Earlier this month, Anthropic reached a settlement agreement with a group of authors who had sued the company for using their works to train its AI models. The copyright cartel celebrated the deal, and some suggested it was proof that AI training requires a license. That framing obscures a key fact: Anthropic had already won on the question of whether training AI was a fair use. So what’s up with this settlement? Here are 5 things you need to know:

1. 1. It has almost nothing to do with fair use and AI training. Judge Alsup was clear that AI training is fair use. He found that training AI models is “spectacularly transformative” and further determined that acquiring print copies on the used book market and digitizing them for the purposes of AI training is also fair use. That ruling was not impacted by the settlement. 
   2. The settlement doesn’t “set a precedent” for anything, but especially not for licensing AI training. This private settlement cannot bind anyone but these two parties, and Judge Alsup ruled that a market to license AI training is “not one the Copyright Act entitles Authors to exploit.” 
   3. The settlement is really about data sourcing. The class action was only allowed to move forward because of how Anthropic sourced some of its training data: from so-called “shadow libraries,” which make digital copies of copyrighted books freely available online. The settlement is only for books that were sourced in this way. 
   4. Copyright’s exorbitant statutory damages provisions were likely much more important than the merits in driving Anthropic to settle. A defendant who loses a copyright infringement case can be compelled to pay anywhere from $750 to $150,000 per work infringed. Multiplied by a class of up to 500,000 works, this creates a potential death sentence even for a leading company like Anthropic. The amounts involved in this settlement are thus not a reflection of the true market value of copyrighted content in the age of AI, but rather a windfall to these plaintiffs extracted in the shadow of draconian statutory penalties.
   5. Even the data sourcing question at the heart of the settlement remains wide open. Within days of Judge Alsup’s ruling excoriating Anthropic for using shadow libraries, another federal court ruled exactly the opposite. In Kadrey v. Meta, Judge Chhabria reasoned that fair uses are always unauthorized, so the use of unauthorized copies should be permitted when the ultimate purpose is fair. The plaintiffs in the Anthropic case were aware of this opinion, and knew they might end up losing on appeal—that’s why they settled. If the case was a slam dunk for them, they could have continued to litigate in pursuit of those sky-high statutory damages.  

Mistaking this one-off private settlement for a blueprint for AI licensing requires twisting the facts and ignoring the actual rulings in this case. It’s understandable that companies and policymakers are hungry for guidance on these important issues, but they should look to the rulings themselves for that guidance, not to private arrangements extracted under threat of heavy penalties.