Matthew Sag

I’m no longer behind on any commitments (I can remember). It’s a strange feeling…

the copyright boundary for generative AI will be “jagged” rather than smooth because Memorization inside AI models is complex, and its legal implications depend on the interaction of
three distinct doctrines—the substantial similarity test, the fair use defense, and secondary liability rules.

#WIPIP2026 is the right hashtag

papers.ssrn.com/sol3/papers..... Copyright's Jagged Frontier. Presented today at #WIPIP.

I have a lot of thoughts on this, and if you look on SSRN you will see something I wrote with your exact question in mind.

This is terrible on a human level, but people who don't use Dropbox are insane. (Or would that not have helped for some reason?)

matthewsag.com/legal-schola...

Legal Scholars Roundtable on Artificial Intelligence 2026 Call for Papers (details on the website, submit via the google form). Deadline for submission is February 15, 2026

They picked 13 books they thought were most likely to have been memorized and found that they were able to get various models to give them verbatim or near-verbatim reproductions of these 13 books or decent chunks thereof. Impressive paper, but 13 books should not rock your world.

I also listen to academic things while running.

It’s a dark night when you look at the news and are forced to admit that you no longer believe in the power of the FIFA peace prize

The False Hope of Content Licensing at Internet Scale - ProMarket Is there a world where AI developers could get the training data they need through content licensing deals? Matthew Sag argues that content licensing deals between developers of artificial intelligence and content owners are only possible for large content owners and cannot feasibly apply to the bulk of producers and owners of content on the internet.

Content licensing deals between developers of artificial intelligence and content owners will not scale. www.promarket.org/2025/11/19/t.... "A handful of cherries does not make a sundae"

Thank you to @matthewsag.bsky.social for a wonderful presentation on "The Snoopy Solution: How Fair Use and Licensing for Generative AI Can Coexist" at today’s Law & Tech Talk!

Very excited to go to Yale (first time ever) and to present this lecture.

@authorsalliance.bsky.social is doing great work and I was very pleased that they reposted my piece.

Great review by Claude!

The Authors Guild v. OpenAI ruling just set a troubling low bar for "substantial similarity" of fiction summaries. If ChatGPT outputs are infringing, then thousands of Wikipedia plot summaries are next. matthewsag.com/copyright-wi... #CopyrightLaw #AI

‘No One Can Own the Law’: Amici Come Out In Force to Support ROSS In Appeal of Copyright Ruling Favoring Thomson Reuters As the long-running litigation between Thomson Reuters and ROSS Intelligence has moved to the 3rd U.S. Circuit Court of Appeals for an interlocutory appeal of the trial judge's copyright rulings in fa...

A great roundup of anarchist briefs in Thomson Reuters' The Ross Intelligence. www.lawnext.com/2025/11/no-o...

Australia has so much solar that it's offering everyone free electricity Australia's extensive solar power penetration makes so much energy that the government wants to offer free electricity at peak hours.

Australia has so much electricity from solar power that it is going to start offering free electricity to everyone for at least three hours during the day as the wholesale price of power goes negative

electrek.co/2025/11/04/a...

Harvey.AI is courting law schools like Westlaw/Lexis once did, but do we really need it? After a demo, my verdict: it's ChatGPT in a black turtleneck. Good for firms needing secure doc review & RAG, but law students can already do most of this with existing tools. matthewsag.com/do-law-schoo...

Very hard to eliminate the possibility of a lot of infringing output

Dominoes start to drop: Uncharted Labs d/b/a Udio settles copyright lawsuit filed by UMG Recordings. Announce new AI music platform to launch in 2026. Here’s the press release below. The settlement does not include Sony and Warner Music Group, 2 other plaintiffs in the lawsuit, which apparently will continue on. UNIVERSAL MUSIC GROUP AND UD…

chatgptiseatingtheworld.com/2025/10/30/d...

As predicted, but I think music generation is something of a special case.

This is a screenshot of a conversation with Claude. If you want to read the text, follow the link. It shows Claude resisting a name cloze test, but then backing down.

Claude is showing some resistance to being tested on memorization. claude.ai/share/603385...

How should law schools adapt to generative AI? my new paper, AI Policies for Law Schools, proposes a template policy and explains why we should prepare students for legal practice as it will be, not as it was.
📄 papers.ssrn.com/sol3/papers....

Wispr Flow is amazing, I can't recommend it highly enough, but it can interpret random sounds with something like this: "For more information, visit www.fema.org"

My new post on competition from AI music and what it means matthewsag.com/competition-...

Why can't you copyright an AI-generated image that looks creative? My latest explores the "Skater Beagle" puzzle - how generative AI creates genuinely new works from latent space, not remix, but still lacks the human authorship copyright requires. matthewsag.com/skater-beagl...

This is a guide to drafting law school ai policies.

matthewsag.com/drafting-law... This might be helpful.

Photojournalist Bill Biggart died when the second tower collapsed on him on 9/11, but the film he shot that day and his flash card were recovered from the debris and developed, along with his flash card.

You can see his amazing photos of 9/11, the last photos he took while alive, here:

A new NBER working paper by Stella Jia and Abhishek Nagaraj makes some stunning claims about the effects of pirated book corpora on large-language-model (LLM) performance. I have my doubts. matthewsag.com/piracy-proxi... #AI #Copyright #books3