Anthropic’s $1.5 billion copyright settlement, announced just weeks ago as a landmark resolution to claims that the AI company trained its models on pirated books, is running into trouble. A federal judge has delayed approval after objections surfaced that the deal might shortchange authors while making their lawyers very, very rich.

The settlement was supposed to resolve a class action brought by authors who alleged Anthropic scraped their copyrighted works without permission to train Claude. Under the proposed terms, Anthropic would pay $1.5 billion over several years, with $320 million of that going to the plaintiff lawyers as fees. That’s a 21% cut, which isn’t unusual in class actions. What is unusual is how little individual authors might actually see.

The math doesn’t work for most authors

Critics of the settlement point out that the payout structure heavily favors a small group of bestselling authors while leaving everyone else with token amounts. If you wrote a midlist novel that ended up in Anthropic’s training data, you might be looking at a check for less than $100. Maybe much less, depending on how many class members file claims.

Meanwhile, the lawyers who negotiated the deal stand to collect $320 million in fees regardless of how many authors actually benefit. That’s raised questions about whether the legal team was more interested in securing their payday than maximizing recovery for the class.

Several authors have filed objections arguing exactly that. They claim the settlement was rushed through negotiations without adequate discovery, meaning nobody actually knows the full scope of what Anthropic used or how valuable those works were to the company’s model training. Without that information, it’s impossible to know if $1.5 billion is reasonable or if Anthropic got off cheap.

The judge pumps the brakes

The presiding judge appears to share some of those concerns. Rather than rubber-stamping the settlement at a scheduled fairness hearing, the court delayed approval and ordered additional briefing on the fee arrangement and the adequacy of class representation.

That’s a problem for everyone involved. Anthropic presumably wanted this case behind them, especially as other AI companies face similar copyright litigation. The plaintiff lawyers want their fees. And many authors, even those who think the settlement is inadequate, would probably prefer some money to years of additional litigation with no guarantee of a better outcome.

But the judge’s role in a class action isn’t to make everyone happy. It’s to ensure the settlement is fair, reasonable, and adequate for the class as a whole. And when the lawyers are walking away with $320 million while the median class member gets a nominal payment, that’s a hard standard to meet.

What happens next

The court has given objectors until early June to submit detailed briefs explaining why the settlement should be rejected. Anthropic and the plaintiff lawyers will then have a chance to respond. A new fairness hearing will be scheduled after that, probably sometime this summer.

If the judge rejects the settlement, the case goes back to square one. That could mean more discovery, more motion practice, and potentially a trial. It could also mean Anthropic might be willing to negotiate better terms if the current deal falls apart.

More likely, the parties will try to salvage the settlement by restructuring the payout formula or reducing the legal fees. Courts have significant power to modify fee awards even when they approve the underlying settlement. Cutting the lawyers’ take from $320 million to, say, $200 million would free up an extra $120 million for authors without Anthropic paying another dime.

The objecting authors have a harder path. To blow up the entire settlement, they’d need to convince the judge that the deal is so fundamentally flawed that no modification can save it. That’s a high bar, especially when many class members probably just want their money.

Why this matters beyond Anthropic

This case is being watched closely by every AI company training models on copyrighted content, which is to say every AI company. OpenAI, Meta, Google, and others all face similar lawsuits. If Anthropic’s settlement gets approved despite the objections, it sets a template for how much these cases are worth and how they get resolved.

If it falls apart, it signals that courts won’t accept sweetheart deals that prioritize legal fees over actual compensation to rightsholders. That could push other cases toward trial, which would finally give us judicial clarity on whether training AI models on copyrighted works is fair use. Nobody knows the answer to that question yet because none of these cases have reached a verdict.

For now, the authors objecting to Anthropic’s settlement are betting that a judge will force a better deal rather than let the lawyers run away with the bag. We’ll know by summer whether that bet pays off.