3 Problems with the Anthropic Settlement

Over the weekend, news broke that the AI company Anthropic reached a $1.5 billion settlement in a class action lawsuit filed against it by authors.

The settlement, if accepted by the court, would cover approximately 500,000 books and pay authors roughly $3,000 for each allegedly infringed work. Anthropic does not deny any wrongdoing, but touts that it would be the “largest copyright recovery ever.”

On the surface, those numbers are eye-watering. Anthropic is right to highlight this as the largest copyright recovery ever and $3,000 per infringed work is not a small amount.

The settlement also does not give Anthropic the right to use pirated books moving forward. It requires Anthropic to “destroy” the copies it has in its possession, and that any infringements past this settlement could result in a new lawsuit.

But, despite those significant victories, there are a lot of reasons for authors to be skeptical. Not only will the vast majority of authors not benefit from this settlement, but it could set a precedent that would greatly benefit large AI companies.

In short, Anthropic is just settling past alleged wrongdoing. It’s making a strategic purchase, and it’s one that they know will benefit them for a long time to come.

To that end, here are just three of the bigger problems with the settlement.

1: It’s About Piracy, Not Training

Previously, US District Judge William Alsup, the judge in the case, had ruled that using copyright-protected material to train AI systems was a fair use. However, that decision remains and continues to be controversial. Even the US Copyright Office disagrees.

However, with that element of the case set aside, Alsup had ordered a trial to determine Anthropic’s liability for the pirating of copyrighted books.

Anthropic, like many other AI companies, had used data sets such as Library Genesis (LibGen) and Pirate Library Mirror. This settlement pertains solely to the alleged piracy. It says nothing about how the books were actually used.

In short, this settlement isn’t about AI; it’s about piracy. It’s actually closer to the Internet Archive settlement from December 2024 than it is to anything related to AI.

2: Who is (and is not) Included

Ultimately, this settlement only applies to a very small group of authors. Though 500,000 works is a lot, it’s only a tiny fraction of those who were actually injured.

First, this only applies if your books were downloaded as part of one of the pirate datasets. If your work was scraped from your website, obtained through a partnership or accessed in any other way, this does not apply to you.

However, according to Judge Alsup, over 7 million digitized books were illegally downloaded by Anthropic. So why is the settlement class roughly 1/14th of that?

The reason is simple. Only works that were registered timely with the US Copyright Office qualify for this settlement. If your work is not registered with the US Copyright Office, you are pretty much out of luck.

But, even if you did, you might still not be involved. You must have either registered it within three months of publication or within five years of publication, and before either June 2021 or July 2022, depending on the library at issue.

I have long been a critic of the registration requirement in the United States. As a result of that requirement, millions of authors are being denied the opportunity to participate in a settlement. There’s no clearer demonstration of how unnecessary this requirement is and how harmful it is to creators.

Worst of all, the authors who suffer the most are international authors and independent ones. They are the ones least likely to have registered their work in a timely manner.

Even if we (like the settlement) just look at the piracy aspect, the vast majority of authors will receive nothing for their works being infringed. Even if you think $3,000 is adequate, the vast majority of creators won’t see it.

3: Legal for a Price

There’s not much doubt that $1.5 billion is a lot of money. However, just a few days before this announcement, Anthropic announced that it had raised $13 billion on a $183 billion valuation.

Even if Anthropic paid the entire settlement upfront (and not over 2 years, as is the current plan), it would have only been about 12% of its latest fundraising and less than 1% of its current valuation.

Anthropic, despite never being profitable, can easily afford this. Apple, Microsoft, Meta, OpenAI and XAI are likely in a similar position. However, it’s more than the valuation of most AI companies.

This settlement won’t require large AI companies to change their behavior. While Anthropic will have to destroy copies of the books, the training has been done. There’s no mention of Anthropic “unlearning” what it did from those datasets. Training, as Judge Alsup ruled, is a protected fair use.

In short, this won’t be a major setback for Anthropic when it comes to AI training. Their AI models will not be affected by this settlement.

However, it will ensure that smaller companies cannot replicate their success, essentially limiting this form of AI training to those who can pay the piracy tax.

Is There Good News?

To be clear, I’m not blaming the plaintiff lawyers for taking this deal. Their hands were largely bound by a combination of the judge’s previous ruling and the absolutely insane registration requirement.

Without either a successful appeal or a major shift in US copyright law, this is likely the best that they can do.

Instead, the point of this is to curb any enthusiasm over the settlement. Behind the eye-popping numbers, there’s not a lot to get excited about. Unless you meet a slew of very specific requirements, this settlement won’t do anything for you.

Still, there is some good news. First, if your work is in LibGen or other pirate datasets, it will be deleted regardless of whether you’re in the class. It’s a small victory, but one nonetheless.

More importantly, other AI companies, including Meta, OpenAI, and Microsoft, have also utilized these databases. This sets the stage for future settlements with those companies and still more payouts.

However, the most important aspect is that this settlement only addresses Anthropic’s piracy prior to the settlement date and only in the United States. Other lawsuits are examining further issues, and the US isn’t the only country with ongoing AI-related lawsuits.

In short, this isn’t the end of the road. That is, most likely, many years down the line.

Bottom Line

Anthropic did not agree to this settlement because it would or could destroy them. They agreed to it because the plaintiffs were in a weak position, and they knew this deal was heavily in their favor.

As impressive as it is on paper, it won’t help the vast majority of creators and only addresses the issue of piracy, not AI training or output.

Instead, the settlement all but guarantees Anthropic’s long-term survival and puts large AI companies in a much better position when dealing with creators.

A company like Anthropic would not jump to spend $1.5 billion without expecting to reap huge dividends down the road. This wasn’t a move to make things right with creators; it was a tactical decision to ensure the company’s future.

To be clear, I don’t blame either company for reaching the deal. Both sides were simply playing the hand they were dealt.

The problem isn’t with this particular agreement, but the bigger picture. While many legal questions remain unanswered, it’s becoming increasingly clear that new laws will be necessary in the United States to protect creators from exploitation.

Still, this is just the first settlement, and it may not even be approved. There’s a lot more in this space to come.