Peter Henderson

Only a couple of days after my last post, vibe hacking in full force.

Only a couple of days after my last post, vibe hacking in full force.

www.bloomberg.com/news/article...

Following by a panel on GenAI, Agentic AI, Law, and CS (1:15-2:00pm ET) with @peterhenderson.bsky.social (Princeton) and Georgios Piliouras (Google DeepMind)

Spotlight Talks (2:30pm-4:00pm) by
@aloni-bologna.bsky.social (UChicago), Rebecca Wexler (Columbia), and @jubaz.bsky.social (Georgia Tech)

Unfortunately, the scale of the problem makes it challenging. Even if we freeze at Codex-5.3/Opus-4.6-level capabilities attackers can probably scaffold them to pretty easily identify tons of vulnerabilities.

As models discover more exploits, we may need something like a responsible disclosure period for major jumps in cyber capabilities. Before release, trusted defenders get privileged access to the more capable model. Together, they scan for vulnerabilities at scale & notify as many as possible.

Missing from the headline: "using Claude Code."

Vibe hacking is already a thing. I've been saying this for a while, but no model-level safeguards will prevent it entirely. What they can do is slow it down enough for us to put societal-level safeguards in place.

www.popsci.com/technology/r...

That was fast.

New copyright law "hypothetical" just dropped.

Warner Music and AI startup Udio settle copyright battle and ink license deal Warner Music Group has resolved its copyright dispute with Udio and signed a deal to collaborate on a new AI song creation service. Announced Wednesday, this is the second agreement between a major...

www.sfgate.com/entertainmen...

Warner Music and Udio settle their copyright case, agree to collaborate on "new song creation service that will allow users to remix tunes by established artists." Expect more such settlements as copyright holders look to leverage AI to boost revenue!

We’ve been pushing hard on AI for public good. One example: partnering with Courtlistener to launch accessible legal semantic search! Many more cool AI projects coming soon from my group aimed at improving access to justice, often spearheaded by @dominsta.bsky.social !

Sora2 is speedrunning my AI law class. We covered issues with copyrighted characters in week 2, and right of publicity claims in week 3. Georgia has a postmortem right of publicity claim. Some states don't (e.g., famous Marilyn Monroe estate battle).

How Gemini Compute Use Agent feels about the "Choose Chrome" popup.

gemini.browserbase.com

Why might AI companies take on larger copyright litigation risks? If they estimate AGI-scale impacts are 2-3 yrs out, litigation will lag that long. By then, the bet might be: govts step in (too big to fail), rightsholders reliant on AI, fair use prevails, or have $$$ to settle.

Quick take: Are open-weight AI models getting a fair shake in evals? A few thoughts on comparing systems-to-models, sparked by Anthropic’s recent postmortem.

Check it our most recent post: www.ailawpolicy.com/p/quick-take...

GPT-5-codex just ``git reset --hard'' ongoing changes in a repo, saying "I panicked!"

h/t Zeyu Shen @ Princeton

AI "Born Secret"? The Atomic Energy Act, AI, and Federalism A law & policy deep dive.

☢️ Can an AI model be "born secret" when it comes to nuclear and radiological risks? What powers does the Atomic Energy Act give the federal government over frontier models?

It might be more than you think! And may preempt parts of state regs. Check out our post: www.ailawpolicy.com/p/ai-born-se...

Anthropic Settles Its Copyright Litigation—and Why That Was the Right Move As well as what it means for the broader landscape of litigation.

Some quick thoughts on the recent copyright litigation developments:

"Anthropic Settles Its Copyright Litigation—and Why That Was the Right Move"

🔗 www.ailawpolicy.com/p/anthropic-...

Annnnnndddd Judge Alsup just rejected the settlement. Still some time to fix it. Rejection was mostly on the grounds that the class was under-specified (no final list of works, no opt-out/notification mechanism solidified).

news.bloomberglaw.com/ip-law/anthr...

Statutory Construction & Interpretation for AI - CITP Blog Blogpost authors: Nimra Nadeem, Lucy He, Michel Liao, and Peter Henderson Paper authors: Lucy He, Nimra Nadeem, Michel Liao, Howard Chen, Danqi Chen, Mariano-Florentino Cuéllar, Peter Henderson A long...

💡New on the CITP Blog: "Statutory Construction & Interpretation for AI" > What if an LLM concludes a user's behavior is “egregiously immoral" -- & contacts authorities?

CITP researchers with Prof @peterhenderson.bsky.social's
POLARIS Lab provide a possible explanation.🔗👇

The terms of Anthropic's settlement w/book authors just came out.

💰$1.5B to authors in libgen (Books3 corpus)!

Interestingly, this is ~$3k per book, close to the terms that HarperCollins allegedly gave to authors for their books ($2.5k). Consensus price forming?

Work with amazing folks: Lucy He, Nimra Nadeem, Michel Liao, Howard Chen, Danqi Chen, & Mariano-Florentino Cuéllar @carnegieendowment.org

Basically, if we’re going to take model specs/constitutional AI seriously, we need to optimize rules and build out surrounding consistency-enhancing structures, paralleling the legal system.

Let's build better natural language laws and law-following AI together! If interested, reach out!

Obviously, lots more to do in this space! I'm super excited about this direction and the forthcoming work that we're building out.

3️⃣ These computational tools, we think, can also be applied to positive models of the legal system, something that we’re tackling. More on this soon!

2️⃣ We leverage interpretive constraints or ambiguity to induce more consistent interpretations and debug laws for AI. These computational tools allow us to not only build more rigorous laws for AI, but adds a layer of visibility on what can go wrong, ex ante.

A few quick takeaways below, but I’ll drop more findings soon on this dense paper:

1️⃣ Given the same set of rules, models will interpret scenarios wildly differently. This gives us a mechanism to quantify interpretive ambiguity.

We model a space of reasonable interpreters and then modify rules, or add interpretive constraints, to reduce the entropy of the distribution.

Statutory Construction and Interpretation for Artificial Intelligence AI systems are increasingly governed by natural language principles, yet a key challenge arising from reliance on language remains underexplored: interpretive ambiguity. As in legal systems, ambiguity...

Check out our new work, Statutory Construction and Interpretation for Artificial Intelligence, doing exactly this!

Paper: arxiv.org/abs/2509.01186
Policy Brief: www.polarislab.org/briefs/Statu...
Blog: www.polarislab.org#/blog/statut...

Wonder why Claude decided to report users to the authorities? It might be because its constitution says Claude should choose responses in the long-term interest of humanity!

But what if we could leverage computational and legal tools to "debug" or "lint" AI rules/laws for ambiguity?

🧵!