A. Feder Cooper

This is a really great community of researchers, and every accepted paper gets a generously long talk slot to present.

CFP: computersciencelaw.org/2026-2/2026-...

Main track deadline (archival and non-archival): September 30, AoE

The NeurIPS position track didn't take a large number of extraordinary papers that surpassed the acceptance bar, limiting the acceptance rate to an unusually low 6%.

If you have a rejected paper at the intersection of ML and law, consider submitting to ACM CSLaw '26.

Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy, Research, and Practice We articulate fundamental mismatches between technical methods for machine unlearning in Generative AI, and documented aspirations for broader impact that these methods could have for law and policy. ...

Our paper "Machine Unlearning Doesn't Do What You Think" was accepted for presentation at NeurIPS

Congrats @afedercooper.bsky.social and @katherinelee.bsky.social, who led the effort

arxiv.org/abs/2412.06966

One more week to submit to CSLaw '26!!

at least 100k with all the appendices 🙃

AI Copyright Lawsuits with Pam Samuelson | Scaling Laws

For an update on the state of play in the generative AI copyright cases, try this podcast: shows.acast.com/arbiters-of-...

15 days left to submit to the CSLaw '26 main track! (archival and non-archival)!

The WSJ carelessly spread anti-trans misinformation The Wall Street Journal’s fuckup while covering Charlie Kirk’s killing needs more than an editor’s note.

did a little media criticism www.theverge.com/politics/777...

was just looking for @seantcollins.com’s “goofy at the crucification” post and google is so cool now

Generative AI can be like a search engine, a website, a library, an author, or like any number of other things copyright has a well-developed framework for dealing with.

Prematurely accepting one of these analogies to the exclusion of the others would mean ignoring numerous relevant similarities

5. Generative AI does not make the ordinary business of copyright law irrelevant:

Courts will still need to make plenty of old-fashioned, retail judgments about individual works.

6. Analogies can be misleading: Generative AI systems blur the boundaries between things that were formerly distinct.

4. Fair use is not a silver bullet:

Generative AI scrambles past assumptions about ML and fair use. Some generations will infringe, and that could impact the fair use analysis at previous stages of the supply chain.

3. Design choices matter:

Every actor in the generative-AI supply chain is in a position to make choices that affect their copyright exposure, and others'.

2. Copyright concerns cannot be localized to a single link in the supply chain:

Decisions made by one actor can affect the copyright liability of another, potentially far away actor in the supply chain

Part III provides broader lessons. They (and the piece in general) have held up really well, in spite of how fast this landscape is changing:

1. Copyright touches every part of the generative-AI supply chain:

Every stage from collecting training data to alignment can make use of copyrighted works

Part I has the relevant background on ML, generative AI, and the complex, interconnected supply chain involved in the design, construction, deployment + use of generative-AI models & systems

Part II goes through the very many touch points between US copyright law and the generative-AI supply chain

We worked really hard to make this piece accessible and useful to both copyright and ML experts, alike.

And we've been thrilled to hear over the last two years how our paper has helped teachers, journalists, and other interested readers better understand these issues.

TALKIN' 'BOUT AI GENERATION: COPYRIGHT AND THE GENERATIVE-AI SUPPLY CHAIN | The Copyright Society We know copyright

After 2 years in press, it's published!

"Talkin' 'Bout AI Generation: Copyright and the Generative-AI Supply Chain," is out in the 72nd volume of the Journal of the Copyright Society

copyrightsociety.org/journal-entr...

written with @katherinelee.bsky.social & @jtlg.bsky.social (2023)

The Bartz v. Anthropic settlement is the polar opposite of the Google Books settlement: a discrete one-time payment for past copying, on a discrete and closed-ended class, and making no attempt at all to deal with a larger forward-looking issues.

not again…wrote a (really) long law review paper that started as a short response to some wildly inaccurate comments he made in a post about copyright and language models reproducing training data they’ve “memorized.” It’s 2 years later and that post still isn’t fully successfully debunked 🫠

But it’s really expensive in others. E.g., NeurIPS ‘25 is slated to reject 300+ papers that ACs voted to accept bc of physical resource constraints. This is all bad for science, and will hit junior scholars the hardest.

(I also have no idea to responsibly recruit students in this ecosystem.)

Also don’t think acceptance at these venues is a sign of quality anymore. Great, influential, unpublished arxiv paper seem to mean more. They’re dealing with the same slop + scale problems, without enough qualified reviewers. Drawing arbitrary lines is a “cheap” way to solve them in some sense.

oh, not all of this evidence is ML experiments.

But I agree in general. I also don’t think I’d know how to solve this problem, but this particular line drawing seems like it could go not so great for some really good work (like @mariaa.bsky.social’s). It’ll be a wait and see.

my current understanding @jacyanthis.bsky.social is this is targeting a new flavor of paper. A lot of good positions at ICML have supporting evidence. versions of these papers (even if not accepted) are arguably typical scientific papers (which often have positions), could go up as normal preprints?

(Having done some work on this, I think this type of clean room practice isn’t actually super feasible in practice. Even if you get curation like this right—a huge if—I’m finding that these types of counterfactuals are really brittle in practice)

Definitely not imo. Makes sense not to rely on them. (Tbh, I don’t really get what the conferences are going for here. I had a NeurIPS position paper bounced out at at desk rejection for having too much science in it. So did a buddy of mine. I wish I were kidding.)

Oh yes, that makes a lot of sense. So it’s not necessarily the case that this work will all be modded out? It’s just that a subset of things that seem like they fit this new “position paper” pattern will? (I can also wait to read your blog post.)

I appreciate the enormous amount of work arxiv folks put in to manage the system and the sheer (accelerating) volume of paper. My concern is that there won’t be a robust policy for dealing with false positives, which can have serious consequences for interdisciplinary work (esp. for junior people)

Not to mention the enormous amounts of ML paper slop on arxiv, stuff that is full of stuff that is just completely wrong, but has enough stuff that resembles math or experiments in it.

My point is, this seems like an arbitrary line to draw about what is good enough for arxiv

What happens to the CY computing papers (that aren’t submitted to these extremely noisy venues) involving rigorous methods from social sciences, law, etc. that some computer scientists think are “position papers” bc they involve methods + skills they don’t recognize/ aren’t equipped to evaluate?