The new world of Artificial Intelligence is upon us. Although it has been around in some form for decades, the sudden onslaught and easy access of ChatGPT and competitors have made it a commonplace topic for many professions. Wine is no exception. It’s been fascinating to watch a whole new side of the wine business start to grow. Having just undergone a coding boot camp, I am especially curious about the intersection of wine and coding. This past spring, ChatGPT passed the theory part of the Introductory Court of Master Sommelier exam with a 92% success rate, then the Certified Sommelier exam with 86%, and the Advanced Sommelier Exam with 77%. As the Master Sommelier Exam is oral, that exam, to my knowledge, has not been attempted yet with AI. Besides the clickbait aspect of AI passing exams that people spend months and years preparing for (I can attest to this and have the pile of flashcards to confirm), innovators and researchers have also been tasked to do a few more things in the service of wine such as wine reviews, AI sommeliers, and nanomagnets that can “blind taste” a wine.
So far, AI, like ChatGPT, can offer up tasting notes on any wine you want, so long as it has been written about and information is published online, or information about wines from a certain region is available for it to utilize. What it doesn’t do is fact-check the information or the sources nor vet the wine professionals – meaning a writer who may not necessarily enjoy or understand the wines they are writing about. Would I use ChatGPT to prepare for exams? No. I could see its future use for generating random theory questions, but until the data source can be completely vetted, I would hesitate to rely on it.
Speaking of theory and wine knowledge, I’ve seen a few AI Sommelier software startups peek their heads out into the world. Utilizing a database of sommelier knowledge, they attempt to create that first-hand experience of a trusted sommelier utilizing a smaller, more controllable database. Although this AI technology would utilize a more curated primary source for its recommendations, arguably, it would not understand the nuances and grey areas of being a sommelier. A person told me they love wines with a “watery finish.” After asking a few key questions, I saw that they were looking for a wine with high acid levels; acid makes your mouth water - wines like Riesling or Sangiovese were what this person liked and wanted to drink. I translated their sensory perception of watery finish because of my training, tasting thousands of wines, and navigating wine service table-side for numerous years. What this person wanted in this example was a Chianti. Would an AI Sommelier be able to translate what the person wanted? Not initially, no, it would take someone inputting this material into the source bank. It will be interesting to see how long it will take for the various ways people describe wine to be collected and compiled for AI usage.
Nanomagnets “tasting wine” – this was the most fascinating usage for me. Wine aficionados and sommeliers use deductive analysis for blind-tasting wines; the neural networks in our brains take in the information from each sip when tasting wine, and the synapses weigh the importance of each data point – the acid level, ripeness, and quality of the fruit, tannin levels, etc. As this information gets processed by our brains, we make an educated guess about the wine’s identity. Scientists want to take this process and teach it to AI; to do so, they have designed computer versions of our neural networks to take in and process information as we do. Pour yourself a glass of wine; it will get a bit dense for a moment – but very worth it. Our brains can complete this deductive process using less wattage energy than a computer, so researchers are looking at energy-efficient ways to replicate our deductive analysis. Scientists at the National Institute of Standards and Technology (NIST) have tried magnetic tunnel junctions (MTJ), which is a device that is good at mathematics similar to what a neural network uses – meaning they store data where computations are done, not in a separate part of the program – making them quicker and energy efficient. To see if MTJs could be used as a neural network, they trained it to prepare for a blind wine examination – behold the power of the human brain comprehending wine! For this test, researchers trained the MTJ on 148 wines from a data source of 178 wines made from three types of grapes. The researchers kept to 13 key points for each wine – alcohol level, color, flavonoids (anthocyanins – which help pigment wine and tannins), ash, alkalinity, and magnesium. Each data point was assigned to be rated between 0 and 1. The MTJ was then given the tasting exam on 178 wines, 30 of which it had never seen before. It passed with a 95.3% pass rate and only missed 2 of the wines it had never “tasted” before. I wish I had that success rate in my blind tasting. They intend not to build a Super Sommelier but to use this neural pathway-like technology for industrial uses that could process data energy-efficiently and possibly at the source – perhaps with drones.
It’s a brave new world with AI biting at our heels to find its role in our lives. The industrial applications of the MTJ technology are fascinating, as is the methodology of blind-tasting wine used to train the computer program. AI Sommeliers could be useful for general information. Still, nothing compares to a wine-savvy human that can translate in real time what a “watery finish” means and offer recommendations or even ask a few more questions and offer interesting alternatives. ChatGPT can be used as a tool, but it will only be as good as its data source - with none of the charm and wit of a human who labored over their word choice and writes in a moment of inspiration. Language is a human construct; playing with words, meaning, and sentence structure cleverly, thoughtfully, and innovatively is very human. Creativity and ingenuity are ours - I argue that humans will always understand humans better than our future Advanced Sommelier AI overlords.