$1.5 billion. In any copyright lawsuit, that's a number that makes you blink. But in Anthropic's case, some say it's not enough.
The judge has delayed approval of the settlement. Not because the amount is too small — but because how the money is distributed has become the problem.
Who's Objecting
The settlement was reached between Anthropic and a group of writers and publishers over the use of copyrighted works in training Claude models.
Now, some authors are pushing back. Their argument isn't "don't settle" — it's "this split isn't fair."
Specifically, the current payout structure is criticized for lack of differentiation — authors whose works were heavily used by Anthropic and those who were only "possibly" used receive similar compensation.
It's like a restaurant that bought 100kg of beef and 10 grams of beef, then paying both suppliers the same amount because "they both sold beef."
Why the Judge Delayed
The judge's delay in approval signals that the court also sees problems with the distribution scheme. This isn't a routine postponement — it's a substantive questioning of the settlement terms.
Legally, class action settlements must go through a "fairness hearing," where the judge must confirm the settlement is reasonable for all affected parties. If there are enough objectors with strong enough arguments, the judge can reject the approval.
In this case, the objectors' core argument is: the settlement doesn't adequately reflect the degree to which different authors' works were used.
What This Signals
This case reveals a deeper shift: AI copyright disputes are moving from "should they pay" to "how should they pay."
In the early stage, the debate was whether AI companies were infringing at all. Now, Anthropic has agreed to pay $1.5B, which suggests the "should they pay" question has a practical answer — at least for Anthropic, paying is cheaper than litigating.
But "how to pay" is the real challenge.
AI model training data involves trillions of tokens. Precisely tracing every token back to which book, which author, how many times it was used — that's nearly impossible technically. So any payout scheme can only be a rough approximation.
And rough approximations mean someone will feel it's unfair.
Impact on Other AI Companies
This case isn't an isolated legal event — it's an industry benchmark.
If the judge ultimately approves this settlement (even with modifications), other AI companies (OpenAI, Google, Meta) will have a reference template — $1.5 billion might be the price anchor for resolving similar copyright disputes.
But if the settlement is rejected, or the payout is required to increase significantly, the entire AI industry's copyright risk exposure needs to be recalculated.
Anthropic's Position
For Anthropic, the delay isn't good news.
On one hand, the $1.5B cash outlay is already a heavy financial burden (though reportedly much of it is in the form of AWS compute credits). On the other hand, ongoing legal uncertainty means the company's copyright risk hasn't truly been cleared.
More troubling: Anthropic just announced a new $10B compute deal with AWS and is expanding aggressively. If copyright litigation remains unresolved, investors and regulators may reassess the risk profile of that transaction.
My Take
The most interesting part of this case isn't the law itself — it's the structural problem it exposes: the copyright ownership and compensation mechanism for AI training data fundamentally doesn't fit the traditional copyright framework.
A book used to train a model and a book reprinted in a magazine are the same type of "use" in law. But in the AI context, these two "uses" have entirely different impacts. Existing copyright law isn't prepared for this distinction.
Eventually, a completely new mechanism may be needed — perhaps an AI training data licensing pool, a usage-based royalty system, or some form of collective management organization.
Until then, disputes like this will keep coming. $1.5 billion might just be the beginning.
Primary source: Ars Technica reporting