As the Bartz v. Anthropic settlement works its way through the popular imagination, it's important for folks to keep a few things in mind:
1️⃣ Judge Alsup vindicated Anthropic's use of copyrighted material to train its AI models as highly transformative fair use. 🧵 2️⃣ Alsup ruled that acquiring print copies on the used book market and digitizing them for this purpose is also fair use. There is no cognizable licensing market for AI training separate from the lawful acquisition of data in the first place.

OK, having read the opinion last night and slept on it, I'm less concerned about Judge Chhabria's weird market dilution approach to fair use, and encouraged by his other holdings. First, let me share what I think are the main takeaways, then I'll say why I'm not worried. Meta-takeaway (no pun intended) 1: Judge Chhabria rules in favor of fair use on every mainstream, non-bizarro argument. Training is transformative. The use of pirated copies is largely OK. Using entire works is OK. Market harm due to lost revenue from licensing uses like this are per se irrelevant.

The COPIED Act is back, and it's still hot garbage. It dictates the design of every tech from spell check to Instagram filters AND gives every thin-skinned politician, celeb, and pundit a "don't parody me!" protection spell for everything they publish. commerce.senate.gov/services/files… In a nutshell, it creates a new form of copyright without any of copyright's pesky public interest stuff, most importantly:
- No term limit -- it lasts forever and can be applied retroactively to content that is centuries old, by anyone (not just the author or copyright holder).

Four things to know about the RIAA AI lawsuit, a 🧵:
1) The legal claim is *not* that the AI-generated songs are infringing. The similarities, producer tags, etc cited in the complaints are only meant to prove the AI was trained on songs owned by the labels. It's the training that the labels say is infringement - the copying that happens behind-the-scenes, to teach an AI model the vocabulary of music, the meaning of "rockabilly" or the characteristics of "trap music." The model learns those unprotected ideas by analyzing a lot

The 5th Circuit gave #FairUseWeek a fitting climax on Friday by affirming a district court opinion tossing troll Keith Bell’s lawsuit and ordering him to pay attorneys’ fees for the public school he sued over a tweet. This is great news for so many reasons. 1/x First, it continues to build the dossier against Bell, whose disgraceful shakedown scheme has cost public schools and non-profits tens of thousands of dollars in needless settlements. As more and more courts see him for what he is, more victims will be emboldened to defy him.

Hi all! Thrilled to give a Twitter-ized version of the overview talk I sometimes give about the Code of Best Practices in Fair Use for Software Preservation. Thanks to @fletcherdurant for inviting me. Read the full Code here: arl.org/focus-areas/co…. 1/15 Software preservation is a really important part of digital preservation, and its value is only going to increase. Almost everything we make, now, is written in a language only computers running the right software can read. #prestc19 2/15