’ palates), and the results of a short survey of the user’s food preferences, its algorithms then predict the user’s individual preferences and suggests wines accordingly.
On a smaller scale, there are also wine-related custom versions of ChatGPT that have been created by users, including GrapeGPT and Elizabeth, Master Sommelier, which can chat to you about wine and make recommendations based on your preferences. The latter was created by sommifyAI, a Helsinki-based company that’s marketing its own AI sommelier based on the knowledge of Julie Dupouy-Young, who won the ASI Best Sommelier of Ireland competition four times between 2009 and 2018.
Viva la revolución
Then there’s the written word. ChatGPT has been trained using a significant portion of anything that has ever been written about wine, potentially making it a powerful tool for wine education. Last year, OpenAI announced that GPT-4, the latest iteration of the AI, had used that accumulated knowledge to achieve pass marks in the three theory exams that Master Sommelier candidates must pass in order to qualify.
But for now, there are clearly limits to its abilities. ‘I tried asking ChatGPT to write some tasting notes and they were absolutely horrible, a complete waste of time,’ says Catena. ‘Words like “enticing” and “voluptuous” making my precious, terroir-based wines sound like sex toys.’
Those limitations are likely to have a short shelf life. We’re currently in the early days of an AI-driven revolution. A wide range of industries, including the wine industry, are racing to keep up with the rapid acceleration in AI’s capabilities, looking for new ways to utilise and benefit from this powerful new tool. What impact it will have on the world of wine is impossible to predict at this early stage, but it’s a safe bet that things will never be the same again.
