Legal Education @ Model Velocity

On Working with Wizards

Ethan Mollick's shift from "co-intelligence" to "wizards" — the observation that recent models are increasingly opaque, producing sophisticated outputs through processes users can't see into. The best short framing of what it's like to actually work with current AI. Read this first.

On AI, Universities, and Higher Education

Jesús Fernández-Villaverde, a Penn economist (SAS), wrestling publicly with what AI does to the case for traditional higher education. A sequence of threads from September 2025 to March 2026 — "Is AI the biggest change in education since the printing press?" (yes), followed by "twelve arguments for traditional higher education" and an evaluation framework. A Penn colleague working through our problem in real time.

All In: Embedding AI in the Law School Classroom

Gregory Duhl on embedding AI throughout a required doctrinal course — reimagining legal education by integrating AI as a learning enhancer rather than a threat to be managed. A concrete peer-institution account of what curriculum-level AI integration looks like.

What Most Law Schools Are Doing About AI

My read of where U.S. law schools sit on AI integration in early 2026 — drawn from Bloomberg Law's 2026 Path to Practice survey, recent peer-school announcements, and conversations with administrators at other schools. Four visible patterns (electives, embedded integration, certificates, policy updates) and two observations about why public signal isn't a reliable measure of institutional depth.

AI-Powered Lawyering

Minnesota/Michigan-led randomized controlled trial testing whether recent AI tools (RAG + reasoning models) materially improve legal work. Findings: statistically significant productivity gains of 50–130% across five of six legal tasks, with real quality improvements over earlier studies of GPT-4. The paper tested mid-to-late 2024 models; read the direction, not the specific numbers.

Grading Machines: Can AI Exam-Grading Replace Law Professors?

Six-author empirical study of AI performance grading real law school exams across four subjects at top-30 U.S. schools. Pearson correlations between AI-assigned and faculty-assigned grades up to 0.93 when AI is given a detailed rubric. The patterns of disagreement are more illuminating than the headline finding.

Turning Risks of Cheating with AI into Opportunities for Better Teaching

John Lande reframes the AI cheating problem as a teaching-design problem. The core move: if students can get AI to produce the answer you were going to grade, the problem isn't the AI — it's that the assessment was measuring something an AI can do. Redesign the assessment.

Can AI Hold Office Hours?

Ouellette et al. test AI models on answering 185 law-school questions about a patent-law casebook. Tests GPT-4o, Claude 3.5 Sonnet, and NotebookLM — the same tool stack used to build the very NotebookLM on this page. Read it and then chat with this course's NotebookLM to calibrate where reliability actually sits now.

Large Language Scholarship

Frazier and Rozenshtein argue that generative AI is already reshaping legal academia itself — law review submissions polished or partly generated by AI, articles written in weeks rather than months, detection practically impossible. Their prediction: disclosure rules won't hold, and scholarly norms will shift toward accepting AI-assisted work as routine. The outlier in this module — every other piece is about AI in legal *practice* or *pedagogy*; this one is about AI in legal *academia*.

Building the Future of the Legal Profession

A public summary of a proposed Penn Carey Law AI Initiative — four pillars (Research, People, Building, Pedagogy), the adaptive-architecture thesis, and the five-year direction. Described at a level appropriate for a general audience; the full concept document is in discussion with the Dean and development team.

Forging the Future: AI at Penn Carey Law

Penn Carey Law's public account of its AI work — AI integration in the 1L Legal Practice Skills curriculum, the AI Law Lab, and institutional access to ChatGPT EDU and Harvey AI for students. External news coverage of what the school is doing on AI.

Legal Tech Lab (working title)

A full-year 4-credit course launching in AY 2026-27. A small cohort of students building access-to-justice tools with community partners — focused on what scalable AI can actually do at the access-to-justice frontier. The goal is students who don't just think about the law but build with it.

AI 1L Orientation — Fall 2025

The AI session I co-run in the incoming 1L orientation — what Penn provides, permitted use as the school currently frames it, and the ethical questions I want students thinking about from day one. One piece of the institution-level work law schools are working out right now.

What Curriculum-Level AI Integration Actually Looks Like

A summary of Penn Legal Practice Skills' approach to integrating AI into the 1L curriculum. The key move is sequencing: AI-free baseline → structured instruction → three anchor experiential modules → reflection. Shared-platform rationale (transparency + equity + FERPA) follows from the design. The pattern generalizes beyond LPS.

AI Law Lab Bootcamps — Corporate & Litigation Tracks

The AI Law Lab's Spring 2026 bootcamp sessions. Two parallel tracks (Corporate, Litigation) — track separation matters because corporate and litigation workflows diverge enough that a generic course shortchanges both. Case-file method, practitioner integration, shared ethics thread through every session.

AI Law Lab Bootcamp — Q&A with the Instructors

Penn Carey Law's news feature on the Spring 2026 bootcamp, told from the instructors' perspective. Meghana Bhimaro and Lakshmi Prakash (both L'25 W'25) describe the course's design philosophy — technical competency plus the judgment to know where AI adds value and where it falls short. An external view that complements the internal syllabus summary.

Something Big Is Happening — AI and the Future of Legal Education

My most recent flagship talk on this subject — delivered February 19, 2026. Opens with Shumer's viral essay, walks through the agent-vs-chatbot shift, demonstrates Claude / Cowork / Claude Code live, and closes with the Slackbot and the exam-grading pipeline. The closest thing to a single-deck version of this whole course.

Teaching with Generative AI — Fall 2024

Earlier faculty retreat slides on practical AI use in law teaching — the classroom pilots, the ethical and legal questions, what to try. Useful as a companion to the Feb 2026 deck: the same subject, eighteen months earlier. The gap between the two decks is itself data about how fast this has moved.

Heron — the Intro to IP Teaching Assistant

A Slack bot I built for my Intro to IP course. Students know it as Heron — named for Heron of Alexandria, the ancient mathematician who bridged theoretical knowledge and practical invention. Answers student questions with citations back to the assigned course materials. Uses RAG with embeddings in a vector database. In this course's usage, students reach for Heron far more often than Canvas for Q&A.

Built entirely with Claude Code, by me. No engineering team, no vendor. The parallel to Shapiro's Claude-native law firm is direct: Shapiro built his practice around Claude; I built the teaching infrastructure for my course around Claude Code. Both are available to individuals now. The demonstration is narrow but concrete: course-specific AI teaching infrastructure no longer requires a vendor.

law-faculty-skills — The Claude-Native Law Professor

An open-source collection of Claude Code and ChatGPT skills for law faculty teaching tasks — MCQ generation, essay exam generation, class problem creation, slide review, full class prep, and more. Shapiro built a Claude-native law firm. This is the pedagogical analog.

AI Law Lab Resource Menu

Living directory of guides: AI attribution policies, sample syllabus language, prompt engineering for legal work, legal AI tool overviews, and how to build a Virtual TA with Custom GPTs. Maintained by the AI Law Lab; updates propagate automatically.

Creating a Virtual TA with Custom GPTs

Step-by-step guide to building a course Virtual TA using OpenAI's Custom GPTs on the ChatGPT EDU platform. Based on the one I built for Intro to IP before the Slack version.

Exams Destabilized — Data Slides

Excerpt from my own Fall 2025 talk to the faculty on rethinking assessment — the "destabilized" framing is mine, not a faculty conclusion. The core data slides only: current Penn exam format breakdown (in-class vs. take-home, exam lengths, format distribution). A baseline for what's actually being assessed right now.

1. If AI can generate sophisticated legal reasoning, what are we testing on exams?

The traditional law-school exam — an issue-spotter, a close-reading question, a policy essay — is a proxy for "can this student do the analytical work a junior lawyer is expected to do." If the analytical work is increasingly AI-assisted in practice — and, as Module II's empirical work suggests, increasingly within reach of current models at the level of complexity a typical exam tests — the proxy starts to drift. This doesn't mean we should ban AI on exams or permit it unconditionally. It means we have to decide what we're actually measuring — capability without AI, capability with AI, or something new. In my read, most law schools are making that choice by default rather than by design.

2. Does a "Claude-native" 1L curriculum still have 1L courses as we know them?

The 1L year has looked substantially the same for generations. Civil Procedure, Contracts, Torts, Property, Constitutional Law, Criminal Law, Legal Writing. If a first-year student using Claude can produce the memo-and-brief output that the LRW program was designed to produce, the question isn't whether we still teach LRW. It's what LRW teaches that an AI-assisted student genuinely still needs — and what part is now redundant. Same question for each doctrinal course. The answer is probably "the core analytical skills remain essential, and the vehicles for teaching them change." Saying that is easier than doing it.

3. What happens to rank-in-class if AI-assisted output is indistinguishable from unassisted?

Every top school has a grade distribution that employers rely on. If the variance in student output collapses because AI smooths out the bottom of the distribution, the grade becomes less informative about student quality. The grade still sorts, but what it sorts on shifts. Firms and clerkship committees will adapt; the question is whether we lead that adaptation or chase it. The schools that lead it will have more say in what the next generation of sorting looks like.

4. Who pays for the tools — the firm, the school, the student, or do we stop caring?

Right now, at most top schools — in my observation — institutional AI provision is uneven: general-purpose tools (ChatGPT, Claude) for faculty, less consistent for students at scale; legal-specific platforms adopted unevenly. Anecdotally, students who pay for Pro tiers tend to outperform students who don't on AI-assisted assignments. This is an equity question, a curriculum question, and a budget question at once — and every school is answering it differently. The answer probably converges toward "the school pays," because the alternative is either mandated opt-out or a structural disadvantage for students who can't afford a subscription. The timeline on that convergence is unclear, and the cost compounds.

5. How do law schools maintain pedagogical identity when every school adopts the same base-model technology?

A Penn Carey Law graduate has historically been a specific thing. A Harvard graduate, ditto. Some of that is brand, but some is training — faculty, cases, clinical opportunities, peer group, specific pedagogy. If every top school runs similar AI-augmented courses on the same underlying models, what differentiates the graduates? The answer is in the specific pedagogical choices, the research culture, and the institutional values — and institutions will have to be more deliberate about those than they historically needed to be.

6. What does "professional judgment" mean when the generation step is automated but the selection step is the actual lawyering?

This may be the deepest one. Every lawyer spends their career learning what the generation step feels like — drafting arguments, spotting issues, synthesizing cases. That apprenticeship is what junior associate work has traditionally been. If the generation step becomes mostly automated, the apprenticeship has to focus on the selection step — which argument is actually persuasive, which issue actually matters, which synthesis is actually correct. That's a different skill, taught through different exercises, assessed differently. We've never taught it as explicitly as we'll need to, because we've relied on years of post-graduate practice to develop it. That runway is shortening.