Monday, October 26, 2020

Should we worry about Artificial Intelligence in wine writing?

Recently, Richard Hemming published a short article called Are wine writers redundant? Most of the actual text was produced by the HAIMKE Artificial Intelligence (AI) system, under the guidance of Yoav Shoham. Yoav said: "Cute, I think", while Richard said: "Terrifying, I'd say." These are two extreme viewpoints, of course, and reality is presumably somewhere in between. But where?

To me, the important point to note is that two human beings were actually involved in producing the article — and none of us would bother reading it if there had been no people involved at all. Being human is a rather human activity, which means that it is based on social interactions. Humans interact with other humans by choice. If things are done by machine, then they have no especial human interest.


After all, isn't that the fundamental cause of the current second wave of the Covid-19 pandemic, in which infections are increasing to levels not seen for several months? People are thoroughly sick of all those months of "social distancing", and having to look like masked bank robbers making a getaway. They want some social human contact, such as they previously had in their lives (and which they one day will have again). So, the restaurants and bars are full, and the young people are partying again. The virus loves it. (For an interesting discussion of the psychology of our current pandemic responses, see: Acedia: the lost name for the emotion we’re all feeling right now.)

So, should we worry about AI in wine writing? Only if the readers want no social contact with another human being. A blog post concerns human communication, for example, as do YouTube videos, FaceBook messages and Twitter tweets. The world is full of noise, of course, but in amongst it there are some real attempts at communication.

You see, the thing about AI is that it is and it isn't. It is artificial, and it isn't intelligent. Even though we call it "machine learning", the computer doesn't actually learn anything. All that most AI systems can do is take a massive amount of input data, and create an organized database that synthesizes the observed patterns of relationships among the different parts of the data. Then, when the system is later given one of the components as a new input, it produces a closely matching output. Impressive, sure, especially when the output looks natural; but human beings are not quite that mechanical — indeed, humans often need only one previous example of something, in order to correctly identify new examples.


We will, of course, be seeing many more uses of AI in the future; and I have written about some of them before (Artificial intelligence in the wine industry? Not yet, please!). Car driving is an obvious candidate, given just how bad most humans are at it. I have always thought that the most remarkable thing about people is their ability to fall asleep at the wheel. Cars can kill, as hundreds of people find out every day; and yet our minds grasp this fact so poorly that they think it is a good time to take a nap while driving. AI has the potential to alleviate the danger of this situation. Even bad AI will probably be better than many human drivers.

So, AI can be used for many things, including unexpected ones. As one interesting example, while there are specialist computer programs for playing the game of chess (eg. Deep Blue, Stockfish, AlphaZero), even a general text AI system can also do it (such as HAIMKE, referenced above) — this is called "natural language processing" (NLP). Chess moves can be written in the form of words (see Chess notation), and so a text-based AI system can be given a large collection of chess games that have previously been played. Even without any further instructions about how chess is played, after it has generated its database, it will have discovered all of The most common chess openings. Then, in response to any given chess move, it will respond with one of those opening gambits, just like a person would. The system knows nothing about chess, not even the concept of a chess board, let alone the rules of play — it simply produces what are apparently appropriate (text-based) responses to any new (text-based) input. Under these circumstances, is this system really "playing" chess?


Anyway, it seems that our children's children will be living in a different world, and this will be an AI one for sure. But I doubt that they will be reading wine writing if they know it was "written" by a computer rather than a person — there is an important difference between creative "writing" and mere "text generation". The latest NLP systems can produce some pretty remarkable performances, such as creating poetry that seems just like the real thing (see GPT-3 creative fiction); but is there much of a market for it? * AI has its obvious uses, but human communication is not one of them.

Mind you, the products of actual Wine Writers often seem far too much like repetition of the same old stuff, year after year; so I guess that the AI people will see reproducing this as pretty easy for a computer to do. Interestingly, this response might very well stimulate more creative forms of human wine writing, which would not be a bad thing. After all, this is what actually happened with chess (AI ruined chess: now, it’s making the game beautiful again).

At least this year the wine writers had some new things to write about, which is something that AI systems cannot (yet) do. However, I hope that we do not have to write about fires and pandemics too often!



* See also: Experimental evidence that people cannot differentiate AI-generated from human-written poetry.

1 comment:

  1. I think you're right that this will spur more creative writing and also a greater reliance on visuals. Formulaic writing (wine reviews, encyclopedic/educational content, etc.) are well within the capability of language models today. Long-form cohesion is still a challenge, but it's not clear how much of that is read anyway.

    I think we'll see a quick decline in the reliability of consumer wine reviews like we've seen on Amazon. OTOH, if professional reviewers are relieved of the task of converting a handful of descriptors to something they think is novel review content [it's not], then maybe they can review more wines from small producers who normally get looked over?

    ReplyDelete