Law & Technology & Economics of AI

The course in four sentences

One claim per session

Each session opens and closes with its claim. If you remember four sentences in five years, make it these.

Session 1 · Architecture

“Before law regulates AI and before markets form, the architecture of AI already decides who can compete, who can scale… and who cannot.”

Session 1 · Economics

“In AI markets, economic success is not rewarded once, it is continuously amplified.”

Session 3 · Regulation

“Every AI regulation is more than a legal text; it is a hypothesis about how technology and markets will evolve together.”

Session 3 · Computational law

“The effectiveness of AI regulation may lie not in its ambitions, but in its capacity to be translated into executable systems.”

Schedule

Four stops, two weeks

Lectures on Thursdays, case studies on Saturdays. Fridays are for your team.

1Thu Oct 22Architecture & economics
Fri Oct 23Last day to drop · Team prep
2Sat Oct 24Case study: the open-weights claim
3Thu Oct 29Regulation & how rules travel
Fri Oct 30Team prep
4Sat Oct 31Case study: the Roosevelt Island AI Act
Registrar’s dateTake-home exam

All sessions run 1:30 to 5:00 pm Eastern Time. In-class exercises need no preparation. Each case study is prepared with your team the day before.

Sessions

Every session tests one claim

Each block names the claim under test, what we do in class, what to read, and five self-check questions. Tick a question when you can answer it from memory. Your ticks stay in this browser only. Readings marked “PDF on Canvas” sit behind a paywall; you will find them on Canvas. Exercises happen in the classroom and need no preparation. Case studies are prepared with your team.

1

Architecture & economics

Thursday, October 22 · 1:30 to 5:00 pm

Lecture 1 · The technical architecture of AI

The claim under test

“Training compute is a good proxy for what a model can do.”

We take the AI stack apart, from chips and cloud to foundation models and applications, and look at how scaling works and where it stops working. We then compare open and closed weights.

Exercisein class · no preparation · 25 min
Using Epoch AI’s public data, find the most capable model trained below the 10²⁶ FLOP line used by California and New York. Then ask what the rule would say about it.

In Epoch AI’s data, October 5, 202624 models above 10²⁵ FLOP, from 14 developers. 4 above 10²⁶.108 models released since June 2025 by OpenAI, Anthropic, Google and xAI have no public compute estimate.

Required

Neil C. Thompson, Kristjan Greenewald, Keeheon Lee & Gabriel F. Manso (2023), “The Computational Limits of Deep Learning”. Ninth Computing within Limits (LIMITS 2023)
Sara Hooker (2024), “On the Limitations of Compute Thresholds as a Governance Strategy”. arXiv:2407.05694

Optional

Sara Hooker (2024), “Big AI, The Compute Frenzy, and Grumpy Models”. Scaling Theory podcast, episode 8 Video · Spotify · Apple Podcasts
Ashish Vaswani et al. (2017), “Attention Is All You Need”. arXiv:1706.03762
Toby Ord (2025), “Inference Scaling and AI Governance”. Centre for the Governance of AI
Thibault Schrepel & Jason Potts (2025), “Measuring the Openness of AI Foundation Models: Competition and Policy Implications”. Information & Communications Technology Law 34(3): 279-304 PDF on Canvas

Self-check0/5

Why can adding compute produce diminishing returns? What capability would a training-compute threshold miss in a model that reasons longer when it answers? Which layer of the AI stack carries the highest fixed costs? What can a developer do with open weights that an API does not allow? Which layer of the stack would you avoid entering as a startup?

Lecture 2 · The economics of intelligence

The claim under test

“Scale turns AI into a natural monopoly.”

AI combines the capital intensity of heavy industry with the increasing returns of digital markets. We look at fixed costs, prices, data, talent, and the business models firms build around them.

Exercisein class · no preparation · 25 min
The strategy game. Teams play firms and a regulator with twelve strategies, from network effects to regulatory barriers. Each team decides which to use, which to ban, and which to lobby for.

A hypothesis under pressureDemirer, Fradkin and Tadelis (2026) report that LLM inference prices fell about 1,000-fold, that open models cost about 90% less than comparable closed ones, and that no single model dominates.

Required

Anton Korinek & Jai Vipra (2025), “Concentrating intelligence: scaling and market structure in artificial intelligence”. Economic Policy 40(121): 225-256 PDF on Canvas
Mert Demirer, Andrey Fradkin & Nadav Tadelis (2026), “The Emerging Market for Intelligence: How Firms Buy and Sell AI”. Journal of Economic Perspectives 40(3): 23-46

Optional

Pierre Azoulay, Joshua L. Krieger & Abhishek Nagaraj (2025), “Old Moats for New Models: Openness, Control, and Competition in Generative AI”. Entrepreneurship and Innovation Policy and the Economy, vol. 4 (NBER WP 32474) PDF on Canvas
Thibault Schrepel & Alex Pentland (2025), “Competition between AI foundation models: dynamics and policy recommendations”. Industrial and Corporate Change 34(5): 1085-1103 PDF on Canvas

Self-check0/5

Why does serving an AI model cost more per user than serving traditional software? Through which mechanism does scaling lead to concentration in Korinek and Vipra’s account? Which finding in Demirer, Fradkin and Tadelis contradicts the natural-monopoly prediction? Which input would be hardest for your startup to secure? Which strategy from the game would a regulator find hardest to detect?

2

Case study · Is giving weights away a durable strategy?

Saturday, October 24 · 1:30 to 5:00 pm

The claim under test

“Giving away model weights is a durable strategy.”

In July 2024, Meta committed to open weights. A week later, the National Telecommunications and Information Administration (NTIA) of the U.S. Department of Commerce made a regulatory claim. It recommended that the government monitor open-weight models without restricting them, and act only if evidence of risk emerged. This case study tests both claims.

Prepare with your team (Friday, October 23)
Each Canvas team has a role. Graded as part of the case-study component.
Every team brings a presentation of about 6 minutes, followed by about 6 minutes of questions from the instructor. The presentation answers two questions. What is your strategy? And should the government keep the NTIA approach, restrict open weights, or adopt a different rule? If you propose a rule, show its text.

MetaPresent the 2024 claim on open weights, what had to be true for it to work, and the indicator you would have watched. Then take the position on the NTIA approach that serves your strategy.
Closed-model labPresent your response to a rival that gives its weights away, and where your advantage must come from. Then say whether leaving open weights unrestricted tilts the market against you.
PolicymakersDecide whether open-weight models should face restrictions. Defend your rule, and name the evidence that would make you change it.
StartupYou build your product on open weights. Present the risks of depending on a model you do not control, and the rule you would ask the government for.

In class (times are approximate). Each team presents, then answers about six minutes of questions from the instructor meant to take you out of your comfort zone. Between presentations, the instructor gives short mini-lectures. Each one introduces a new concept and connects it to what the teams just presented. Expect to be asked whether the concept changes your position.

1:30Opening
1:35Meta teams present
2:00Closed-model lab teams present
2:25Mini-lecture · The economics of open sourceWhy would a firm give away what it paid to build? Lerner & Tirole (2002), read against the two strategies just presented.
2:35The startup team presents
2:50Mini-lecture · Hold-upWhat happens when your product depends on an asset someone else controls? Klein, Crawford & Alchian (1978), applied to the startups’ risks and to the terms of open-weight licences.
3:00Break
3:15The policymaker team presents
3:30The timeline gamein class · no preparation
Six real events are revealed one at a time, the last one from October 2026. After each one, every team decides its next move in its role. Then the class compares with what actually happened and names the assumption that broke. To close, each team says whether its position on the rule survived the game.
4:35Open discussion and questions, until 5:00

Required

Mark Zuckerberg (2024), “Open Source AI Is the Path Forward”. Meta Newsroom, July 23, 2024
Joel Spolsky (2002), “Strategy Letter V”. Joel on Software, June 12, 2002
NTIA (2024), Dual-Use Foundation Models with Widely Available Model Weights, recommendations. U.S. Department of Commerce, July 30, 2024

Optional

Thomas Wolf (2024), “Open-Source AI Scaling Laws”. Scaling Theory podcast, episode 3 Video · Spotify · Apple Podcasts
Yann LeCun (2024), “AI Dynamics and Regulation”. Scaling Theory podcast, episode 5 Video · Spotify · Apple Podcasts
Frank Nagle & Daniel Yue (2025), “The Latent Role of Open Models in the AI Economy”. SSRN working paper

Self-check0/5

In which business does Meta earn its profits? What does “commoditize your complement” mean? What does Zuckerberg’s 2024 case for open weights assume about competitors? What would make open weights lose their value for Meta? Which evidence would, under the NTIA report, justify restricting open weights?

3

Regulation & how rules travel

Thursday, October 29 · 1:30 to 5:00 pm

Lecture 3 · A theory of regulation for AI

The claim under test

“A rule serves the goal written in its preamble.”

Why regulate at all, who asks for regulation, and what counts as capture. We compare rules with standards, liability after harm with licensing before it, and rules that ask firms to write their own safety plans. We close with the Collingridge dilemma and what it implies for proposals to slow AI down.

Exercisein class · no preparation · 30 min
Who pays for a threshold? Each teammate holds one piece of the answer. CS students get the compute data, MBA students the revenue data, law students the legal texts. Nobody can answer alone.

Required

George J. Stigler (1971), “The Theory of Economic Regulation”. Bell Journal of Economics and Management Science 2(1): 3 PDF on Canvas
Neel Guha et al. (2024), “AI Regulation Has Its Own Alignment Problem: The Technical and Institutional Feasibility of Disclosure, Registration, Licensing, and Auditing”. George Washington Law Review 92(6): 1473

Optional

Daniel Carpenter (2013), “Detecting and Measuring Capture”, in Daniel Carpenter & David A. Moss, eds., Preventing Regulatory Capture: Special Interest Influence and How to Limit It. Cambridge University Press, chapter 3: 57-68 PDF on Canvas
Noam Kolt, Michal Shur-Ofry & Reuven Cohen (2025), “Lessons from Complex Systems Science for AI Governance”. Patterns 6(8): 101341

Self-check0/5

Which of Stigler’s four benefits would a licensing regime give incumbents? What must a claim of regulatory capture establish? For one tool studied by Guha et al., which technical condition must hold for it to work? What is the Collingridge dilemma? Why can a rule with high fixed compliance costs favor incumbents?

Lecture 4 · How far does an AI rule reach?

The claim under test

“A rule written in one large market ends up governing AI everywhere.”

We test the Brussels and California effects on AI, then study China’s layered approach. We work through paired cases: the same problem regulated in two places. We end with computational law and ask when compliance can actually be measured.

Three data pointsNew York amended its RAISE Act on March 27, 2026 to adopt California’s frontier thresholds.China’s rules on human-like AI companions took effect on July 15, 2026.In 2024, researchers found audit reports from 18 of 391 New York City employers under Local Law 144.

Required

Anu Bradford (2012), “The Brussels Effect”. Northwestern University Law Review 107(1): 1
Matt Sheehan (2023), “China’s AI Regulations and How They Get Made”. Carnegie Endowment for International Peace

Optional

Scott Singer & Matt Sheehan (2026), “China Is Worried About AI Companions. Here’s What It’s Doing About Them”. Carnegie Endowment for International Peace
Marco Almada & Anca Radu (2024), “The Brussels Side-Effect: How the AI Act Can Reduce the Global Reach of EU Policy”. German Law Journal 25(4): 646-663 PDF on Canvas
Lucas Wright et al. (2024), “Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability”. FAccT ’24: 1701-1713

Self-check0/5

Which of Bradford’s five conditions is weakest for AI rules? Why does product divisibility decide whether a Brussels effect occurs? In what order did China adopt its AI rules? How do Chinese and Californian rules on AI companions differ in enforcement? Why did so few employers publish audits under Local Law 144?

4

Case study · The Roosevelt Island AI Act

Saturday, October 31 · 1:30 to 5:00 pm

The claim under test

“A line drawn by law still holds once the technology moves.”

Roosevelt Island has declared itself a small independent state, and its legislature wants an AI law by the end of the afternoon. The island has a tiny market and a regulator with two staff. Every round of the hearing puts a tool from Lectures 3 and 4 to work, and the mini-lectures add new ones.

Prepare with your team (Friday, October 30)
Each Canvas team has a role, usually two teams per role. Graded as part of the case-study component.
Every team brings a presentation of about 6 minutes, followed by about 6 minutes of questions from the instructor. Legislators also bring their two-page draft.

Island legislatorsDraft a two-page law on frontier models and AI-generated content. For each provision, say whether it is a rule or a standard, and whether it acts before or after harm. One team must use a compute threshold. The other team may not.
DevelopersOne team plays a frontier lab, the other a startup that builds on open weights. Estimate what a compute threshold and a labelling duty would cost you, per model and as a fixed cost. Each team also receives a secret objective from the instructor and pursues it during the hearing without revealing it.
Island regulatorsDesign a compliance check you could run with two staff: what you would measure, with which data, and how you would spot a capable model trained below a threshold.
Other jurisdictionsThree teams, each speaking for one jurisdiction: California, China or the EU. Know what your jurisdiction already requires on thresholds and labels, and the conditions under which you would copy a law written by a small state.

In class (times are approximate). Each team presents, then answers about six minutes of questions from the instructor meant to take you out of your comfort zone. Between presentations, the instructor gives short mini-lectures. Each one introduces a new concept and connects it to what the teams just presented. Expect to be asked whether the concept changes your position.

1:30Opening
1:35The hearing opens: legislator teams present their drafts
1:59Mini-lecture · Goodhart’s lawA measure that becomes a target stops measuring well. Strathern (1997), applied to the compute thresholds in the drafts.
2:09Developer teams testify
2:33Mini-lecture · Raising rivals’ costsA firm can gain from a rule that costs it money, if the rule costs its rivals more. Salop & Scheffman (1983), applied to the developers’ testimony.
2:41Capture auditin class · no preparation
The class guesses each developer team’s secret objective, then applies Carpenter’s test for capture to the drafts. Who were the bootleggers in this hearing, and who were the Baptists?
2:48Regulator teams present their compliance checks and say whether they could enforce each draft with two staff.
3:12Mini-lecture · Responsive regulationAyres & Braithwaite (1992) and the enforcement pyramid, which starts with persuasion and escalates only when a firm does not comply. Applied to a regulator with two staff.
3:22Break
3:37California, China and EU teams present
4:13Mini-lecture · Network effects and standards, with a live label stress testA label is worth attaching only if platforms read it, and worth reading only if most content carries one. Katz & Shapiro (1985). Then an image with Content Credentials goes through a screenshot: does the Verify tool still find them, and how do SB 942, China’s labelling measures and Article 50 of the EU AI Act treat the stripped image?
4:25The vote and the 2031 cardin class · no preparation
A show of hands picks the draft most likely to still work in 2031. Then each student fills in a hypothesis card alone, without AI or devices, for a rule that does not exist yet.
4:40Open discussion and questions, until 5:00

The scoreboard. Drafts earn points when they pass the regulators’ check and survive the instructor’s questions. They lose a point for each capture flag the class confirms. The scoreboard is for fun and does not set grades.

Where the rules standCalifornia’s AI Transparency Act became operative on August 2, 2026, after AB 853 moved its start date.China’s labelling measures took effect on September 1, 2025.New York amended its RAISE Act on March 27, 2026 to adopt California’s frontier thresholds.

Required (short legal texts)

California SB 942 (AI Transparency Act, 2024), as amended by AB 853 (2025), and the definitions in SB 53 (Transparency in Frontier Artificial Intelligence Act, 2025).
Cyberspace Administration of China et al. (2025), Measures for Labeling of AI-Generated Synthetic Content. English translation, China Law Translate
Regulation (EU) 2024/1689, the AI Act, Articles 50 and 51.

Optional

Iyngkarran Kumar & Sam Manning (2025), “Trends in Frontier AI Model Count: A Forecast to 2028”. arXiv:2504.16138

Self-check0/5

What does SB 942 require from a covered provider? Why would New York adopt California’s thresholds instead of writing its own? Under a compute threshold, who carries the heavier fixed cost, a frontier lab or a startup? What evidence would a claim of capture need, according to Carpenter? Is labelling a divisible product, and what follows for a California effect?

Discussed in class

Sam Peltzman (1976), “Toward a More General Theory of Regulation”. Journal of Law and Economics 19(2): 211-240
Louis Kaplow (1992), “Rules Versus Standards: An Economic Analysis”. Duke Law Journal 42(3): 557-629
Steven Shavell (1984), “Liability for Harm versus Regulation of Safety”. Journal of Legal Studies 13(2): 357-374
Cary Coglianese & David Lazer (2003), “Management-Based Regulation: Prescribing Private Management to Achieve Public Goals”. Law & Society Review 37(4): 691-730
David Collingridge (1980), The Social Control of Technology. St. Martin’s Press / Frances Pinter
Charles I. Jones (2024), “The AI Dilemma: Growth versus Existential Risk”. American Economic Review: Insights 6(4): 575-590
Garrett A. Johnson, Scott K. Shriver & Samuel G. Goldberg (2023), “Privacy and Market Concentration: Intended and Unintended Consequences of the GDPR”. Management Science 69(10): 5695-5721
David Vogel (1995), Trading Up: Consumer and Environmental Regulation in a Global Economy. Harvard University Press
Varun Magesh et al. (2025), “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools”. Journal of Empirical Legal Studies 22(2): 216-242
Josh Lerner & Jean Tirole (2002), “Some Simple Economics of Open Source”. Journal of Industrial Economics 50(2): 197-234
Benjamin Klein, Robert G. Crawford & Armen A. Alchian (1978), “Vertical Integration, Appropriable Rents, and the Competitive Contracting Process”. Journal of Law and Economics 21(2): 297-326
Kenneth J. Arrow & Anthony C. Fisher (1974), “Environmental Preservation, Uncertainty, and Irreversibility”. Quarterly Journal of Economics 88(2): 312-319
Marilyn Strathern (1997), “‘Improving Ratings’: Audit in the British University System”. European Review 5(3): 305-321
Steven C. Salop & David T. Scheffman (1983), “Raising Rivals’ Costs”. American Economic Review 73(2): 267-271
Ian Ayres & John Braithwaite (1992), Responsive Regulation: Transcending the Deregulation Debate. Oxford University Press
Michael L. Katz & Carl Shapiro (1985), “Network Externalities, Competition, and Compatibility”. American Economic Review 75(3): 424-440

The tool you keep

The hypothesis card

One page, six lines. You fill it in every session, for rules and for business strategies, and once more for the exam. Drafts save in this browser only.

Worked example

California SB 53

A frontier model is one trained with more than 10²⁶ FLOP, and the heaviest duties fall on developers with more than $500 million in annual revenue. Enacted. Targets the developer.
Training compute tracks what a model can do.
Few firms can afford training runs that large, so the rule reaches a handful of developers.
If compute stops tracking capability, risky models below the line escape the rule, and models above it carry costs without matching risk.
Models below 10²⁶ FLOP that score at frontier level, and the share of new models with no public compute estimate.
The California Department of Technology reviews the definitions every year and recommends updates to the Legislature.

Why it matters for your career

One method, four jobs

If you become	The card teaches you to
A founder	Know at what revenue or compute level your company will cross each threshold, and turn that into a regulatory roadmap your investors can read.
A regulator	Choose the indicator that will show a rule failing, and write the clause that allows it to be corrected.
A lawyer	Draft a provision, then explain it to the engineer who has to implement it.
An engineer	Turn a legal obligation into a technical test, such as a watermark robustness test for California’s AI Transparency Act.

What employers ask for

“Distinguish enacted requirements, proposals, guidance, and emerging expectations.”

“Distinguish evidence from assumptions.”

From OpenAI’s job posting for a Senior Product Policy Lead, Regulation, which also asks candidates to reason “at the model, product, and deployment levels.” That is the hypothesis card.

By the end of the course, you will be able to

Explain how an AI model is built and deployed to a reader from another discipline.
Identify which features of an AI market give a firm a lasting advantage.
Determine which firms a regulatory threshold covers and who bears its compliance costs.
Draft a regulatory provision and justify the choice of instrument.
Predict whether a rule adopted in one jurisdiction will spread to others.

Grading & workload

Two grades, one final mark

Grades follow Cornell Law School’s grading policies. The goal is for as many of you as possible to understand the subject in depth and earn an excellent grade.

40%

Team case studies

Two presentations, on Saturday October 24 and Saturday October 31. You sign up on Canvas for the team of your choice, up to six students. Every member of a team receives the same grade.

60%

Individual take-home exam

An individual case study in which you combine your legal skills with the economic and technical insights from the course. The hypothesis card is the method to use. You download the questions and upload your answer through the Exam4 website, within the number of hours set for the exam. The registrar sets the date: check My Exam Information. Put your Exam ID in the header of every page and never your name or NetID. You may use your course materials and notes. Once the exam has begun and until grades are final, do not contact the instructor about it; write to the Registrar at law.registrar@cornell.edu. A mock exam is posted a few days before. You receive written feedback.

Using AI tools

Cornell’s Code of Academic Integrity treats any AI assistance not permitted in writing as unauthorized.

Take-home exam. The exam instructions state in writing whether AI tools are allowed. The exam is taken under the Law School Code of Academic Integrity and the Law School’s take-home instructions for Cornell Tech courses. Without written permission, AI assistance is not allowed.
Team presentations. AI tools are permitted. State which tools you used and for what, and check every source yourself. A fabricated citation is a violation of the Code.
The 2031 card on October 31. Written without AI and without devices.

Workload

About 25 hours outside class. Two required readings per session take two to three hours. Team preparation on Fridays October 23 and 30 takes two to three hours each time.

Publication opportunity

The best memos and prototypes produced in the course will be considered for publication in the Network Law Review, subject to peer review by its advisory board. Selected students will revise their work with editorial feedback after the course ends. Selection is independent of grading, and publication requires the consent of every author.

Policies

How we work together

Attendance

Attendance is required. If you must miss a session, tell the instructor and your team in advance. An unannounced absence costs 20% of the final grade per session missed. Waitlisted students must attend the first class.

Devices

Laptops and phones must be put away during lectures and case studies, unless the instructor asks you to use them. Students in the online section keep their devices on.

Academic integrity

All students follow Cornell’s Code of Academic Integrity, in force since August 24, 2026. Discussing concepts with classmates is encouraged. Plagiarism results in zero credit and may be referred to the Office of Student & Academic Affairs.

Privacy in class

Do not record, photograph or share interactions involving classmates or the teaching team, and do not share course materials outside the course without written permission. Students who need recordings for learning reasons should speak to the instructor.

Accessibility

Send your Student Disability Services accommodation letter early so arrangements can be made in time. For immediate needs, speak to the instructor after class or write to SDS at sds_cu@cornell.edu.

Wellbeing & observances

Your health matters as much as your grade, and Cornell offers support services for mental health and wellbeing. If a religious observance conflicts with a session, raise it well in advance so we can arrange the work.

Resources

Before, during and after

New to the technology? Watch the primer before October 22. The data sources are the ones we use in class.

Primer · Intro to Large Language ModelsAndrej Karpathy’s one-hour talk. Recommended for law and MBA students. Epoch AI · Data on AI ModelsTraining compute, dates and developers for thousands of models. Used in Sessions 1 and 3. Content Credentials · VerifyFree tool that reads provenance data in images. Used in Session 4. Stanford AI Index 2026Annual data on research, the economy and policy. OECD.AI Policy NavigatorAI policies and initiatives across countries. Our World in Data · AICharts and data on artificial intelligence. Scaling Theory podcastThe instructor’s podcast.