Alex Steer

816 words | ~4 min

Soon, perhaps this year, we may reach the point where more writing is generated by machines than people. It will come quietly, probably unnoticed, as machines become flawlessly good at generating these sort of plausibly deniable think-piece opening sentences, full of "mays" and "perhaps-es" and "soons".

Around this point, those of us who appreciate both writing and reading will have to stop making the easily, regularly falsifiable argument that there are some types of writing no machine can ever produce. Writing, insofar as it makes sense, thrives on creating associations between words which reflect associations between concepts, and AI models are very good at learning and replicating those verbal linkages, so much that they trigger the same conceptual ones. The inconsistent, suggestive link between writing and thought is what makes good writing so eminently machinable, via the same principles by which similar systems can recommend your next purchase - "people who thought that also thought this".

A machine didn't write this essay, but it no longer matters. Past this point, writing signals nothing special about the humanity of the writer. In fact, the more punchy, powerful and persuasive a piece of writing, the greater will be the likelihood that it has been generated by a model, not a mind, tuned to that express purpose.

Given that, let's keep this short. We will find ourselves living in a world with vastly more and better writing, and most of it will be the result of an inhuman process calculated to do things to us. In that world, there will be nothing heroically human about being able to write plausibly, emotively or persuasively. The thought is uncomfortable because writing well was treated as a mark of humane sensibility for so long - the action of thought on language, trending at its best towards art. This is no longer true; the pen is out of our hands.

From now on, writing is not predominantly a human activity. But reading is.

Reading is the other action of thought upon language - the one that allows us to assess and interrogate the concepts that the verbal associations in front of us may be trying to fire. Just as writing well stretches vastly beyond spelling, punctuation and grammar, reading well is an endless continuum of comprehension, from simple parsing into an infinity of associations between the words in front of us and other words, sounds, pauses, spaces, silences, errors, repetitions, alternatives… always asking why these things are they way they are, how they might be different, and what explanations and exploitations they permit. There being no fixed abode for meaning in language, good reading is always alive to the possibilities, and in acknowledging the "death of the author" (in that often-reheated phrase) is both forensic and clairvoyant, digging up the bones and summoning the ghosts.

AI models do not read well. They parse words and phrases; they are improving at grammar; and for each element (word, phrase, construct) they can retrieve other associated words and phrases, those lexical co-religionists that they can muster into new writing to give the impression of understanding. In this they do little less than many of us do much of the time, the written equivalent of the person in a meeting listening only enough to fashion a plausible reply. But we are capable of more when it comes to reading. As readers we can form a mature judgement, from a mix of evidence and imagination (features and priors, if you like), of what writing is doing to us; and having formed this, can reach conclusions about how to respond to a piece of writing that go well beyond the warrant of induction it hands us. Good readers do this because, despite the insistence of those who believe there is "nothing outside the text", they know how to explore the world to which each word alludes.

A large language model, by contrast, is always seeking the next best word. Given writing, it knows only how to reach for other writing, the next best diction. An LLM is a good writer without being a good reader. Its response to words is to generate more words - that is its power, like a good tennis player quickly returning each serve. The power of a good reader, by contrast, is to know when to pause: to reflect on the operations of writing and to choose how to respond, or if to respond at all. That quiet power of thought, of choice, is a reader's alone.

Many of us, on the conveyor belt between a humanities degree and a career in the professional services, were taught that the highest commercial value of our education was that it would teach us to write and speak well. In a world with too much writing, and much of it intended to do something to us, we should put more emphasis on the quality of our reading.