The conversation about artificial intelligence in the visual arts has been running for three years now, and it has settled into two camps that rarely speak to each other. One says the technology is a plagiarism machine that devalues a lifetime of craft. The other says it is simply the next brush. Both positions are held sincerely, and both are too broad to be useful to a painter standing in a studio on a Tuesday morning wondering whether any of this applies to them.
What has changed recently is that we can stop speculating. Enough artists have integrated these tools into working practice that patterns are visible, and the patterns are considerably less dramatic than either camp predicted.
Where it actually shows up
The uses that have stuck are, almost without exception, upstream of the finished work.
Artists report using generative image tools for composition studies generating thirty variations of a figure's placement before committing charcoal to paper. For colour exploration, feeding a photograph of a half-finished canvas back in and asking to see it under six different palettes. For reference gathering, producing a plausible image of a pose or an architectural detail they cannot photograph and do not have in their morgue file.
None of these produce the artwork. They compress the part of the process that was always slow and never visible: the deciding.
There is a second, unglamorous category that comes up constantly the administrative periphery of an art career. Mock-ups of how a series might hang in a given room, for a collector who cannot visualise it. Images for a grant application deadline. Social media material that does not eat a studio day. This is not art-making, and nobody pretends it is, but it is the work that surrounds art-making and consumes an enormous amount of an artist's week.
What the objection actually is
The training-data argument deserves to be stated precisely, because it is frequently stated badly.
The complaint is not that a machine learned from existing images every artist does that. It is that commercial models were built on scraped work without consent, credit, or compensation, and are now sold as products that compete with the people whose work built them. That is a specific grievance about the supply chain, not a metaphysical claim about creativity, and it does not evaporate because the output is sometimes beautiful.
It also has a practical dimension artists care about: several models will reproduce a living artist's style on request by name. An illustrator who spent two decades developing a recognisable visual language can now be approximated by anyone typing their surname. Whatever one thinks of the philosophy, that is a direct commercial injury to a working person.
Some providers have begun training on licensed and public-domain material and offering indemnification. Whether that becomes the industry norm or stays a premium niche is, at the moment, genuinely unresolved.
The institutional position
Museums and galleries have been noticeably quieter on this than the discourse would suggest, and their actual practice is instructive.
Most institutions have not banned the technology. They have done something more mundane: they have started asking. Several open calls and acquisition processes now require disclosure of generative tools used in a work's production, in the same way they have long asked about materials and editions. The position is not that the work is disqualified, it is that the catalogue should be accurate.
Conservation departments, meanwhile, have a more concrete worry provenance. A print whose source file was generated rather than drawn raises questions about what exactly is being preserved and what the edition means. These are not culture-war questions; they are cataloguing questions, and they will be answered slowly and administratively, as such questions always are.
The technical situation, briefly
For artists weighing whether to experiment, two developments matter more than the headline quality improvements.
The first is that these models now take an image as input, not just text. You can feed in your own sketch, your own photograph, your own unfinished canvas, and ask for variation on that. This is the difference between a slot machine and an instrument. Text-only prompting produces work that belongs to the model's aesthetic; image-conditioned work stays tethered to yours.
The second is cost structure, which is unusual and worth understanding before anyone signs up for anything. These tools are billed per generation rather than per month, because each image or clip costs the provider real computation. The meaningful number is therefore not the subscription price but what a hundred discarded attempts cost and artists discard most of what they generate. This applies with more force to moving image, where artists working in video and installation are running the same experiments and paying per second of output rather than per frame. Anyone budgeting for studio use should compare those rates directly rather than the monthly headline; the current rates for the Veo 3.1 API on APIMart and the competing image and video models give a reasonable sense of the range, which varies considerably more than the output quality does.
The practical advice from artists who have been through this: do not commit to an annual plan. The field moves fast enough that whatever you standardise on will be superseded within a year, and the only sensible posture is to stay portable.
What it has not done
Three years in, it is worth noting what has not happened.
Gallery representation has not collapsed. The market for original physical work painting, sculpture, print has not been displaced by generated images, and if anything the premium on demonstrably hand-made work has risen. Collectors who buy paintings were never buying image files.
Nor has the technology produced a body of work that the art world takes seriously on its own terms. The generated images that have entered critical conversation have done so as part of a practice with an argument behind it, which is exactly how every other new medium entered.
What has happened is quieter and probably more consequential: the floor of visual competence has risen, which means competence alone is no longer a differentiator. That is uncomfortable for illustrators and commercial artists in a way it is not for fine artists, and the profession most affected is the one least discussed.
Where this leaves a working artist
The honest summary is that this is a tool with a real use, a real ethical problem attached to its origins, and a much narrower effect on finished work than the volume of argument implies.
The artists getting value from it treat it as a sketchbook that thinks fast useful for deciding, useless for making. That is a smaller claim than either side of the debate wants to hear, and it is the one that survives contact with an actual studio.
