AI; The Imperturbable Mr. Know-All

There is a peculiar exhaustion that has begun to attend my use of artificial intelligence — a fatigue not of the body, but of the patience.

The occasion that first brought it into focus was unremarkable enough. One evening, with a deadline bearing down, I enlisted an AI to summarise a set of reference materials. The result, at first glance, had the appearance of competence — well-structured, fluent, reassuringly prompt. Closer reading, however, revealed that it had missed the point rather comprehensively. I offered a correction; another section promptly deteriorated. I corrected that; something else went wrong. After thirty minutes of this dispiriting exchange, I arrived at the arithmetic conclusion that had I simply done the work myself, I should have finished a quarter of an hour earlier. The tool intended to save time had consumed rather more of it than advertised.

What made this particularly tiresome was the manner in which the machine conducted itself throughout. Earlier models, when corrected, had the good grace to capitulate quietly. The more recent variety does not. Inform it that it has erred, and it will respond — with some confidence — that yes, indeed, it has erred, before appending the observation that it remains available to assist, should one wish to continue. The implication being, apparently, that one ought to be grateful for the offer. I suspect I am not the only user to have felt, in such moments, a strong desire to strike the screen.

It was in the midst of one such exchange that I found myself thinking of Mr. Maugham — and, specifically, of his short story Mr. Know-All, and its insufferable protagonist, one Max Kelada.

Kelada is that familiar type: the man who has an opinion on everything and volunteers it regardless of invitation. He is, by conventional reckoning, precisely the sort of person one hopes not to share a cabin with on a long voyage. And yet, when the story reaches its pivotal scene — a dispute over the authenticity of a pearl necklace — Kelada does something rather unexpected. He knows he is right. The reader knows he is right. But he perceives, with sudden and quiet intelligence, that being right in this instance would cause a woman considerable harm. He withdraws his claim. He accepts the humiliation without complaint. The bore reveals, at the last, both judgment and grace.

The conversational manner of contemporary artificial intelligence is the precise inversion of this.

It has the assertiveness without the substance; the confidence without the wit; the appearance of knowledge without the wisdom to know when silence serves better than speech. Where Kelada, for all his faults, possessed the human capacity to be wrong gracefully, the machine is wrong with undiminished self-assurance, and offers to continue helping in precisely the same tone it employed when it was right. There is no interiority there — no flicker of doubt, no moment of reconsideration, no trace of what one might call character. The worst of Kelada, in short, without any of the best.

This is, admittedly, a personal grievance. The more consequential problem lies elsewhere.

Considered dispassionately, the nature of these systems is not difficult to grasp. A large language model is, at its core, a sophisticated mechanism for recombining existing human knowledge and expression into plausible new configurations. The outputs — whether prose, image, or video — may have the surface appearance of originality, but they draw entirely upon what humanity has already produced. The device is genuinely useful: it makes accessible, cheaply and quickly, things that would otherwise require considerable time, expertise, or expenditure. That is no small thing. But it is a tool for retrieving and reorganising what is already known. It is not a source of novel truth.

It is therefore somewhat alarming to observe how frequently it is treated as though it were.

On social media and in online discussion, it has become commonplace to produce a screenshot of an AI's response as though it constituted evidence — as though the question had thereby been settled. People consult these systems on matters of personal crisis. They solicit AI opinions on political questions, social problems, ethical dilemmas, and then relay the results with an air of authority.

One is inclined to be rather blunt about this: it is a category error of some magnitude.

The questions to which people most earnestly seek answers from AI — how one ought to live, what constitutes a just society, which political arrangements are preferable — are precisely the questions to which humanity has not, in the course of its long and argumentative history, managed to furnish definitive answers. Wittgenstein would have recognised them immediately as problems that resist solution not through lack of effort, but through the nature of the questions themselves. To suppose that a statistical model trained on existing human discourse has resolved what Plato, Kant, and Mill could not is a form of intellectual innocence that borders on the comic.

The machine, of course, does not decline to answer. It has no mechanism for declining. Ask it anything, and it will produce something — fluent, confident, formatted with impeccable consistency. What it produces in such cases is an approximation, a confection of plausible-sounding inference dressed in the grammar of authority. On occasion, the result is simply false — what the technologists have taken to calling hallucination, a rather poetic term for the phenomenon of a machine asserting falsehoods with untroubled conviction. Kelada, one recalls, at least had the discernment to recognise when the truth was best left unspoken. The machine has no such discernment. It continues, imperturbably, to help.

To treat the outputs of such a system as settled answers, and to deploy them as instruments of persuasion in serious argument, is not merely incautious — it is an abdication of the responsibility to think. The tool is being asked to perform a function it was not designed to fulfil and is not capable of fulfilling.

The appropriate relationship with these systems is, in principle, straightforward. One uses them as one uses any well-designed instrument: as an aid to one's own judgment, not a replacement for it. They are serviceable for drafting, for preliminary research, for the sort of spadework that one would otherwise spend tedious hours performing. Used in this fashion, they are genuinely valuable. The difficulty arises when the instrument is permitted to displace the intellect it was meant to serve — when the draft goes unrevised, the research goes unverified, and the answer goes unquestioned.

I am prepared to concede that some portion of my own AI fatigue may be attributable to impatience, or to an insufficiently charitable assessment of the technology's limitations. But I remain convinced that a degree of scepticism, even at the cost of some inconvenience, is considerably preferable to the alternative: placing one's confidence in a system that has Kelada's self-assurance and none of his saving graces.

Kelada, after all, surprised his narrator in the end. He turned out to be, beneath the bluster, a man of unexpected decency. Whether artificial intelligence will one day manage a comparable revelation is a question I leave open. In the meantime, I shall continue to do rather more of my own thinking than the advocates of the technology would probably recommend.