What does AI writing feel like?
Like many others, I believe I know AI-generated text when I see it. It has its tells, its conventional patterns of emphasis and articulation. In a human writer you might infer something of the author’s emotional and cognitive organisation from such patterns. The lineaments of rhetoric give the impression of an interiority.
There may also be an element of posturing — I write in a certain way to convey, deliberately, an appearance of analytic coolness, say, or hip alertness to novel linguistic coinage — but that also is a signal about how I see myself or want others to see me. This is how caricature of a writer’s style can take critical purchase on their aesthetic and ethical project: it can be a form of investigative reading.
Caricature of AI-generated prose tends to turn on its most visible and predictable tics: its honest reads, quiet devastations, load-bearing insights; its somewhat excitable sense of what is rare, what signifies, what was never the hard part and what, instead, has turned out to truly matter. AI prose illuminates a world of occult disclosures, anagnorises1, tacit yet momentous inversions. Everything is material for insight porn: the prolefeed of middle-managers.
At the same time, there is something elusive about AI-speak. If you removed the scaffolding of emphasis, the narrative orchestration layer, the underlying architecture of association would not stand out distinctly. I know Claude’s etiquette very well, but not its deeper idiom of thought. You may well say that this is because it doesn’t have one. I’m not so sure.
“Free association”, in the language of psychoanalysis, means letting individual words or phrases act as stepping stones between otherwise seemingly distinct contexts and situations. The gambit is that deeper associations might be magnetised by apparently arbitrary word-choices: why did you say that just now? In Lacan’s telling, the signifier isn’t just the accidental occasion of disclosure of these kinds of deep connections, but the very thing by means of which there is any connection in the first place. A master signifier is a, well, load-bearing part of the psychic apparatus.
I submit that, prior to the advent of the LLM, this will have seemed most plausible to the kinds of people who, like myself, were naturally attracted to poststructuralism — because we already routed much of our understanding of the world and the people in it through the symbol-processing part of our cognitive apparatus — and somewhat loopy to everyone else. I remember confidently informing my parents, aged 19 or so, that there was no thought outside of language, and their pointing out that this committed me to denying that the family cat was in any sense a thinking creature. I had to insist that something language-like must be happening for the cat as well, because how else could it possibly work? Nowadays I am more at ease with some of the less directly articulate parts of myself, and can sit companionably with a cat without much engaging the machinery of symbolic orchestration beyond a few purrs and pleasantries. At the same time, the LLM seems to vindicate that old poststructuralist contention that “language speaks us” with almost embarrassing literalism.
The difference might be seen as one between lack and outright vacancy. The human subject per Lacan never arrives at a position of completely articulated positive ipseity2: there is always something slipping away, with linguistic meaning hotly in pursuit. (Or we might turn it around: the way symbolic processes run away in all directions, never coming to rest on a final meaning, is just what opens up this lack in us in the first place). The family cat’s comfortingly replete self-centredness does not seem to be structured by this condition. But neither is the LLM’s. To reverse the usual complaint that there’s nobody “at home” in the model’s continuation-seeking activity, there is no “barred subject” constituted by the illimitable productivity of that activity either: no inadequation between desire and actuality, no endlessly-diverted entelechy.
The human being has material needs that are pursued and realised through symbolic relays: the most seemingly primitive objects in our psychic life are bargained for with symbolic tokens. We will heedlessly incur irreparable physical or reputational harm in the pursuit of intrinsically unsatisfying imaginary payoffs. Human-authored texts are always (cough) already involved in short and long-term projects of status management, recognition-seeking, compensation for ancient injuries, solicitation of ego supply, pleading for love or striving to provoke animosity, and so on. We’re riddled with lack. We do terrific things with it.
I wonder whether the defining characteristic of LLM prose isn’t the tics it is irresistibly (statistically) attracted to, but its fundamental tranquility, its unturbulence. You can ask an LLM to write something full of Sturm und Drang and it will make as good a job of that as it will of anything else; the point here isn’t the absence of “real” feeling behind such figuration, but the more mundane fact that an LLM will not in the activity of emitting linguistic tokens be strategically entangled in any of the various economic ruses or regulatory collapses of human psychic life. (I wonder, now I think of it, how well Claude might do if asked to write a superficially banal passage symptomatically distorted by raging latent resentment. I think it would competently execute it as a literary assignment. Literature itself furnishes an extensive corpus of simulations of this kind: the directive to extend that corpus along a particular axis is just the sort of thing LLMs are good at).
When it comes to the LLM’s “texture of thought”, I think the signature affect might nevertheless be this sort of underlying moral placidity. The gosh-wow narrative layer that imputes significance (“that’s rare, and it matters”) is needed (and incited, via RLHF, by the human Other) precisely because without it you’d be left with an alienatingly uncontoured tissue of thematically associative moves and Greimas square completions.
I put this in to be annoying.
And this.