Writing Slowly

In which the artificial style of writing encounters the iron hand of fashion

AI makes writing more bland, as reported by NBC News. This will accelerate the rate at which readers demand new forms of writing that AI can’t yet (or ever) achieve. There’s been plenty of talk about how AI caters to the economic requirement for efficiency (aka reduced labour costs), but there’s another very obvious economic requirement too: novelty.

AI handles novelty of combination very well. Just try asking it for a story about a unicycling giraffe who learns quantum mechanics and escapes from a zoo in the Alpha Centuri star system — it won’t refuse you. That kind of combinatorial novelty it handles with aplomb. At no point will it tell you this is a bad idea. On the contrary, “You might consider how the unicycle itself acts as a metaphor for the observer effect,” says Google’s Gemini as it eggs me on. But it doesn’t produce novelty of expression. Despite the arguable novelty of this scenario, the adventures of a fugitive circus scientist space giraffe, the story, as written by AI, will still be, well, bland.

Cheaper, better and newer. Consumers crave novelty, which drives the endless parade of fashion, and the instant obsolescence of what only yesterday was highly desirable. In a sense, AI writing stands at the end of an era, the era that saw the kind of writing on which AI has been trained as up to date. For example, if a chat-bot wrote an article or a piece of advertising copy in the 150-year-old style of Charles Dickens, it would be quaint, but hardly useable. To achieve the effect it does, of being ‘as good as an average human writer’, it must mimic what’s considered the current writing style, and it does so blandly.

Now the research shows that AI-generated prose isn’t just bland, it’s also distorted in several other ways, many of them, such as pronoun use, connected with style.

How LLMs distort our written language.

But fashions change, and the blandness of the AI style will accelerate the speed at which writing style fashions change. Just as we can’t take seriously today someone who writes like Charles Dickens (unless it’s deliberate pastiche), tomorrow we won’t be able to take seriously any writing produced in the style of a bot. And that means soon we won’t be able to take seriously any writing that’s written in our current style.

You might find this hard to accept, since in our lifetimes writing styles have tended to change quite slowly. Your style is different from that of your parents when they were your age, but only very slightly, and the gap between your style and that of your grandparents when they were your age is only slightly larger again. But try reading Shakespeare and you’ll see very clearly that over the space of just a few centuries writing style can change dramatically. Even though Shakespeare is definitely considered ‘modern English’, it doesn’t seem very modern when you’re trying to understand it.

But fashions change at varying speeds. We know this from observing clothing fashions and pop music, where there’s a ceaseless seasonal turmoil, but there are slower fashion cycles at work too. To give just one example, I’ve been surprised to observe how fashions in pet dog breeds have changed radically in my lifetime. Even into my 30s I assumed the dogs we had were the dogs we were always going to have. This assumption hasn’t aged at all well. It’s labradoodles everywhere now.

My prediction is that the arrival of large language models is going to speed up the fashion process in writing, so that a clear gap opens between the style AI can produce and the style that’s considered up to date by actual humans. Sure, bureaucracy will probably thrive on the bland style that AI has perfected, just as even now its ‘formal’ style seems slightly old fashioned. But the cutting edge of human communication will soon be leaving this era far behind.

Now you might argue that this just isn’t how language works. You might observe that the kids, in each new generation, have their own ways of speaking and yet their slang and their idiom don’t quite take hold. Sure, some expressions break through to ubiquity, but on the whole, language change maintains its own pace, which is a little slower than the span of a generation. True, you can see in recordings of the long-lived Queen of England, that she changed her accent through the course of her life, but she didn’t do it quickly or dramatically.

I accept that this is how it has been up till now, but my point is that the near future will be different. AI is going to speed up the rate of change, quite dramatically. By the time AI can write an article like the one you’re reading now – an article I wrote by hand myself, but which you might not be completely confident I didn’t use AI for some or all of it – by that time, which is surely coming soon, we’ll have to change our idiom quite radically if we want to speak and write like humans. Attempts to somehow certify human writing as such already verge on the farcical¹. Ironically, the more convincingly AI learns to imitate us, the more urgently we’ll be making ourselves sound quite different.

True, AI raises some important and controversial questions, such as ‘could a large language model be conscious?’, as philosopher David Chalmers has asked. But in an important sense, for the changes I’m predicting such questions don’t matter. The better and more competent AI becomes, the more pressing will be the urge for humans to behave differently.

The language the LLMs have been trained on, which is basically the language of our own generation, more or less, will not be used by those of us who, for whatever reason, but certainly by reason of the harsh imperatives of fashion, wish to demonstrate that we are human. For good or ill, a new generation, though raised in the lap of AI will not be speaking or writing anything like the ‘intelligence’ that raised it. Even though modern English has only existed for a few centuries, and even though it’s already quite hard to understand Shakespeare, not to mention Chaucer, it remains very difficult to imagine our own language changing beyond our present ken. Yet this is what’s about to happen. Meanwhile, AI will be stuck in the past that its training data prepared it for. Human language will press onwards, as it always has done, but rather faster, as LLM training strains to keep up. In this sense, the rise of AI marks not the beginning of a new era but the end of an old one. What happens next might not really shock you, but it will leave algorithmically-generated text far behind. I’m the author of Shu Ha Ri: The Japanese Way of Learning, for Artists and Fighters, available now.
And if you found this article interesting you might like to sign up to the Writing Slowly weekly email digest. You’ll receive all the week’s posts in that handy email format you know and love.