Jamie Goode:  How soon will AI replace wine critics?

Tasting and rating wine is an imprecise science. Try as we might, it’s hard to get the universal ‘truth’ on a wine. As we taste, we strive to get to the true ‘flavour’ of the wine, and each time we come back to it we find something new. And then we have the challenge of capturing our perceptions in words, in order to codify them and communicate them. I know that when I get together with my friends and taste interesting wines blind, in the absence of any other information understanding a wine fully from just what is in the glass is hard. It’s often only when we know the identity of the wine that we can begin to understand it properly and to enjoy it fully.

At the same time, we live in the age of the wine critic, where many people are making a living from delivering their verdicts on new-release wines, hoping to become (or solidify their reputation as) the go-to expert for a particular wine region. They sell their scores, both to consumers and increasingly also to wineries. But there are often divergent opinions, and sometimes it seems more of an art than a science.

But what if we could turn rating wine into a science? Artificial intelligence (AI) promises a lot – why shouldn’t it enter the space of perception also? In order to answer this question, first we need to look at how this might work, and what is needed of any AI tasting and rating system.

The first challenge is one of identification and separation of different samples. Before we can think of digital olfaction and gustation, we need to satisfy the far more simple task of discrimination among wine samples.

There are a couple of rather different ways of going about this, one of which is obvious but unpromising, and one of which is more complicated and less obvious, but perhaps more interesting.

The first is to use a collection of measuring devices to detect what is there in the wine. This is the traditional chemistry approach to looking at wine flavour. Identify the flavour compounds with an odour activity value (OAV) of 1 or more (where OAV is the concentration of the odorant in wine divided by its threshold), and look at the non-volatile flavour components (alcohol, acids, sugars, polyphenols) and maybe the contribution of the non-volatile matrix that affects how the aromas are released. Then after breaking the wine into its components, you have the basis of wine flavour chemistry, and you can analyse each wine for the levels of these various compounds. This might help you to give a chemical signature of types of wines, but such a reductionist approach will only get you so far, and an electronic nose on its own can’t tell us much about the perception of wine – which is multimodal and much more complex.

The second way is to use machine learning, which attempts to mimic how we learn to perceive the world around us. We learn to recognize combinations of odour molecules, and from this create odour objects, through lots of experience. Take the smell of coffee. Whether it is instant coffee, or gourmet coffee, or even coffee beans that are freshly roasted, we recognize the smell of coffee. There are hundreds of different smell molecules involved, and the brain must have ways of recognizing these combinations, even though they may differ significantly of various occasions, they are still recognizable as coffee. Even someone who has never smelt coffee has the ability to learn to recognize coffee with repeated experiences. This sort of object recognition is widely accepted in the visual system, but less so by the olfactory community who have traditionally focused on single molecule odours and their perception. Somewhere in the brain a module must be saying if we have this, this and this smell, then its coffee. Interestingly, we can detect coffee even in a situation where there are competing smells, such as pizza and burgers. All those molecules are flying around and getting into our noses, and we are able to split the various patterns apart. It’s quite a feat.

Think about wine. As we taste wine, the first impression is, oooh, this is wine. Then we use our abilities to dig a bit deeper. It’s a Pinot Noir, we might think, because we recognize certain patterns of smells and tastes. What we are doing is computationally very complex, and as well as the sensory perception, we are relying our knowledge and experience as we taste, and we probably also do templating, where first we decide what sort of wine it is, and then we look for certain elements in the wine that we associate with that type of wine. Is there oak there? Is there some greenness? This directed searching no doubt influences our perceptions.

For an AI-assisted mechanical device, there would have to be quite a lot of learning. The device will not only have to pick out wine relevant tastes and smells, but also develop its own body of experience, and must learn to recognize odour objects. This is quite a task.

Then comes the quality assessment element, which brings together a whole new level of complexity. In fact, it’s so bewilderingly complex, I wonder whether it is possible. But then I look at my iPhone, and 50 years ago the existence of such a device would have seemed impossible, too. So if we are to have AI critics, then I think there’s a very great need for lots of good data to feed the machine-learning model.

And, slightly cheekily, as I suggested in my novel The Wine Critics, replicating the scores generated by wine critics to a level that no one would spot them as machine-generated, wouldn’t be all that hard anyway, in these days of score inflation and legions of critics of somewhat mixed ability. The best critics are no doubt safe: they do a great job that will be hard to replicate. The rest?

• Jamie Goode is a London-based wine writer, lecturer, wine judge and book author. With a PhD in plant biology, he worked as a science editor, before starting wineanorak.com, one of the world’s most popular wine websites.