Resisting the seductions of AI fluency

Most pitches I now receive are clearly leeched from large language models (LLM). About once a week I come across an article in a magazine, or a newspaper, that clearly depends on artificial intelligence. One of the more subtle clues is the way writers rely on it to generate references, especially to a particular cultural artefact, or historical gobbet. Why should that matter? Surely, it is much like using the Encyclopaedia Britannica, or Google, to give the writer the information they need? It is true that using a reference without an awareness of why it is being invoked, and what role it plays in the argument being made, is common to most writers, especially those who lack confidence in their own voice.

Early in my career, I would often think it cleverer to find a quotation to back up my argument than put it in my own words, perhaps a hangover from the kind of standard paragraph structure taught at British schools, such as ‘Point, Evidence, Explain, Link’. Very few of the great essay writers in the Anglophone tradition follow the rubric. In France, the culturally dominant model is thèse-antithèse-synthèse, a system which would have confounded the essay’s greatest exponent, and inventor, Michel de Montaigne.

Large language models are trained on huge swathes of articulate human speech, on trillions of words. Models are only models, partially accurate representations of the raw stuff of life, and large language models are only approximations of human writing. The most grating quality of LLMs is the compulsion to tie every sentence up neatly. Human thoughts are long and short, bitty and abstruse. Sometimes they need lots of words; sometimes very few. But a large language model always pursues clarity even when clarity compromises meaning. Its general statements are usually vacuous because the machine is proceeding along a preconceived route.

Conceptual models are thoroughly useful things – human beings operate on extremely limited and basic information. We need maps. Partly, we need models to grapple with the act of thinking itself; the notion of the ‘psyche’, or the ‘ego’, or the ‘id’, are merely models of aspects of the mind. And yet, the challenge is to go further, in effect, to ‘dare to know’. This task is as much to do with the process of unlearning, unknowing, and reformulating the stuff of the classroom.

Relying on large language models to shortcut the difficult work of finding the right reference, or to produce polished, seamless prose, leaves the writer close to a state of self-imposed immaturity. A large language model is like a complacent pupil who has mastered the rules, and thinks the work is complete.

To be a writer at all, and especially to write well, with spirit and flair, the act of writing must be at least close to the summit of life’s pleasures. It must relieve the portion of suffering allotted to each of us. In William James’ ‘Letter to Thomas Wren Ward’, an address to a depressive friend of his, he talks about the need for human beings to make a ‘nick’ in reality. Some writers hack great portions from the raw matter of the world; some leave just a scratch – but all born writers must be somehow convinced that even these little nicks matter, and that they are made, mysteriously, in their own image.