AI and artists

An image tool produces a finished illustration in a few seconds — the kind of picture that would once have taken an illustrator a day, and been paid for. The models behind those tools learned from enormous collections of pictures pulled off the open web, and those collections contained work that living artists made, own, and never handed over.

This is not an accusation that has to be proved. It is how the largest image datasets were assembled: automated crawlers gathered hundreds of millions to billions of pictures from websites, portfolio pages, stock libraries and forums, each one paired with whatever caption or alt text happened to sit beside it. Nobody wrote to the illustrator whose portfolio was included. There was no form, no fee, and for a long time no way to ask to be left out.

Artists have been consistent about what bothers them, and it is worth hearing precisely. The objection is less that machines can make pictures than that their own work became the raw material for a commercial product, and that they were not part of that decision. A parent who hears this and does not want their household’s money or attention spent on it is responding to the part of the situation that is plainly true.

Large image models are trained on datasets built by crawling the public web. Several of the best known are open research datasets whose contents can be searched, and searching them is how a great many illustrators found their own work listed. Copyrighted images were in there in quantity. Permission was not obtained image by image, because at that scale nobody attempted to obtain it.

It is also established that a trained model can reproduce a recognizable style closely enough to fool people who know the artist’s work, and that early systems could sometimes emit near-copies of specific images that had appeared many times over in their training data. That second behavior has a name in the field — memorization — and it is documented, not rumored. Later systems added filters aimed at it. The underlying capability did not go away.

The commercial effect is not hypothetical either. Illustrators, concept artists and stock photographers have lost work that used to be routine: spot illustrations, covers for smaller publishers, background art, marketing images — the mid-range commissions that made a freelance career hold together between the good jobs.

Trade bodies and surveys of creative workers have reported falling commission volumes and downward pressure on rates. The people affected are mostly not famous. They are the ones whose income was assembled out of many small jobs, and small jobs are exactly what a good-enough generated image displaces first.

Nothing further down this page softens any of that. It is the part of the concern that is simply correct, and it belongs on the record before anything gets more complicated.

The most direct thing is also the least technological: buy from artists. A print, a commission, a book whose illustrator is credited on the cover, a small payment to someone whose work you follow. Commission volume is the thing under pressure, so money arriving as a commission is the money that counts most. If your child wants a picture of the character they invented, paying someone to draw it is a better afternoon than a prompt, and they will still have it in ten years.

It is also entirely reasonable to be selective rather than absolute. Plenty of people who will use AI to tidy up a work email will not use it to make a picture, on the grounds that the picture takes something from a person who would otherwise have been paid for it. That is a coherent position. You are allowed to hold it about images without holding a general view about AI, and you are allowed to change your mind as the licensing situation changes.

Compensation at scale has no working design. Every proposal — per-image payment, collective licensing, a levy on model providers, an opt-out register — meets the same arithmetic: an enormous number of contributors, most of them impossible to identify from what was scraped, and a value that cannot be traced back to any one of them. Music took decades to build collective licensing and still argues about the split. Nothing comparable exists for images, and nobody has yet shown that it can be built fairly.

Opt-outs may be worth less than they sound. Several tools now let an artist mark their work as excluded from future training, and some jurisdictions are writing that into law. But a model that has already been trained does not forget, retraining from scratch is enormously expensive, and an artist who opts out today was already inside whatever was built before they did. Whether an opt-out is a real right or a gesture depends on facts that are not yet settled.

The entry-level question is wide open too. Junior illustration work — the small, repetitive, unglamorous jobs — is where people used to learn the craft, build a portfolio and get paid while doing it. If that rung is removed, it is not obvious where the next generation of senior artists comes from, or whether the field reorganizes around something else. Every confident answer to that is a forecast, including the optimistic ones.

So if you finish this and conclude that your family will not use generative image tools, you have reached a legitimate conclusion from accurate information. That is not a failure to understand the technology; it is a judgment about consent and about whose work built the thing, made by someone who now knows what the arguments are. Being able to explain that to another parent is the point. Agreeing with any of it was never required.