
law&ai, mostly legal evals | PhD'ing in law visiting fellow @stanfordlaw prev research @cambridge_uni, mle @freshfields
Bay Area
Joined August 2022
- Tweets30
- Following173
- Followers1.1K
- Likes207
Pinned Tweet
We have discovered a substantial number of vulnerabilities in the German legal system using LLMs.
These may enable access to confidential documents, travel without an otherwise required visa, enforceable leave for soldiers at NATO’s JSEC, and more.
Similar to computer systems, the written law can contain vulnerabilities that invite exploitation. In our blog post, we share a selection of vulnerabilities, and present potential exploits for each. To our knowledge, none of these have been previously recorded. We share only those vulnerabilities whose exploitation we believe would not cause serious harm.
blog post: felixringe.com/legal-exploit…
Thank you all for the overwhelming response! I'm glad so many of you took the time to share your thoughts. It's getting very late here on the West Coast, so I'll sign off for tonight, but I'll be back tomorrow!
xitter.cf/felixringe/status/2115…
We have discovered a substantial number of vulnerabilities in the German legal system using LLMs.
These may enable access to confidential documents, travel without an otherwise required visa, enforceable leave for soldiers at NATO’s JSEC, and more.
Similar to computer systems, the written law can contain vulnerabilities that invite exploitation. In our blog post, we share a selection of vulnerabilities, and present potential exploits for each. To our knowledge, none of these have been previously recorded. We share only those vulnerabilities whose exploitation we believe would not cause serious harm.
blog post: felixringe.com/legal-exploit…
1
18
1,349
And many thanks to @MeganPHMa, @JulianNyarko, Esmie Hurd, Chijioke Mgbahurike and everyone else at liftlab at @StanfordLaw who gave feedback and comments on an earlier draft!
35
28,637
felix ringe retweeted
It’s getting hard to build good legal evals. So many legal tasks involve judgment and discretion.
I built LegalForecastBench to test frontier models on a high-level legal task with objective ground truth: predicting how judges will decide real motions.
felix ringe retweeted
AI can do essentially all the technical parts of empirical social science research now---but there's still an essential role for human researchers and universities.
It's been 9 months since we started exploring how AI was going to transform empirical social science research. Here's a periodic update on how my thinking has evolved since then:
(1) The most powerful AI models, including the kinds of multi-agent systems delivering recent math results, are now capable of doing essentially any kind of technical empirical social science research you could desire. I no longer write any data collection, munging, or analysis code directly.
(2) Norms are evolving slower than I expected, but it is increasingly seen as weird if someone posts a working paper without evidence that they have used AI to verify that the results are directly reproducible with their code and data. This is a great sign.
(3) AI has, if anything, gotten WORSE rather than better at writing. As a result, the best papers combine AI-driven empirical analyses with human writing (though AI is quite helpful for things like managing bibliographies, finding relevant web links, etc).
(4) Human direction is still essential for pointing AI at the most interesting and most valuable empirical questions, and for interpreting the results, making sense of them, and communicating them.
We are going to need institutions that are independent of the labs that can verify knowledge claims and direct AI-driven research towards the most fundamental questions of knowledge that AI labs won't be positioned to pursue effectively themselves. This is a tremendous opportunity for places like Stanford, and I'm excited to see how it evolves.
Here's proof that Claude Code can write an entire empirical polisci paper.
To validate my claim that AI agents are coming for polisci "like a freight train", today I had Claude Code fully replicate and extend an old paper of mine estimating the effect of universal vote-by-mail on turnout and election outcome...essentially in one shot.
After careful prompting, Claude Code:
(1) Downloaded the old paper's repo and replicated the past results, translating our old Stata Code into Python
(2) Crawled the web to get updated official election data and census data
(3) Ran new analyses extending the results through 2024
(4) Created new tables and figures
(5) Performed a lit review
(6) Wrote a wholly new paper
(7) Pushed the whole thing to a new github repo
The whole thing took about an hour.
This is an insane paradigm shift in how empirical work is done.
It also validates the point that several people including @BrendanNyhan made yesterday---it's going to be especially easy to scale observational research with AI.
Thanks to @alexolegimas, @arthur_spirling , and many others who gave me feedback. .
I believe it will become standard practice to have AI agents reproduce papers before publication.
My ICML AI4Law workshop paper is now on arXiv with such a reproduction: given only the input data and a description of the methodology, I had agents reproduce both the paper’s reported scores and its benchmark datasets for less than seven (!) cents.
I have laid out the broader case for this approach in a blog post (link below). It also includes a guide to creating such reproductions and discusses some failure modes observed in the agent trajectories (available on @harborframework’s Harbor Hub).
I first came across this idea in @ahall_research's “The 100x Research Institution,” in which he briefly proposes it as one of many ways in which AI might change research and its institutions, and somehow it has stayed with me ever since.
Lastly, thank you to the organisers and everyone who stopped by to discuss my poster at @stanfordHAI’s Empirical Methods in the Age of AI conference this weekend!
blog post: felixringe.com/reproduce-bef…
ICML AI4Law paper: arxiv.org/pdf/2609.26945
The 100x Research Institution: freesystems.substack.com/p/t…
2
535









