Anthropic Has a Copyright Problem

Tuesday was a bad day for the AI company Anthropic.

The company published the now-deleted 2.1.88 release of Claude Code, its AI coding tool. But with the intended package came a nearly 60 MB file that contained over 500,000 lines of source code. According to Anthropic, this was the result of “human error” and that they were taking steps to prevent it from happening again.

However, for many programmers, once was all it took. They began pouring through the code and exposed many of Anthropic’s secrets, including a “Tamagotchi-style pet” and how the tool tracks user frustration. 

But most damning was the implementation of an “undercover mode” where Claude will not reveal itself as an AI agent. Though this feature was inactive, its presence at all was deeply disturbing to many.

However, it was what happened next that really attracted the headlines. Anthropic began filing copyright takedowns, mostly with GitHub, to try and contain the leak. Initially, the takedowns targeted an “entire network of 8.1K repositories” but were eventually scaled back to just one repository and 96 fork URLs.

But, even with the scaled-back version of the takedown, Anthropic has a myriad of copyright-related problems. The biggest of which is that it’s unclear if this code can be protected by copyright at all.

Who (or What) Wrote the Code?

Right off the bat, there’s a very serious question: Is Anthropic’s code protected by copyright at all? 

The reason is fairly simple. According to both the US Copyright Office and recent court decisions, wholly AI-generated works are not eligible for copyright protection. That’s because, under US law, only humans can hold copyright protection, and if an AI is making the creative decisions, there is no copyright protection.

To be clear, we have no idea how this code was written. Anthropic, as a company, does employ human developers, and they likely had some hand in this code.

But it would also be odd for Anthropic, a company that heavily promotes its product to developers, to have not generated some or all of it. It’s very likely that some of the files were wholly AI generated, some were written by humans, and others were a hybrid of the two.

It would be a tremendous and tedious mess to go through this code and determine what is and is not protected by copyright. Still, it seems likely that at least some of the code being targeted is not protectable. 

This makes the broad DMCA takedown extremely dubious legally.

But then comes the next question: why does Anthropic suddenly care so much about copyright at all?

Copyright for Me But Not for Thee

When it comes to copyright, Anthropic’s reputation is less than stellar.

Like other AI companies, Anthropic used other people’s intellectual property to train its AI systems. When sued, they fought tooth and nail to have that behavior declared fair use and even won a victory on that point in California. 

However, even with that victory, they reached a $1.5 billion settlement over their use of pirated books, agreeing to pay authors up to $3,000 per work.

Still, the debate over AI training is still undecided, and there will likely be many more rulings in many more countries before any kind of norm is settled upon. All that said, it’s impossible to ignore that Anthropic has been fighting hard for the right to use anyone else’s content, including their code, as they see fit.

So, when their code gets leaked, it smacks of irony for them to aggressively try to protect it using copyright law. However, the very arguments that they have made and the very acts that they have enabled are going to make that impossible.

First off, with their arguments, there’s nothing to stop other AI companies from using this code for training. If AI training, even without permission, is fair use, then there’s not much they can do. All they can really demand without falling into hypocrisy is a pittance if it is determined the code was pirated.

But then there’s the twist: many are using Claude itself to translate or rewrite the source code. As we saw with the recent Malus project, AI tools can quickly and effectively rewrite code in a way that generates a “legally distinct” work. 

So, even if their code is protected by copyright, their arguments and their own tools make enforcing those rights nearly impossible. As one of the company, Systima, said, “The source is, for all practical purposes, permanently public.” 

To that end, Anthropic only has itself to blame.

Bottom Line

To me, this brings back memories from January 2025, when DeepSeek made its debut and OpenAI accused it of distilling their models. The schadenfreude ran deep as many lined up to mock OpenAI for complaining about the very thing they had done to human creators.

Anthropic has been eager to flout copyright law. Whether you feel AI training is fair use or not, there is something galling about taking the work of millions of humans and turning it into a product that is designed to replace them. It’s even more galling to then run to copyright when it’s your work that’s being misused.

Anthropic is getting a taste of the world that it helped create. While this probably will be nothing but a footnote in the story of AI, it’s still a very telling moment.

Anthropic wants all the benefits of copyright protection but none of the responsibilities. It wants to profit from the entire history of human creation, but protect its own intellectual property.

Whether they can do that legally or not doesn’t matter; it’s still a hypocritical stance to take and one that is being rightly mocked. 

The sad part is that it’s unlikely anything will be learned or will be changed because of it.