University of Pennsylvania Carey Law SchoolWhat This Is
This is Polk Wagner. I'm Deputy Dean for Academic Affairs and Innovation at Penn Carey Law, and I spend a lot of my time on the question this page is built around: what does legal education look like in a world where AI capability is compounding faster than the institutions that are supposed to respond to it.
The next twelve minutes are my framing of the whole course. What's on this page, why it's organized the way it is, and what to do with it.
The Shapiro Moment
Start with Zack Shapiro. In February of 2026, Shapiro, a solo practitioner at a small firm called Raines LLP, published a thread on X describing how he'd rebuilt his transactional practice around Claude. Not Harvey. Not CoCounsel. Not Spellbook. A general-purpose AI, combined with custom skills he'd written himself, doing most of the work a team of junior associates used to do.
The thread got over seven million views as of April 2026. It made Shapiro, briefly, the most talked-about lawyer on the internet. And it changed the conversation about AI in legal practice in ways that matter for people who train lawyers.
Here's why it mattered. Everybody in legal technology had been talking about AI-native law firms in the abstract for two years. Shapiro's thread was concrete. A working practice, represented clients, closing deals, making money — running on a stack that looked nothing like what law firms have looked like for fifty years. The argument stopped being hypothetical.
But — and this is the thing that matters for how you read the rest of the course — Shapiro isn't the median of BigLaw. My read, informed by the firms I talk with, is that the median is twelve to eighteen months behind him. Maybe more. The firms that have AI advisory boards mostly set them up recently. The firms that have deployed a real model-native workflow at scale are the exception.
So there's a gap. The leading edge of practice is moving fast. The median is moving slower. The question Module I of the course is built around is: how fast is the gap closing? The pieces I've curated for Module I let you calibrate that yourself — Shapiro's thread, Matt Pollins's thoughtful interview with him, Seth Chandler on Claude driving Westlaw and Lexis directly, a fresh benchmark showing current models answering bar-exam reasoning at 89 to 93 percent, and Anthropic's own legal team describing how they cut review times from days to hours.
That's the state of legal practice in April of 2026. Whatever else is true about legal education, it has to account for this.
Why Scholarship Lags Practice
Module II is the scholarly frame. And the first thing to say about the scholarship is that it's slower than the models.
A paper posted on SSRN in March 2025, testing o1-preview and a mid-2024 version of Vincent AI, is already describing models that are two generations behind what's in production now. The empirical results are still useful — they tell you what's possible, what the failure modes look like, what trained users can extract when the workflow is designed well. But the specific performance numbers should be read as lower bounds. Current models do more, reliably.
The direction is consistent, though. The empirical work pulled together in Module II points the same way from different angles. Recent models materially improve legal work quality when the workflow is designed well, and materially hurt it when the workflow isn't. The question isn't does AI help anymore. It's: how do we actually teach people to design workflows that extract the value without the hallucination risk?
Module II pairs the empirical work with faster commentary. Ethan Mollick's essay on "working with wizards" is the best short description of what it's like to use current models — the shift from co-intelligence to something more like conjuring. Jesús Fernández-Villaverde, a Penn economist in the School of Arts and Sciences, is working through the implications for higher education publicly on X. And Gregory Duhl's "All In" case study from Mitchell Hamline shows what a full doctrinal course redesigned around AI can look like in practice.
Module II also includes a short landscape summary I've written — where I see most U.S. law schools sitting on AI integration as of early 2026. The short version: most schools are doing something. Very few have a settled position.
One thing to notice about Module II. The paper titled "Can AI Hold Office Hours?" tests three models — GPT-4o, Claude 3.5 Sonnet, and NotebookLM. That's the same tool stack I used to build the NotebookLM on this page. You can read the paper, then chat with the course's NotebookLM, and calibrate directly where reliability actually sits now. That's the kind of thing the scholarship lets you do if you use it as a live instrument.
What We're Building at Penn
Module III is where this stops being diagnosis and starts being action at Penn.
Three levels. At the institutional level, I'm drafting an AI-initiative proposal for Penn Carey Law — organized around four pillars: research, people, building, pedagogy. The bet behind the proposal is specific: rather than predicting where AI takes the legal profession and building a program bet on that prediction, we'd build adaptive capacity — the institutional infrastructure to observe the transformation in real time and redirect as it evolves. That's a different kind of commitment than a traditional law-school center. The concept document is currently under discussion with the Dean; the course materials include a public summary at the right level of abstraction.
At the curriculum level, programs and courses are already integrating AI intentionally. Our Legal Practice Skills program — the 1L required writing curriculum — redesigned itself around a simple sequence: an AI-free opening assignment to establish baseline skill, then structured instruction in AI ethics and use, then three experiential modules where students use AI intentionally. That's AI as pedagogy, not AI as an add-on. The course materials include a summary of the program's approach.
The AI Law Lab is running two intensive bootcamp tracks this spring — one for corporate practice, one for litigation — built around realistic case exercises and guest practitioner sessions. The track separation matters. Corporate and litigation workflows diverge enough that a generic "AI in law" course would shortchange both.
And at the infrastructure level, I've been building tools. Heron — a teaching-assistant bot for my Intro to IP course, which I built entirely with Claude Code. An open-source collection of Claude Code and ChatGPT skills for exam generation, essay exam design, class problem creation, slide review. The repo is called law-faculty-skills. That's the pedagogical analog to what Shapiro built for his practice.
Here's the parallel worth sitting with. Shapiro built a Claude-native law firm. I've been building the faculty-side version at individual scale — Heron, the skills repo. Not in the sense that Claude is teaching the class — the judgment is still mine, and the judgment is still the whole job. But in the sense that the tools carry the routine work so the judgment has more room. That split is the point.
None of the pieces in Module III are finished. All of them are running. The point isn't that we've solved any of this — it's that we've been in motion long enough to have data on what works.
The Open Questions That Matter
Module IV is six questions I don't yet have clean answers to. But before the six, the question they're all angles on.
Law schools have used work — and writing especially — as the central mechanism for teaching for a long time. Cases read, briefs written, memos drafted, exams sat. The work itself was the pedagogy. You learn to think like a lawyer by doing what lawyers do.
AI changes that on both sides. It makes the work cheap to produce, which in principle lets meaningful repetitions rise by orders of magnitude. And repetition is how judgment develops. But AI also does the work — or something that looks like the work — which lets a student outsource the wrestling the pedagogy was trying to induce. The student's role changes from producer of text to supervisor of text. And assessment has to test the supervisory judgment, not just the text.
So here's the deeper question running under Module IV: how do we train lawyers who can 10x their productivity with AI and still maintain the standards of careful, deep analysis the profession has always expected? The answer isn't pick one. It's teach both, in a form where each reinforces the other. What that looks like concretely is the work the next five years of pedagogy design have to do. Here are six angles on it.
First: if AI can generate sophisticated legal reasoning — and current models can, as the benchmarks show — what are we testing on exams? The traditional law-school exam is a proxy for "can this student do the work a junior lawyer is expected to do." If the work is increasingly AI-assisted in practice, the proxy drifts. In my read, most law schools are making that choice by default rather than by design. We should be designing.
Second: does a Claude-native 1L curriculum still have 1L courses as we know them? The 1L year has looked substantially the same for generations. If a first-year student using Claude can produce the memo-and-brief output that the LRW program was designed to produce, the question is what LRW teaches that an AI-assisted student still needs. The answer is probably "the core analytical skills remain essential, and the vehicles change." But saying that is easier than doing it.
Third: what happens to rank-in-class if AI-assisted output is indistinguishable from unassisted? The grade still sorts, but what it sorts on shifts. Firms and clerkship committees will adapt. The schools that lead that adaptation will have more say in what the next generation of sorting looks like.
Fourth: who pays for the tools? Right now, it's uneven. In my observation, students who pay for Pro tiers outperform students who don't on AI-assisted assignments. The answer probably converges toward "the school pays." The timeline on that convergence is unclear, and the cost compounds.
Fifth: how do schools maintain pedagogical identity when every school adopts the same base-model technology? Brand, training, faculty, pedagogy, clinical opportunities — all of those still matter, but we'll have to be more deliberate about them than we've historically needed to be.
Sixth, and this may be the deepest one: what does professional judgment mean when the generation step is automated but the selection step is the actual lawyering? The traditional apprenticeship taught junior lawyers what generation feels like — drafting, issue-spotting, synthesis. If generation becomes mostly automated, the apprenticeship has to focus on selection — which argument actually persuades, which issue actually matters, which synthesis is actually right. That's a different skill. We've never taught it as explicitly as we'll need to.
Use the NotebookLM on the page to take positions on any of these. Ask it to defend one side, then the other. The gap between the two responses tells you where the sources actually disagree — which is where the real argument lives.
Housekeeping
Three things before I let you go.
One. The course lives at polkwagner.com slash legal-education. It's an evolving resource — I speak on this topic often and I update the page as the conversation evolves. If you're hearing this after a talk, the link has everything we discussed, plus the open questions and the NotebookLM.
Two. The colophon, linked from the top of the course, documents the stack, the process, and a reproducible recipe. If you're a faculty member thinking about building something similar for your own topic, start there.
Three. Email me. Pwagner at law dot upenn dot edu. I want to hear what you're building, what's working, what isn't, experiments that flopped. The best ideas I've seen come from colleagues sharing what they've done. Drop me a note anytime.
Thanks for listening.