Post

The action around OpenAI's case and, uh, fretting over mass extinction or whatever, might have distracted us, but the Doe v. Github AI case had a meaningful ruling last week. It's the case about MSFT Co-Pilot code-writing AI, with claims that focus on outputs.

We've been watching the 9th Cir to see what they would say about "Copyright Management Information" (CMI), a weird quasi-copyright thing from the DMCA that may be a dark horse candidate for most stupid copyright law to apply to AI (a tight race!).

As part of its package of anti-piracy measures, the DMCA includes a provision that says you can't remove CMI from a copy of a protected work. The goal was to penalize pirates who would strip out watermarks or other markings that rightsholders use to indicate their ownership.

It's stupid to apply this rule to AI training and other computational fair uses because removal of CMI is an important part of non-consumptive textual analysis. It's got nothing to do with facilitating piracy - the copies never see the light of day!

As I pointed out in a recent blog post, even AI critics strip CMI as part of their textual analysis work. It's just best practice in computer analysis, whether for AI training or for other purposes. recreatecoalition.org/without-fai...

Happily, the 9th Circuit avoided disaster, siding with Github, @EFF and other amici who had urged the court to be reasonable in applying the law about CMI. While the court avoided deciding the core question about training, it repudiated the Does' dumbest argument.

As @joemullin explains, the court recognized that creating a new work with genAI is not the same as making a copy. So when a genAI tool outputs code that might include a snippet that matches the training data, that output is not a "copy" that has CMI "removed" - it's a new work.