Peter Henderson
@peterhenderson
Assistant Professor, leading the Polaris Lab @ Princeton ( Researching: RL, Strategic Decision-Making+Exploration; Law
To evaluate our system we collected a new benchmark of representative queries, with the help of real appellate-level OPD attorneys—we're immensely grateful for their time! These queries are often different from existing academic benchmarks and help make evals more real!
Before doing this we conducted structured interviews to identify concerns. One identified risk related to hallucinations. As a result, we focused on an extractive-first system, with optional generative summaries for only the facts of the case to situate the retrieved content.
Until now, finding the right material across that collection could take hours. The AI-enabled Resource Library lets NJOPD staff enter a legal question and quickly find relevant passages from a secure, closed collection.
Public defenders manage large caseloads and tight deadlines. They also rely on a wealth of shared knowledge: past briefs, memos, directives, and lessons from colleagues about what has—and hasn’t—worked in court. Here's an example of an AG directive memo.
🚀Last week we announced the launch of a new AI tool in partnership with the New Jersey Office of the Public Defender! AI can empower civil servants and improve public services, and this project provides a blueprint for doing so responsibly. 👇Lots of takeaways, below!
Btw, we actually covered almost exactly this scenario in our piece a few years back. But: negligence and products liability are (maybe) shifting—at least for co's that take few safety precautions—so there may be some updating needed. www.journaloffreespeechlaw.org/hendersonhas...
This is wild, but really great investigative journalism. I guess all the jailbreaking papers trying to prevent "How do I build a bomb?" queries have been vindicated... So much for it being an unrealistic scenario... www.nytimes.com/2026/07/10/u... casp.ac/reports/ai-e...
"Altman told staff that the government would be 'approving access customer by customer during this preview period'" Opaque de-facto licensing regimes are a recipe for corruption. Not sure this is the best place to have landed on AI governance... www.theinformation.com/articles/tru...
A surprisingly detailed accounting of how Grok is used in DoW, presumably along with other models in Maven. Maven workflows "enabled" US forces to hit 2000 targets in Iran over 96 hours. storage.courtlistener.com/recap/gov.us...
Forcing people to leave the country to transfer to a green card, will hurt families, communities, and United States innovation. If you aren't aware, adjustment of status can take many months—if not years. I hope this policy is reconsidered.
if you try to get Claude to speak Armenian it just outputs "delays"! Seems like glitch tokens are still unresolved. Interesting (kind of sad?) to see Opus thrown into a loop.
This is a challenging legal problem for NeurIPS (and other conference participants)! You might be wondering how this is possible given the First Amendment? I wrote a quick explainer on the current status quo of relevant First Amendment cases & law to get you up to speed. 🔗👇
I feel this urgency too. But this is all so utterly avoidable with good policymaking. No one should be left behind because they didn't accumulate capital in 2026. There are so many people who aren't plugged into these conversations or are simply not in a position to do anything about it.
I’m really excited about our new paper! I think we will ultimately need to draw on expertise from both law and AI to get alignment right, and this paper lays out that vision in more detail. arxiv.org/abs/2601.04175
Many legal scholars talk about lock-in effects for LLMs from the conversation history/memories (akin to social media). But if an LLM can access the info & is capable, you can just ask the LLM to give it to you, making it far easier to switch providers than social media. Good example of that here.
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...
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...
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
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
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...