It\u2019s an uncanny experience. A social media post written by an acquaintance, yet sounding quite unlike them. It sounds polished. Punchy. Dramatic, even. Lots of full stops instead of commas. There are weirdly repetitive sentence structures: \u201cThis isn\u2019t an X. It\u2019s a Y.\u201d Or: \u201cNo As. No Bs. Just Cs.\u201d Questions are posed, then immediately answered. There are long hyphens, known as \u201cem dashes\u201d \u2014 the kind you\u2019re not even sure where to find on your keyboard. Slowly, you start to realise. This isn\u2019t your friend writing at all \u2014 it\u2019s a chatbot.<\/p>\n
From one angle, it\u2019s quite funny. For decades, university humanities departments have been hammering the idea that authorship doesn\u2019t matter; that texts are free-floating signifiers with no fixed meaning. They are wrong, of course. If it were true, lecturers wouldn\u2019t bother writing anything down themselves. But now we stand on the brink of the world they imagined. Texts are drifting about the internet, untethered. We read intention into them, only to the degree that we see meaning in tea leaves or cloud pictures. The voice may seem highly personable but there is literally nobody behind the words.<\/p>\n
Whenever I notice the style, I silently remonstrate with the person who summoned it to the screen. What is the point in pretending to be elaborately direct and \u201cauthentic\u201d when it is not actually you at the helm? This must be what Dorothy felt like when she pulled back the curtain in The Wizard of Oz. But there is also some relief in knowing one can spot the signs. The really worrying prospect is what happens when we can\u2019t.<\/p>\n
To produce text, chatbots rely upon large language models<\/a>, or LLMs: computational networks that learn patterns from exposure to huge amounts of data. A recent article in The New York Times analyses the deathless prose of LLMs in some depth. Alongside grammatical tics, there are vocabulary giveaways. One such word is \u201cdelves\u201d, common in Nigerian English but now present in thousands of academic texts in the biomedical sciences too.<\/p>\n Another telltale phrase is \u201cI rise to speak\u201d. By August, this Americanism had appeared almost five times more often in UK parliamentary speeches than in the previous year. It seems AI is no respecter of regional borders. If a word or phrase is popular in the massive datasets it mines for content, chances are it will eventually pop up in your repertoire too.<\/p>\n But let\u2019s skip a few years ahead and suppose that the machines work out how to eliminate offputting signs of their own presence. And let\u2019s also assume something that seems very likely: reliance upon them becomes even more widespread. At the moment, the norm is still to write the thing oneself. Those who use AI to produce<\/a> emails, essays, articles and speeches are freeriding on the general faith that a piece of writing still stands for thoughts originating in the stated author\u2019s head. Once that trust goes, I fear that a profoundly alienated future awaits us.<\/p>\n It is not just that new generations face mental atrophy, though they obviously do. Writing is a form of thinking on the page. As you write something down, then edit and rearrange it, you literally build a thought. When a machine spews out fully-formed paragraphs for the passive reader to repurpose and call his own, this crucial skill gets lost.<\/p>\n But never mind that: what will happen to the practice of communication itself? In the scenario I am predicting, there will still be lots of text but there won\u2019t be much writing. Writing, strictly speaking, requires thoughts passing from one mind into another via the intermediate use of symbols. That is the spirit in which most reading still occurs. Even if it is just the water company sending you a template letter, still there is an implicit belief that a person constructed the original boilerplate. And when it comes to more traditional forms of individual expression \u2014 essays, personal emails, novels, speeches and so on \u2014 understanding them as having deliberate origins is fundamental to how they are received.<\/p>\n For to read these kinds of writing is to enter into a kind of relationship with the author, however brief. Readers believe that they are encountering the thoughts of a particular person (or less frequently, a group of co-authors). As they process the content, they are getting a better picture of who the author is. Is it someone funny? Reliable about facts? Good at noticing connections? Or someone who exaggerates? Cherry-picks? Likes to whip up a crowd? Think of your favourite columnist, or even your least favourite. Don\u2019t you feel you have a sense of the sort of person they are; and what weight, accordingly, to grant to their words?<\/p>\n Essentially, the process is no different with a speech from an MP; a student essay; a love letter. You aren\u2019t just extracting impersonal information, regardless of source; you are simultaneously making connections to the person who offered it. By the time LLMs are near-ubiquitous and also (let us suppose) undetectable, we will be stuck in a kind of epistemic purgatory: knowing that the practice is rampant but being unable to tell where, exactly. Every communication will be under suspicion, irrespective of provenance. People are already feeling disassociated, distrustful and paranoid. This isn\u2019t going to help.<\/p>\n