Column Editor’s Note: Throughout history, wine has been a blend of art, science, and sensory delight—an ancient craft steeped in tradition. Today, the addition of AI adds a fascinating modern twist, offering precision and personalization like never before. In this Recreations in Randomness column article, Jing Cao uncorks how AI is reshaping wine evaluation, pouring new possibilities into the consumer experience.
Keywords: artificial intelligence (AI), wine quality evaluations, wine recommendation systems
Winemaking and human civilization have been intertwined for millennia. The earliest evidence of winemaking dates back to around 6,000 BC in the territory of the modern-day country of Georgia. This suggests that winemaking began shortly after humans started cultivating grapes for consumption, making it one of humanity's oldest and most enduring practices. Through time and space, wine has seamlessly integrated into various aspects of modern life, from celebratory clinking glasses to the quiet contemplation of a solo sip, this fermented beverage transcends mere refreshment. It is a cultural touchstone, a culinary companion, and a global industry generating billions annually.
Winemaking and wine consumption are deeply rooted in tradition and heritage. Many regions around the world boast centuries-old winemaking traditions, passed down through generations as a form of artistic expression. Yet, each technological advancement has revolutionized the wine industry. For instance, the internet, in less than 3 decades, has transformed how winemakers share information, consumers learn about and purchase wines, and the industry markets its products. Furthermore, in recent years, the rapid development of artificial intelligence (AI) has been revolutionizing the wine industry. An interdisciplinary science combining computer science, statistics, and domain-specific knowledge, AI enables machines to mimic human intelligence to solve complex problems and automate decision-making. This powerful technology has been and will likely continue transforming both wine production and consumer behavior.
Before consumers can savor the fruits of a winemaker's labor, successful production relies on careful vineyard management and a meticulous winemaking process. Let us start with the beginning of the process. Grape production is highly sensitive to changing environmental conditions, and viticulture is significantly impacted by climate change. In recent years, advances in satellite or drone remote sensing, field sensors, robotic technologies, and automated weather stations have generated vast amounts of data to support AI algorithms in pattern recognition and make recommendations for precision viticulture (Mohimont et al., 2022). In the subsequent winemaking stage, a sophisticated microbial process, AI systems can enhance quality control and optimize fermentation. These systems analyze the chemical and sensory attributes of wines and recommend ideal conditions. For example, Napa Valley’s Palmaz Vineyards has adopted the Fermentation Intelligence Logic Control System (FILCS), a fully algorithmic fermentation control system. FILCS continuously monitors the fermentation tank environment, leveraging its measurements and historical data to predict the likelihood of undesirable outcomes. It then proactively alerts the winemaker when necessary.
While AI has demonstrated transformative potential in vineyard management and winemaking, this article focuses on AI’s impact on the wine consumer experience. Specifically, we will explore how AI can revolutionize wine consumption through wine quality evaluation, personalized recommendations, and the enhancement of wine-tasting experiences. For example, AI-powered virtual sommeliers can provide personalized wine-tasting experiences, offering tailored information on flavor profiles, food pairings, and historical context.
Once wines are produced, wine quality evaluation and certification carry essential importance to both consumers and wine producers and sellers. Traditionally, wine quality evaluation involves human experts to conduct wine tasting, which is a meticulous process that relies on a combination of senses, experience, and established protocols. The sensory evaluation includes three stages. The first is sight: experts examine the wine's color, clarity, and viscosity, looking for any abnormalities. The second is aroma, which is a crucial stage. The taster swirls the wine to release its aromas, identifying specific compounds and their intensity. They might describe notes of fruits, flowers, spices, or even earthiness. The third is taste. Small sips are taken, allowing the taster to assess the flavors across the tongue. Balance between acidity, sweetness, tannins (in red wines), and alcohol is a key factor. Mouthfeel, the texture and weight of the wine, is also evaluated. Formal tastings often follow a structured format. Wines are tasted blind (without knowing labels) to minimize bias. Standardized scoring systems are used to evaluate various aspects and arrive at a quality score.
Previously, wine quality evaluation, which usually involves human experts to conduct wine tasting, can be time-consuming, expensive, and unreliable. One of the most influential studies conducted to show this was the one by Richard Hodgson, a winemaker and retired statistics professor. He was curious about why some of his own wines would do really well in some places (i.e., winning a gold medal) and not do anything in others. As a data person, he designed an excellent experiment conducted from 2005 to 2009 at the California State Fair Commercial Wine Competition. With the assistance from the program manager, Hodgson placed three samples poured from the same wine bottle (displayed with different IDs) in one flight of judged wines to see if the judges ranked each sample consistently. Though it is expected there is a considerable amount of subjectiveness in wine tasting, the result from Hodgson’s analysis is still shocking (Hodgson, 2008, 2009). Hodgson found that “Fewer than 10% of judges could judge the three wines and maintain consistency in their scores” (2008, p. 106) and “Some of the same wines received ratings that ranged from no award (the worst category) to gold (the best category)” (2009, p. 5). When the study, “An Examination of Judge Reliability at a Major U.S. Wine Competition” (2008) was published in the Journal of Wine Economics, it received significant media attention and created a stir among wine judges and within the wine industry.
The results of the study are partly due to the fact that our knowledge of the physiological basis for the taste system is considerably less developed (Brody, 2012). In terms of wine tasting, we should understand that wine appreciation is inherently subjective. What one person finds delicious, another might find unpleasant. This makes it tough to have a universally agreed-upon standard for ‘good’ wine. Wine competitions often involve tasting dozens of wines in a single day. By the end, judges can be fatigued and their palates dulled, affecting their ability to accurately assess each wine. Things like what a judge ate for breakfast, the order of wines to be tasted, or even the room temperature can influence how they perceive a wine’s taste and aroma.
Wine assessment can take various forms, notably, objective quality assessment and subjective preference evaluation. Objective assessment primarily focuses on identifying and distinguishing between different types of wines (e.g., Champagne vs. Prosecco, Cabernet Sauvignon vs. Bordeaux). This approach centers on measurable and quantifiable characteristics such as chemical composition, origin, and production methods, aiming for a neutral and unbiased assessment based on factual information. In contrast, subjective wine assessment evaluates a wine’s overall quality and its ability to satisfy the palate. This approach relies heavily on personal sensory experiences and individual preferences. The combination of objective and subjective assessments ultimately contributes to tasting scores, reviews, and personal opinions about a wine’s quality and enjoyability.
In recent years, different types of data related to wine quality evaluation have become available, including physicochemical test measurements, sensory test scores from human experts, and their unstructured tasting notes. Wine quality evaluation with AI leverages rich data mines to complement the traditional methods with a layer of data-driven insights and predictive capabilities. Specifically, AI can analyze vast data sets encompassing everything from grape chemistry and vineyard conditions to historical weather patterns and fermentation processes. This allows for identifying patterns and relationships that might be missed by human evaluation alone.
For example, Tiwari et al. (2022) claimed that wine quality, which is often perceived as a subjective and qualitative measure, could be related to the wine chemistry parameters through mathematical and machine learning (ML) approaches. The authors used wines’ chemical and physiochemical compound measurements to train feedforward deep neural networks (FNNs) to predict wine quality. FNNs are powerful AI tools for various tasks because they can learn and represent complex patterns in data. The universal approximation theorem states that a sufficiently large FNN with one hidden layer can approximate any continuous function to arbitrary accuracy. With multiple hidden layers (i.e., deep neural networks), FNNs can learn even more intricate mappings between inputs and outputs. The study shows that the predicted wine quality indices are in good agreement with the wine experts’ perceived quality ratings.
With the advancement of natural language processing, the previously underexplored wine review documents have become a rich data source in wineinformatics, which is a new data science application with wine as the domain knowledge and different types of wine-related data sets as data sources (Chen et al., 2018). Chen and Cao (2024) proposed an interpretable sentiment analysis framework to provide word-specific sentiment measures for wine review documents. They found that wine review vocabulary has its own words to express sentiment, which can be different from everyday language. For example, ‘stained’ in wine reviews implies a deeper and more complex flavor profile, which is typically considered to be of high quality. Similarly, ‘carpet’ and ‘fabric’ are associated with wines having a rich velvety texture that is often indicative of high-quality wines. On the other hand, words such as ‘quick,’ ‘breezy,’ and ‘easygoing’ indicate a lack of sophistication and depth, making them descriptors for less desirable wines. Bender et al. (2024) constructed WineSensed, a large multimodal wine data set that consists of visual perception, language, and flavor information. The authors then trained an ML algorithm to generate concept embedding across different data modalities. The authors claimed that the ML algorithm effectively combines data from multimodal sources, which in this case includes wine labels, wine reviews, and wine tastings, and it is capable of making more accurate predictions on wine quality than when it only uses separate data sources.
Wine reviews or tasting notes constitute a specialized genre in oenology, with their own rhetoric and language. Crafting professional wine reviews has been a distinguished accomplishment for wine experts. Recent research shows that AI algorithms are capable of generating wine reviews without having to open a bottle. Using approximately 125,000 reviews published on the website of Wine Enthusiast from 1996 to 2016, Carlson et al. (2023) trained a transformer-based deep learning neural net architecture through the linkage of review text to a set of features such as winery name, alcohol percentage, and price, to learn the general structure and style of wine reviews. Note that transformers are a revolutionary AI technique that excel at understanding context. When translating a word, humans consider not just the word itself but also the surrounding words. Similarly, transformers analyze the context of a word within a sentence, allowing them to generate more coherent and relevant text. Upon training, the model developed by Carlson et al. is able to write (generate) new reviews given unseen wine features. The algorithm passed a kind of Turing test, which means that the machine-writing reviews are largely indistinguishable from those written by wine experts.
Among all the implementations of AI in wine industry, maybe the biggest impact comes from AI recommendation systems for wines. Wines belong to the category of experience goods whose quality cannot be assessed prior to consumption (Nelson, 1970). When it comes to choosing wines, consumers often face a daunting task of figuring out what they prefer, where wine-related factors such as taste, mouthfeel, aroma, and color will blend simultaneously with contextual factors such as environmental, cultural, and psychological considerations to create a holistic tasting experience (Pedroza & Herrell, 2022). Traditionally, wine consumers had relied heavily on wine experts or wine sommeliers for their recommendations. However, the tasting experience is subjective and highly personal. A highly praised wine may not be appreciated by all consumers. It is the subjective nature of wine quality assessment and the lack of consensus regarding quality information that contribute to distrust among wine customers regarding wine price, quality, and experts’ opinions.
In the not-too-distant future, many consumers may choose to put the task of choosing wines in the hands of a wine AI recommendation system (AI-RS). In general, a recommendation system is a system that produces individualized recommendations as output or has the effect of guiding the user in a personalized way to interesting or useful objects in a large space of possible options (Burke, 2002). The more the data AI can access, and the more efficient the algorithm, the more accurate the AI-RS delivers. This is because the ultimate goal of AI-RS is to recognize patterns hidden in vast amounts of data—something that can be challenging for people to tease out but is feasible for the AI-RS to accomplish. In this case, a wine AI-RS can effectively use different types of information, such as wine experts’ evaluations, feedback from other consumers, and the user’s demographic information and wine-purchasing history, to come up with a recommendation tailored just for the user.
Such wine AI-RS apps are already available on the market. For example, Vivino, the world’s largest wine marketplace, launched its AI wine recommendation app Match for You in 2021. The algorithm takes advantage of Vivino’s more than 200 million wine ratings and reviews. Once a customer has rated at least five wines, the algorithm will begin to identify the customer’s unique taste profile. It will provide a match score on all 13 million wines (almost all the wines on the market) in its database, customized for the customer. Another AI-RS app, DRINKS Recommend, was introduced by DRINKS, a company that specializes in alcohol technology and e-commerce enablement. The app was built on DRINKS’s patented AI models, called PAIR (i.e., Predictive AI Retailing), which incorporates wine prices, regions, varietals, tasting notes, descriptions, label features, and how a product is perceived by customers to develop recommendations tailored to online merchants and their customers. The app was deployed on select pages by two major online wine retailers, bringing in an excess sale of 250,000 bottles a month.
Even an ordinary ChatGPT chatbot can play the role of your personal AI sommelier. According to a study published by OpenAI (2023), the newest version of their product has already passed the Introductory, Certified, and Advanced Sommelier theory exams—without ever taking a sip. (The Certified Sommelier Examination is a rigorous assessment conducted by the Court of Master Sommeliers. It evaluates a candidate’s proficiency in several key areas related to wine and beverage service.) Just for experimenting purposes, we asked ChatGPT this question: “I am celebrating my 20th wedding anniversary, I like red wine, price range $20–30, mouthfeel not too dry. What do you recommend?” The following is what it returns; the readers can judge whether ChatGPT has done a good job.
Bread & Butter Pinot Noir:
Price: Around $15
Description: This Pinot Noir offers a smooth and approachable mouthfeel with flavors of ripe red berries, cherries, and a touch of spice. It’s not overly dry, making it a great choice for celebrating without breaking the bank.
Josh Cellars Cabernet Sauvignon:
Price: Approximately $16
Description: Josh Cellars’ Cabernet Sauvignon is rich and fruit-forward, with notes of blackberries, plum, and a hint of vanilla. The tannins are well-balanced, providing a satisfying mouthfeel without being overly dry. It pairs wonderfully with hearty dishes and cheese platters.
One Hope California Cabernet Sauvignon:
Price: About $19
Description: This Cabernet Sauvignon from California boasts flavors of dark cherries, blackcurrants, and a touch of cocoa. It has a velvety texture and a medium body, making it a crowd-pleaser for your anniversary celebration.
In summary, AI will probably play an increasingly important role in shaping the future of the wine industry, from adopting robots, drones, and sensors in vineyards to better manage grape growth and navigating issues caused by climate change, to employing AI-powered virtual wine critics and judges and relying on AI-powered virtual sommeliers that may know the customer well enough to recommend the perfect wine. Throughout millennia, humans have woven an intricate tapestry of winemaking and wine consumption. This ancient art form, steeped in tradition, blends professional expertise with an alluring aura of secrecy and romance, leaving an indelible personal imprint in the whole process. Should we relinquish the control of making wine, choosing wine, and tasting wine to AI or just put AI in the supporting role? We may argue, for example, that we should not rely too much on AI wine recommendation tools, for we might enjoy using our own eyes, noses, and mouths to compare, smell, and taste wines, and enjoy being surprised or even disappointed by the choices, which constitutes the whole experience of wine consumption. Surely, AI can render great help in some aspects, but some areas of experience might be best left to ourselves.
The author would like to thank the AE and two reviewers for their constructive suggestions.
The author has no financial or nonfinancial disclosures to share for this article.
Bender, T.,Sørensen, M. S., Kashani, A., Hjorleifsson, K. E., Hyldig, G., Hauberg, S., Belongie, S., & Warburg, F. (2024). Learning to taste: A multimodal wine dataset. ArXiv. https://doi.org/10.48550/arXiv.2308.16900
Brody, H. (2012). Editorial: Taste. Nature, 486(S1). https://doi.org/10.1038/486S1a
Burke, R. (2002). Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12, 331–370. https://doi.org/10.1023/A:1021240730564
Carlson, K., Kopalle, P. K., Riddell, A., Rockmore, D., & Vana, P. (2023). Complementing human effort in online reviews: A deep learning approach to automatic content generation and review synthesis. International Journal of Research in Marketing, 40(1), 54–74. https://doi.org/10.1016/j.ijresmar.2022.02.004
Chen, B., Velchev, V., Palmer, J., & Atkison, T. (2018).Wineinformatics: A quantitative analysis of wine reviewers. Fermentation, 4(4), Article 82. https://doi.org/10.3390/fermentation4040082
Chen, C., & Cao, J. (2024). Interpretable sentiment analysis using the attention-based multiple instance classification model: An application to wine reviews [Manuscript submitted for publication].
Hodgson, R. T. (2008). An examination of judge reliability at a major U.S. wine competition. Journal of Wine Economics, 3(2), 105–113. https://doi.org/10.1017/S1931436100001152
Hodgson, R. T. (2009). An analysis of the concordance among 13 U.S. wine competitions. Journal of Wine Economics, 4(1), 1–9. https://doi.org/10.1017/S1931436100000638
Mohimont, L., Alin, F., Rondeau, M., Gaveau, N., & Steffenel, L. A. (2022). Computer vision and deep learning for precision viticulture. Agronomy, 12(10), Article 2463. https://doi.org/10.3390/agronomy12102463
Nelson, P. (1970). Information and consumer behavior. Journal of Political Economy, 78(2), 311–329. http://www.jstor.org/stable/1830691
OpenAI. (2023). GPT-4 technical report. ArXiv. https://doi.org/10.48550/arXiv.2303.08774
Pedroza, M. A., & Herrell, R. (2022). Exploring emotions as a new quality parameter in wine. Wine Business Journal, 5(1), 6–26. https://doi.org/10.26813/001c.31663
Tiwari, P., Bhardwaj, P., Somin, S., Parr, W.V., Harrison, R., & Kulasiri, D. (2022). Understanding quality of pinot noir wine: Can modelling and machine learning pave the way? Foods, 1(19), Article3072. https://doi.org/10.3390/foods11193072
©2025 Jing Cao. This article is licensed under a Creative Commons Attribution (CC BY 4.0) International license, except where otherwise indicated with respect to particular material included in the article.