Forum
Three Thousand Dollars a Work: AI Training Data Gets a Price

For three years, arguments about AI and creative work have mostly been arguments about principle. In the first week of September, they acquired numbers.
On 5 September, Anthropic agreed to pay $1.5 billion to settle a class action brought by authors whose books it had downloaded from pirate libraries. The proposal works out at roughly $3,000 per book for around 500,000 works, and it still needs the approval of the court. The day before, Warner Bros. Discovery sued Midjourney in federal court in Los Angeles, alleging that the image service lets subscribers generate pictures and videos of characters such as Superman, Batman and Bugs Bunny. It is the third major studio to do so, after Disney and Universal filed a joint suit in June.
One case is about books and the other about images. Both point to the same development: the costs of building and running generative AI are moving from the abstract to the balance sheet.
What the settlement does and does not say
The Anthropic case followed a ruling in June by Judge William Alsup, who found that training on lawfully acquired books could be fair use, but that downloading and keeping pirated copies was a different matter. The settlement resolves claims over those pirated copies. It is not a judgment that training itself requires payment, and it does not set a legal rate.
What it does set is a reference point. Every negotiation over licensing creative work for AI will now take place in the shadow of a widely reported figure per work. That figure was reached for books, under specific legal facts, but people in other fields will quote it anyway.
Why image makers should pay attention
Visual artists have their own lawsuits, and they move at their own pace. But two lessons from this week apply directly.
First, how data was acquired matters as much as what it was used for. The distinction Judge Alsup drew between bought and pirated copies is one that photographers, illustrators and stock agencies will recognize. A dataset assembled from licensed collections is in a different position from one scraped without regard to source.
Second, outputs are becoming their own legal front. The studio suits against Midjourney focus heavily on what users can generate: recognizable characters, on demand. Whatever happens with training claims, companies are under pressure to show they can stop their tools from reproducing protected characters and styles that belong to someone else.
What this means for working designers
Most designers will never negotiate a training license. They do choose tools, and they deliver work to clients who carry risk. That makes a few questions worth asking before a project starts.
- Can the tool’s provider explain where its training data came from in terms a client would accept?
- Does the provider offer contractual protection for commercial use of outputs, and on what conditions?
- Does the workflow avoid generating recognizable characters, logos or living artists’ signature styles?
- Is there a record of what was generated and how it was edited, in case questions arise later?
These are not legal formalities. They are part of professional practice, in the same way that using licensed fonts and cleared photography became routine.
A note for competitions
Awards that accept AI-assisted work depend on the same discipline. The International AI Award rules require entrants to own, or have permission for, everything they submit, and they warn that non-compliant entries may be disqualified. As settlements and lawsuits make the cost of careless sourcing visible, statements like these will be read more closely by organizers and entrants alike.
The shift underway
The size of this settlement will shape behavior far beyond publishing. AI companies now have a vivid reason to license, document and filter. Creative professionals have a reason to understand the value of their own archives. And everyone in between, from agencies to studios to independent designers, has a reason to treat the provenance of their tools as seriously as the quality of their output.



