Paying artists for training data is a corporate peace offering that solves the wrong problem. The industry is trying to treat a fundamental shift in the nature of creative labor as a simple licensing dispute, as if we can just slap a price tag on the essence of an artist’s identity and call it a day. It is a sanitization effort (mostly to avoid a class-action nightmare) designed to make the vacuuming of the internet look like a standard B2B transaction. By framing this as a payment issue, the labs are successfully distracting us from the fact that they are essentially selling the tools to replace the very people they are now offering to pay.
The idea of royalties for generative AI is a logistical joke. We are talking about models that blend billions of parameters into a latent space; the influence of a single illustrator on any specific output is a mathematical ghost. To suggest a royalty system implies we have a way to track attribution at the token level, which we simply do not. How do you actually quantify the contribution of one illustrator among five billion samples? Who decides the unit price for a “style”? If a model is trained on ten thousand artists to produce a generic “digital art” look, does each person get a fraction of a cent every time a user prompts for “pretty forest”? The friction here isn’t just about the money—it’s about the fact that the accounting for this would be a nightmare of GPU-heavy auditing that no lab actually wants to implement.
It is a bit like paying a master chef a one-time fee to let you record every flavor profile they’ve ever created, then using that data to build ten thousand automated kiosks that mimic their signature dishes perfectly. Sure, the chef got a check, but they didn’t sell a product—they sold the ability for the market to render them obsolete. The payment isn’t a partnership; it’s a buyout of their future relevance. You aren’t paying for a license to use the work; you are paying for the right to automate the person.
The power dynamic is skewed entirely toward the labs. As The Verge points out, artists have spent years sounding the alarm about theft, and the response has been a slow pivot toward “well, what if we just paid you?” This isn’t a gesture of goodwill. It is a strategic move to create a legal shield. If a lab can point to a handful of licensing deals, they can argue in court that a “market rate” for training data exists, which effectively kills the argument that the training was an unauthorized appropriation of intellectual property. Once a price is established, the conversation shifts from “you stole my work” to “you underpaid me,” which is a much easier fight for a corporate legal team to win.
The irony is that the developers building these tools often pride themselves on transparency and open-source ethics, yet the training sets remain black boxes. We are asked to trust that the royalties are fair without seeing the weights, the training logs, or the actual data curation process. It is a classic “trust us” move from companies that have spent the last two years scraping every public pixel without asking. (Or maybe they just don’t care—see below.) We’ve seen this pattern before in the music industry, where streaming payouts became a way to pacify artists while the platforms captured all the actual value.
By Q4, we will see the first major licensing deal between a top-tier AI lab and a creative collective collapse because the royalty math simply doesn’t scale. The labs will realize that paying for every single influence is too expensive to maintain their margins, and the artists will realize that the payments are a pittance compared to the loss of their livelihoods. The middle ground doesn’t exist because the two sides are arguing about different things: one side wants a fair wage, and the other wants a legal indemnity waiver.
The royalty model is a distraction.
