The AI ​​sommelier writes wine reviews without ever taking the stopper of a bottle

In the world of wine reviews, it is important to write visually. Just think about this: ‘While the nose is slightly closed, the palate of this non-dry Riesling is filled with juicy white grapefruit and mandarin flavor. It is not a deeply concentrated wine, but it is nicely balanced by a touch of lemon-lime acid that lingers.

When you read the description, you can almost feel the cool glass of sweat in your hand and taste a citrus burst on the tongue. But the author of this review never had that experience – because the author was a piece of software.

An interdisciplinary group of researchers developed an algorithm capable of writing reviews about wine and beer that is usually indistinguishable from a human critic. The researchers recently published their findings in International Journal of Research in Marketing

The team hopes this program will be able to help beer and wine producers gather a large number of reviews or provide human reviewers with a template from which to work. The researchers say their approach could even be extended to reviews of other “experience products” such as coffee or cars. However, some experts warn that these types of applications can be exploited.

Theoretically, the algorithm could have produced reviews on anything. But a few key features made beer and wine particularly interesting to researchers. First, “it was just a very unique dataset,” says computer engineer Keith Carlson of Dartmouth College, who helped develop the algorithm used in the study. Wine and beer reviews are also a good template for AI-generated text, he explains, because their descriptions contain many specific variables, such as region, grape, variety, fermentation style, and production year. These reviews also tend to rely on a limited vocabulary. “People talk the same way about wine and use the same words,” Carlson says. For example, connoisseurs routinely toss around adjectives like “oak,” “floral,” or “dry.”

Carlson and his co-authors trained their program with a decade of professional reviews – about 125,000 in total – from the magazine Wine enthusiast† They also used nearly 143,000 beer reviews from RateBeer’s website. The algorithm processed these man-made analyzes to learn the general structure and style of a review.

To generate its own assessments, AI was given the details of a particular wine or beer, such as the name, style, alcohol content and price of the winery or brewery. Based on these parameters, AI found existing reviews for that beverage, extracted the most commonly used adjectives, and used them to make its own description.

To test the program’s performance, team members selected one human and one AI-generated rating for each of 300 different wines and 10 human ratings and one AI rating for each 69 beers. They then asked a group of human subjects to read both the machine-generated and man-written reviews and checked whether the subjects could distinguish what was what. In most cases, they could not. “We were a little surprised,” Carlson says.

Although the algorithm seemed to be good at gathering many reviews and summarizing them in a single, coherent description, it has some significant limitations. For example, the algorithm may not be able to predict the taste profile of a beverage that has not been tasted by human taste buds and described by human authors. “The model can not taste wine or beer,” said Praveen Kopalle, marketing specialist at Dartmouth and co-author of the study. “It only understands binary 0s and 1s.” Kopalle adds that his team would like to test the predictive power of the algorithm in the future – by letting it guess what an unclassified wine would taste like, and then comparing its description with the description of a human reviewer. But so far, human reviewers, at least in the beer and wine world, are still essential.

AI-generating language is not new and similar software has already been used to make recommendations for online review platforms. However, some sites allow users to hide machine-generated reviews, and one of the reasons for this is that this type of language generation may have a disadvantage.

For example, an AI that writes reviews can be used to synthetically amplify positive reviews and drown out negative ones, or vice versa. “An online product review can really change people’s minds,” notes Ben Zhao, a machine learning and cybersecurity expert at the University of Chicago who was not involved in the new study. With this kind of software, a person with bad intentions can “completely blackmail a competitor and financially ruin their business,” Zhao said.

But Kopalle and Carlson see more potential for good than harm in developing software that generates reviews, especially for small business owners who may not have enough time or knowledge of English to make product descriptions on their own.

We already live in a world shaped by algorithms, from Spotify recommendations to search engine results and traffic lights. The best thing we can do is continue with caution, Zhao says. “I think people are incredibly easy to manipulate in many ways,” he says. “It’s just a matter of identifying the difference between proper use and abuse.”

Leave a Comment