ChatGPT sommelier suggests wines at Tasca da Esquina – 04/05/2023 – Food

It is unlikely that someone who has never drunk wine would suggest wine labels to pair dishes in a restaurant. But whoever is offering this consultancy at the traditional Portuguese restaurant Tasca da Esquina, in São Paulo, is not exactly a sommelier —in fact, it is not even a person, but the ChatGPT.

The traditional Portuguese food house has adopted a system that uses the most talked about artificial intelligence of recent times to suggest wine options that match the ordered food. As ChatGPT uses all the knowledge available on the internet to try to answer any question asked of it, the tool is based on this information to suggest harmonizations.

The system, called “virtual sommelier”, was created by the company Atena, which provides software for operating and managing restaurants. Tasca da Esquina claims to be a pioneer in implementing this tool.

Atena’s CEO, Francisco Gioielli, explains that artificial intelligence uses APIs to enable communication between the platform on which the customer chooses the dishes and the GPT Chat, which will suggest the wines. API, according to him, is “the language and environment that allows two systems to talk”.

In practice, the customer arrives at Tasca da Esquina, points his cell phone at a QR code and chooses the dishes of his choice. But, contrary to what usually happens with restaurants with virtual menus, the order can now be placed on the consumer’s own device, and that’s when the virtual sommelier comes into action.

Each dish is already associated with some drink options, such as red, white, rosé or green, and the customer can choose one or more of them. Then there are four price range options, up to R$300, up to R$600, up to R$900 or without limit. After selecting the value, ChatGPT presents three final labels.

One of the founders of Eno Cultura, a school that has trained sommeliers since 2013, Paulo Brammer, accompanied the report on a visit to Tasca da Esquina to check out the artificial intelligence wine recommendations.

Contrary to what one might imagine about a sommelier evaluating a tool that could even replace this function, Brammer’s opinion on the novelty is, in parts, positive. The first point he praised was the ease of use of artificial intelligence.

In general, the sommelier considered that the ChatGPT suggestions were correct, but some combinations left something to be desired, according to him.

During the visit to the restaurant, a shoulder of lamb paired with red wine was served, which is within the expectations of experts. The issue is that the dish has mint in the seasoning, something that does not seem to have been considered in the virtual sommelier’s suggestion.

For Brammer, the most suitable label in this case would be a cabernet or some wine that has notes of eucalyptus or menthol. The sommelier believes that the problem occurred due to the lack of personalization of the dishes, which are described without all the details so that the tool can find the most appropriate suggestions.

It is when the customer chooses options that are very different from each other that the tool has the most problems. This difficulty is not exclusive to the tool —Brammer says that even good sommeliers also need a lot of experience to know how to balance opposite flavors.

When ordering, for example, filet mignon, which pairs well with red wines, and gratinated shrimp, which goes best with white wines, the virtual sommelier suggested only red wine options. For the flesh-and-blood professional, the solution would be a rosé “more full-bodied, maybe even with a little wood”.

The lack of personalization also applies to the texts that accompany the three labels suggested by the virtual sommelier. Brammer considers that they fail when they suggest, for example, that the alcohol in the wine moderates the flavor of the dish, which has no technical basis.

The paragraphs of descriptions generated by artificial intelligence are generic, says the real-world sommelier, and it is clear that they were not written by someone who understands the subject.

Brammer explains that the justification given by a real sommelier is not only useful to link a dish to a wine, but above all to connect the producer to the consumer. This experience also deepens the uses of the label to the context in which it is being consumed, such as dinner with friends, a business meeting or even the temperature of the day.

Despite helping customers and restaurants at first, Brammer believes that the relationship that happens between the sommelier and the customer is something that the tool will not be able to replace.