The machine at the easel

In 2022, an image produced with a text-to-image program won first prize in the digital art category at a state fair in Colorado. The man who entered it had typed a long series of prompts, selected and refined the results, and printed the final version on canvas. Many artists were furious when the news spread; others merely shrugged. The episode was trivial in itself, yet it crystallised a question that has since moved from the margins of the art world to its centre: when a machine generates an image, what, if anything, has the human contributed?

One familiar response is to dismiss such images as theft, pure and simple. Generative models are trained on millions of existing works, very often without their creators' knowledge or consent, and they can imitate a living painter's style with unsettling accuracy. To the extent that these systems profit from labour that was never paid for, the complaint seems justified, not to mention the very real threat they pose to illustrators who already struggle to earn a living. It would be naive to pretend that the question is purely aesthetic when, for many working artists, it is a matter of putting food on the table.

Yet the charge of theft, however emotionally compelling, is harder to sustain than it first appears. Human artists, after all, also learn by absorbing the work of others; nobody paints in a vacuum. Defenders of the technology argue that a model does not store the images it was trained on, any more than a student stores every canvas she has studied. Critics, for their part, reply that scale changes everything: a student absorbs a few hundred influences over a lifetime, whereas a model swallows millions in a matter of weeks and can then reproduce a style on demand, for anyone, at almost no cost.

Then there is the question of authorship. Some users insist that crafting prompts is a creative act in its own right, comparable to the work of a photographer who chooses a subject, a frame and a moment. The comparison is not without merit. When photography emerged in the nineteenth century, it too was dismissed by many critics as a mechanical trick requiring no real skill, and it was decades before it was widely accepted as an art form. By the same token, it is conceivable that prompt-writing will one day be regarded as a legitimate discipline, with its own masters and its own standards.

That is not to say that the analogy holds in every respect. A photographer still has to be somewhere at a particular moment, and the camera records something that actually happened. A prompt, by contrast, delegates almost every visual decision to a system whose inner workings even its designers only partly understand. Hence the unease many people feel: the more of the process the machine takes over, the harder it becomes to say where human intention ends and statistical guesswork begins.

Notwithstanding these doubts, the technology is unlikely to disappear, and outright bans would not only be difficult to enforce but would also risk throwing the baby out with the bathwater. A more promising approach would be to tackle specific harms rather than the tool itself: requiring consent before an artist's work is used for training, compensating creators whose styles are exploited commercially, and labelling machine-made images clearly, thereby allowing audiences to make informed judgements. None of these measures is straightforward, admittedly, but each addresses a concrete injustice rather than a vague anxiety.

Perhaps, then, the most revealing question is not whether AI-generated images count as art, but what our discomfort tells us about what we valued in art in the first place. If we prize a painting partly because a particular person struggled to make it, an image produced in seconds will always feel somewhat hollow, however striking it looks. If, on the other hand, we care only about the finished object, the machine has already won the argument. Most of us, it seems, have not yet decided which of these we believe.