I have taught several different courses on law and AI at Emory Law, and I’m currently teaching “Current Topics in the Law and Policy of AI.” I’m not wild about any of the Law & AI textbooks currently available (but they are getting better), and I think most people in the field are still teaching their own bespoke courses, like I do. What is missing from those courses is a general introduction that sets the stage for all of the material that follows, and so I have tried to fill that gap with a simple online Primer.
Part I is the background. What is AI, how we got here, how large language models work, why prediction is not the same thing as truth, and why systems that had improved slowly for decades suddenly became astonishingly capable.
Part I also covers what has happened since ChatGPT: the long 2023, the long 2024, and the turn since 2025 toward deployment and state regulation.
Part II is the policy. Why anyone wants to regulate AI at all, taking the benefits, the harms, and the real uncertainty between them seriously. How AI changes problems the law already knew about: the declining cost of malicious activity, surveillance through accumulation, inference, biased proxies, opacity, manipulation, distribution, and concentration. And how to think about law and AI, including where existing law is a starting point rather than an answer, the difference between regulating uses, capabilities, and domains, and what to make of arguments about artificial general intelligence.
In total, the primer runs to about 8,000 words, divided across six reasonably modular chapters.
I primarily wrote this for my students. But if you are putting together a Law and AI syllabus, or you just want to follow the argument without a computer science degree, help yourself.
Generative AI has made every form of assessment other than a supervised exam (written or oral) essentially worthless as a measure of what a student knows. I will expand on that below, but it’s hard to see how it can seriously be a matter for debate any more. The obvious response is to bring exams back into the classroom. We should. Many of you have already. But we also need to think more carefully about whether those exams are an appropriate response to the current environment because generative AI is not the only development threatening the integrity of our assessments.
In-class exams don’t work the way they did in the 1990s when I was in law school. The standard long-fact-pattern exam, which asks students to spot and apply as many issues as possible under time pressure, has always rewarded speed in addition to understanding. Once upon a time, that might have been defensible, but it now interacts perversely with an accommodations system under which a third of a class may have extra time.
The good news is that some small changes can address both the challenges of generative AI and our current accommodations environment. Supervised, closed-book exams with a word limit that actually constrains would ensure that we are assessing students on their understanding of the law, not ChatGPT’s or Claudes. This format also makes sure that exams reward understanding more than typing speed and that extra-time accommodations do what they are meant to do instead of the opposite: create a more level playing field, regardless of disability, rather than creating an uneven playing field and perverse incentives.
We have it within our power to respond to generative AI in a way that also makes our exams fairer and more meaningful. We should all make the necessary changes now.
Take-home exams are dead
I don’t have any easy answers for what to do with legal writing, seminars, or experiential courses, but for ordinary doctrinal courses, take-home exams are dead. Some of you may resist this conclusion because you once fed a torts or contracts exam to ChatGPT and received a mediocre answer.
But there are two things we all need to understand as to why that is no longer any kind of answer. First, 2023 was a long time ago in AI terms. In January 2023, ChatGPT earned a C+ average on four real Minnesota law finals. Two months later, GPT-4 passed a complete Uniform Bar Exam, scoring above every state’s cut score. (OpenAI’s “90th percentile” claim was inflated; the pass was not.) Law professors at Maryland have tested several iterations of GPT on their own finals and found grades steadily improving to the point where OpenAI’s o3 earned three A+s, one A, one A-, two B+s, and a B across eight exams. Their 2026 follow-up found performance had plateaued at A-range.
Second, all those results come from very simple prompting. Our students are likely more resourceful. The enterprising ones are not typing “answer this question”; they are dissecting the casebook, the syllabus, and their notes, and feeding those to various models along with much more structured prompts. If you give me 24 hours, the course materials, and some time with Claude Code, I am certain I can ace any doctrinal law school exam while learning close to nothing. If I can, so can our students.
A natural experiment
If you want to see what happens when students use generative AI on their exams, consider Roberto Serrano, an economics professor at Brown, whose story Inside Higher Ed reported this month. This spring, after a December mass shooting at Brown left students anxious about classrooms, Serrano gave his Welfare Economics class a take-home midterm, the first in nearly two decades teaching the course. Enrollment nearly tripled, presumably because all of a sudden students developed a passion for Pareto optimality. The midterm average was 96 percent on a deliberately harder exam, against a historical range of 65 to 80. Forty of the 86 scored a perfect 100. Serrano ran the exam through ChatGPT, and the result mirrored several students’ answers. When he made the final in-person and proctored, eighteen students dropped the course, and a further nine stayed enrolled but skipped the exam. The final’s average was 48.6 percent, a historic low; nineteen students failed even after he lowered the passing line.
The chart shows the gap, student by student: for most of the class, the midterm sits near the top, and the final lands 30 to 50 points lower. Obviously, this is not one or two students having a bad day. Look carefully at the figure, and you will see three students who stand out, because their performance scores are moderately close together (S1, S22, and S31). S22 would have had the worst midterm score in the class if Serrano had not voided that assessment. The cost of AI-corrupted assessment falls hard on honest students like S22; whether that matters more than the lost learning these figures also suggest depends on how you balance these incommensurables.
Your take-home exam has close to zero validity as a measure of legal knowledge or skill.
False solutions
You might think that technology is coming to save you. AI detection tools can tell you something about a cohort, but as a means of identifying which individual student cheated, the tools are inherently unreliable. AI detection tools are nothing like plagiarism detection tools—they don’t compare new inputs to a database of prior works looking for a match; they cast auguries and scrutinize the shadow of a digital Punxsutawney Phil to diagnose AI drafting (not really, but they may as well). It is telling that OpenAI shut down its own detector after six months for “low rate of accuracy.” I would rather toss a coin than use an AI detector on a single piece of work (because in the former case I would not kid myself that I had learned anything), and I am not alone. A peer-reviewed evaluation of fourteen AI-text detectors found them “neither accurate nor reliable.”
The problem with the obvious solution
The obvious answer is to go back to a world of in-class exams using software to deactivate students’ access to the internet. “Electronic BlueBook,” “Softtest,” and other similar platforms have this functionality. We should all do this, but its not enough. We need to think harder about exam conditions.
Since time immemorial (or at least since The Paper Chase), law school exams have used time scarcity to produce variation in answer quality that we then map onto a grade distribution. To make sure they get a good distribution, law professors invariably pack their exams full of issues and sub-issues that are almost impossible to address in the time allotted. The theory was that students who understand more would do more with the time they had. That made sense in the blue-book era when everyone wrote at about the same speed and had the same time to complete their exams. Both predicates are false.
Typing speed varies, a lot
Most law schools moved from handwritten exams to exams typed on student laptops around twenty years ago or more. At the time, I don’t think anyone gave much thought to the fact that there is a lot more variation in typing speeds than in handwriting. Students average around 23 words per minute in handwriting, with the fastest writing twice as fast as the slowest. But according to a study at BYU a few years ago, typing speeds among law students ran from 21 to 108 words per minute: a 5x advantage rather than a 2x advantage.
It’s hard to imagine that anyone can defend the proposition that class rank should be determined by a typing test—except that revealed preference indicates that some of you do.
An arms race we built
Extra-time accommodations have become very, very common. In my Property class last semester, a third of the students had them, mostly time and a half, occasionally double time, and colleagues around the country report much the same.
Some of this reflects good news. It tracks a broader, more humane understanding of who struggles and why, and less stigma about saying so. I have no interest in relitigating any of that. I don’t want to demonize students seeking accommodations. I am dyslexic and have ADHD, and I remember how hard it was fighting to have my teachers and professors evaluate me on what I had written as opposed to my spelling. But I also know that if I had been given 50 percent more time than my fellow students in an open-book exam, that would have over-corrected.
A system that hands a large advantage to anyone with the right paperwork will not stay honest for long. Once students perceive that extra time confers a significant advantage in a highly competitive environment, the incentives are obvious. The legal standard for a disability is, by design, not a demanding one, so the rational move for a student under pressure is to find a name for whatever they are feeling and get in line. I have heard first-hand from students who feel this pressure and are wrestling with this dilemma. Many commentators see accommodations as a strategic tool for competitive advantage, with growth concentrated at the most selective and expensive institutions, but it’s important to understand that strategic is not the same as dishonest. No dishonesty is required, given the prevalence of ADHD, anxiety, etc.
Extra-time accommodation started out as a remedy for unfairness, but usually not a terribly well-thought-through one. Of course, I have no idea in any individual case whether a student should be accommodated with extra time, and if so, how much. I also suspect, based on my conversations with the university department that determines these accommodations, that they don’t either. When I pushed for an explanation from DAS last year, I was told that each evaluation was an “individualized assessment.” When I and another professor asked, “on what criteria?” we were offered the phrase “individualized assessment” as though it answered that question. It doesn’t.
We cycled through the “what criteria?” – “individualized assessment” loop several times until it emerged that the primary determinants of how much extra time to award were a combination of (a) what the student had asked for, (b) what they had received at a prior institution, and (c) what the relevant medical professional/therapist had recommended.
Our disability offices do their best, but their focus is compliance with the law as it has been explained to them. As far as I know, they are neither teachers, psychologists, nor psychiatrists, and thus they are poorly positioned to second-guess accommodations recommended by mental health professionals, already agreed to by some undergraduate institution. They also seem to have trouble understanding how something as benign-sounding as time and a half interacts with a curved, time-pressured exam.
How extra time distorts the results
I don’t see the accommodations landscape changing in a hurry, but we could do a far better job designing assessment so that extra-time accommodations are more equalizing than distortionary.
The unfairness of extra time is at its worst in open-book, points-accumulation exams (i.e., exams where you get points for saying something correct that matches the rubric, with no deduction for irrelevant or even incorrect observations). Not long ago I sat down with a student unhappy with his grade in one of my courses. He said he struggled with my 1,500-word limit on the three-hour exam, which contained one essay and a battery of multiple-choice questions. I asked, offhandedly, how much he would have written without a word limit. Six thousand words, he said.
Six
Thousand
Words
This student had an accommodation for double time in exams, and his strategy was: “you just dump your outline to make sure you hit every issue.” I don’t doubt that the condition behind his accommodation was genuine. But we are failing students like this, failing students who don’t have accommodations and failing the legal profession. Exams that are a race to accumulate correct statements, with no penalty for wrong ones and twice as much time for some as for others can’t be a valid measure of achievement at law school or legal ability. In my experience, clients want clarity and synthesis, the exact opposite of the “dump your outline” mentality.
My sympathies are entirely with people who struggle to read and write in conventional ways. But accommodations should be calibrated and focused on removing artificial impediments to performance that don’t reflect real-world conditions. We shouldn’t just be haphazardly giving some students who are stressed, depressed, or have learning difficulties an advantage over others (many of whom are probably also stressed, depressed, or struggling with learning difficulties).
When our unaccommodated students take their exams, they look around the room at the empty seats and do the math. Of course, every cohort contains students whose accommodations are unambiguously warranted. But poorly designed exams manufacture both unfairness and, just as corrosive, the perception of unfairness, among students who can count as well as I can.
I don’t have a perfect solution, but we can do a lot better.
The fix: supervised, closed-book exams with constraining word limits
All doctrinal courses should be assessed by supervised, closed-book exams with constraining word limits.
That regimen will not fix admissions or the accommodations pipeline. But it will dilute the incentives for an accommodations arms race and make our grades a fairer reflection of relative ability. When extra time no longer buys a materially better grade, far fewer students have any reason to chase a diagnosis, and those who genuinely need the accommodation get it without the side-eye.
Word limits
Word limits that actually constrain are vital.
You should set your word limit so the exam is not a typing test. This means setting it well below what students can physically produce (1,500 words, say, where the average student could type 2,500–3,000 in three hours). If you think it takes 3,000 words to write an A answer in your exam, you should give all your students six hours. If that sounds as awful to you as it does to me, write a simpler exam and give them 1,500 words. The worst way to set a word limit is to imagine what a fast typist would write in the standard time allowed in the course of a perfect answer. That will result in a limit that is basically no limit at all.
Word limits force students to prioritize among issues, which is a more important skill than typing fast. It’s a skill lawyers practice most days. With proper word limits, the equalizing effect that extra time is meant to have is preserved, but the unfair advantage is substantially reduced. Not eliminated, because nothing is perfect, but substantially reduced. Proper word limits would also mean that students who type slower than average have very little need or incentive to seek accommodations in the first place.
Closed-book
Closed-book exams reward internalized knowledge. You may feel they unfairly privilege memorization over some idealized form of “true, knowledge-free understanding,” but recall and understanding actually tend to go hand in hand.
More importantly, open-book exams pour gasoline on the unfairness of extra-time allocations. How can you make a fair comparison between one student with two hours and barely enough time to look at their outline, and another student with four hours and the luxury to consult an outline they may not have even written? If you really think it is unfair to test students on their ability to remember the elements of adverse possession and the like, you can include that information in the exam.
Supervised
All our exams should be properly supervised to ensure that what is meant to be a closed-book exam is really a closed-book exam. Accommodations that allow students to take their exams in private rooms are incredibly resource-demanding and undermine the integrity of the exam process if that leads to students taking exams unsupervised. Students have reported to me that other students have bragged about using AI in unsupervised rooms. Whether this hearsay is true or not, if students believe it is true, we have a serious problem.
What to do about it
This letter is written for my fellow law professors, and the message is simple. Generative AI and the shifting accommodation landscape both demand a response. There are things that should be done at the institutional level, and I hope we do them. But the most immediate remedy is already in front of us, and it is entirely within our power.
You do not need a committee or a faculty vote. You can change your syllabus now, for the courses you will teach next semester.
I have made this an open letter because every law school around the country seems to be facing this exact same problem. We, as a faculty, have been having versions of this conversation for quite some time, but it is also something very much on the minds of our students.
Sincerely,
Matthew Sag
Jonas Robitscher Professor of Law in Artificial Intelligence, Machine Learning, and Data Science, Emory University
Starting in academic year 2026–27, Emory Law will offer a formal concentration in Artificial Intelligence and the Law — a structured academic pathway for J.D. students who want to develop real expertise in one of the most consequential areas of modern legal practice.
The concentration requires students to complete at least 12 credits drawn from three core areas: foundational courses at the intersection of law and AI, privacy and technology law, and intellectual property. Students can also supplement those courses with approved electives, internships, and externships. There’s no competitive application process — students simply need to satisfy the requirements and indicate their interest in their final semester. Those who do will have “Artificial Intelligence and the Law Concentration” listed directly on their transcript.
Which courses are currently available?
For more authoritative information on concentration and how it works, go to the Emory website. I wanted to highlight the relevant courses offered in the 2026-2027 academic year.
The following is subject to change, but to the best of my knowledge Emory Law students will be able to take the following courses relevant to the concentration:
AI Fundamentals
Current Issues in Law and AI (Fall, Prof. Matthew Sag)
AI and Legal Writing (offered in Fall and Spring)
Law and Film – AI and Law (Spring, Prof. Ifeoma Ajunwa)
AI and Intellectual Property
Copyright Law (Fall, Prof. Matthew Sag)
Intellectual Property (Spring, Professor Margo Bagley)
AI Privacy and Health Law
Privacy Law and AI (Fall, Prof. Ifeoma Ajunwa)
Genetics & the Law (Spring, Prof. Jessica Roberts)
Privacy Law (Spring, Adjunct Prof. Will Bracker)
Other
Fundamentals of Innovation I (Fall, Prof. Nicole Morris)
Fundamentals of Innovation II (Spring, Prof. Nicole Morris)
What should you take?
My advice is to think about what kind of lawyer you want to be and what kinds of clients that lawyer will work with, work backward from there to figure out what useful competencies and knowledge to build. For example, if you see yourself becoming in-house counsel at a technology company, you will also want a strong background in corporate law and antitrust, in addition to AI and IP courses. It’s also very helpful to know a little something about labor law and secure transactions.
I can’t really tell you what courses you should take, but I can tell you what I would take given my interests and the slightly unrealistic assumption that I was trying to meet all the concentration requirements in one academic year. I would take “Current Issues in Law and AI” (2 credits) and Copyright Law (3 credits) with me in the Fall; Genetics & the Law (3 credits) with Professor Roberts, and Privacy Law (3 credits) with Will Bracker in the Spring, plus AI and Legal Writing (2 credits).
Important note
This blog post is not the final authoritative word on the requirements of the concentration or the courses that will be available in the next academic year. The information presented here is meant to be helpful but not authoritative. It is definitely subject to change.
The 5th Annual Legal Scholars Roundtable on Artificial Intelligence at Emory University School of Law starts at Emory Law tomorrow. The Roundtable features a phenomenal lineup of authors, commentators and participants, including: Katrina Geddes, Andres Sawicki, Gabriel Weil, Jonathan Iwry, Nathan Reitinger, Bryan Choi, Jacob Noti-Victor, Annemarie Bridy, Charlotte Tschider, Michael Froomkin, David Rubenstein, Yonathan Arbel, Yinn-Ching Lu, Nikola Datzov, Christina Lee, Deven Desai, Chinmayi Sharma, Jessica Roberts, Jillian Grennan, Lawrence Nodine, Salwa Hoque, and Andrew Miller.
Papers will address deepfakes, strict liability for frontier AI, robots.txt and web scraping regulation, copyright litigation after generative AI, privacy law’s first principles in the AI age, AI culture wars and federalism, LLM courts, and AI agents’ shadow principals.
As always, this conference is made possible by Emory Law and Emory University’s AI.Humanity.
The roundtable is an invitation only event. But if you missed out this year we encourage you to apply next year.
David Kemp has just released a policy builder designed to help people who are struggling to design an AI policy that is relevant to their specific course.
According to the website, the policy content is “grounded in Sag, AI Policies for Law Schools (2025); Bliss, Teaching Law in the Age of Generative AI (2024); Perkins, AI Now: The Duty to Integrate AI Education in Law Schools (2024); Moppett, Preparing Students for the AI Era (2025); ABA Formal Opinion 512; and state bar guidance. All content is dedicated to the public domain under CC0 1.0 Universal.”
I have not tested this, but it seems like a fabulous idea. If you want to know more about my thoughts on building AI policies for law school classes, I have a paper on this topic on SSRN, the upshot of which is that effective AI policies must be course-specific, enforceable, and focused on teaching students to use AI responsibly as future legal professionals.
This post is a very lightly edited extract from my forthcoming article in the Duke Law Journal, Copyright’s Jagged Frontier (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6319379)
What does AI memorization prove?
Some argue that any evidence of memorization necessarily negates the claim that AI models are transformative. They advance this claim by injecting the term “compression” into the conversation in a way that suggests that AI models like GPT, Claude, and Gemini, are compressed representations of their training data in the same way that an MP3 music file is a compressed version of music from a compact disc.
“[model training is] similar to what’s called lossy compression, which one way to describe it is if you have a giant file and you compress it into a ZIP file, you lose some of the contents of the work, but effectively you’re just actually compressing the file. … it’s actually taking the expressive content of the training data and compressing it down into a model. And that confirms that there’s no actual transformative use going on here … what the model is doing is actually just repeating over and over the training data over and over again.”
— Bartz v. Anthropic, Transcript of Motion for Summary Judgment Oral Argument, May 22, 2025., p44-45 (explaining Plaintiff’s expert’s view)
Alex Reisner (AI’s Memorization Crisis, The Atlantic), for example, draws on the Cooper and Ahmed studies, and argues that the evidence of memorization undermines the learning metaphor and reveals generative AI training for what it really is: “compression.” The upshot is, “Large language models don’t ‘learn’—they copy[.]” See also Ted Chaing‘s famous essay: ChatGPT Is a Blurry JPEG of the Web.
Technically accurate but thoroughly misleading
Associating AI training with compression is technically accurate if you understand the term the way computer scientists do; but it is also thoroughly misleading if you associate compression with MP3s, JPEGs, and Zip files, as most of us do.
AI models learn compact internal representations of their training data which capture whatever patterns that enable more accurate predictions. It is equally valid to label this process as “abstraction”, “learning”, “dimension reduction”, or “compression”; but the compression label invites analogy to familiar media formats such as MP3s and JPEGs.
These formats store approximations of original works that can later be reconstructed in forms that closely resemble their sources and are usually regarded as functionally indistinguishable. Other than hipsters with a taste for vinyl records, consumers interact with ZIP files, JPEGs, and MP3s as functionally equivalent to their uncompressed originals; whatever information is discarded is socially normalized as imperceptible. Side note, I highly recommend Jonathan Sterne, MP3: The Meaning of a Format (2012).
Calling it compression tells you nothing
Training an AI model is nothing like a ripping music into an MP3 format. Calling that process “compression,” tells you nothing about the level of detail of what is learned or the significance of the information discarded. The compression metaphor is further misleading because it implies uniformity and predictability. In conventional audio or image compression, the same categories of information are discarded from every file according to stable and transparent criteria that reflect advance judgments about what matters and what does not. By contrast, memorization in large language models is uneven, incidental, and difficult to anticipate. We know that memorization is more likely when a model is exposed to multiple copies of the same work, and that the timing of exposure during training can matter. Beyond such generalities, however, it is not possible to predict in advance which works will be retained verbatim or to what degree.
The rhetoric of compression is really just an effort to sidestep a difficult empirical question, rather than to answer it. The fact that one thing is memorized to a degree that seems relevant under copyright law doesn’t prove that everything is memorized to a similar degree.
To evaluate whether memorization actually has significance under copyright law requires some kind of qualitative and quantitative assessment of the nature and extent of memorization. But even that statement is overbroad, as I explain in Copyright’s Jagged Frontier, what actually matters in terms of a fair use analysis is not memorization in the abstract, but memorization that finds its way into production.
Why the Line Between Legal and Infringing AI Won’t Be a Line at All
By Matthew James Sag
Everyone wants to know whether training AI on copyrighted works is legal. The real answer is: it depends—and the boundary between what’s permissible and what isn’t will be far messier than anyone expects.
In my forthcoming article in the Duke Law Journal, I argue that the copyright boundary for generative AI will be jagged rather than smooth. Not a clean bright line, but an irregular, context-dependent frontier shaped by the interaction of varying memorization rates across different AI models, divergent legal standards of similarity across different creative media, and the interplay of three distinct bodies of copyright doctrine (substantial similarity, fair use and secondary liability).
Understanding that jaggedness turns out to be essential—not just for predicting litigation outcomes, but for seeing the opportunities that lie on the other side.
The phrase “jagged frontier” will be familiar to many. It comes from the influential 2023 study by Fabrizio Dell’Acqua, Ethan Mollick, and colleagues, who used it to describe the uneven capability landscape of AI itself. It’s a useful concept because it captures the way that AI can be astonishingly good at some tasks while failing at others that seem equally difficult.
I borrow the metaphor deliberately, because copyright law presents generative AI with an analogous problem. The legal boundary between permissible and infringing AI conduct is similarly jagged: not because AI’s capabilities are uneven (though they are), but because the legal standards that determine infringement are themselves uneven across different creative domains. It seems likely that an AI system can cross the line into copyright infringement far more easily when generating music or images of recognizable characters than when generating prose—even when the underlying technology is essentially the same.
Explaining how and why the intersection of copyright and AI leads to a jagged frontier accounts for the first third of the article.
That jagged frontier is only the beginning of the story. Drawing on Ronald Coase’s insight that legal rules are starting points for adaptation and negotiation rather than final allocations, the Article argues that the extensive literature on AI and copyright has focused almost exclusively on fair use while ignoring what comes next. I might have something to say about that in a future post.
Matthew James Sag is the Jonas Robitscher Professor of Law in Artificial Intelligence, Machine Learning, and Data Science at Emory UniversitySchool of Law. His article “Copyright’s Jagged Frontier” is forthcoming in the Duke Law Journal.
Emory Law is proud to host the fifth annual Legal Scholars Roundtable on Artificial Intelligence. The Roundtable will take place on April 09-10, 2026, at Emory University in Atlanta, Georgia. The Legal Scholars Roundtable on Artificial Intelligence (AI) is designed to be a forum for the discussion of current legal scholarship on AI, covering a range of methodologies, topics, perspectives, and legal intersections.
Format Participation at the Roundtable will be limited and invitation-only. Participants are expected to read all the papers in advance and be prepared to offer substantive comments. We will try to accommodate a limited number of Zoom-based participants in exceptional circumstances, but in person attendance is strongly preferred.
Applications to present, comment, or participate We invite applications to participate, to comment, and/or to present from academics working on any topic relating to legal issues in AI. To request to present, you need to submit a substantially complete draft paper.
The deadline for submission is February 15, 2026, and decisions on participation will be made shortly thereafter, ideally, by March 1, 2026. If selected, final manuscripts are due April 1, 2026, to permit all participants an opportunity to read the papers prior to the conference.
To apply to participate, comment, or present, please fill out the google form:( https://forms.gle/h5Vqgj6xpDNSjfFDA).
What to expect from the Legal Scholars Roundtable on Artificial Intelligence The Legal Scholars Roundtable on Artificial Intelligence is a forum for the discussion of current legal scholarship on AI, spanning a range of methodologies, topics, perspectives, and legal intersections. Authors who present at the Roundtable will be selected from a competitive application process, and commentators are assigned based on their expertise. Participants will have an opportunity to provide direct feedback in paper sessions and will have access to draft papers but will be asked not to post papers publicly or share without author permission. Robust sessions involve energetic feedback from other paper authors, commentators, and participants. Our goal is to ensure all authors have the full participation of all workshop participants in each author’s session.
Space is limited and we expect people to stay for the entire conference.
Essential logistics The Roundtable will be held in person on the Emory campus in Atlanta, Georgia. The conference will begin on Thursday morning and run until 1PM on Friday. You can expect to be at the Atlanta airport by 1:45 PM, in time for a 2:30 PM flight or later on Friday. We will pay for your reasonable (economy) travel and accommodation expenses within the U.S. At the roundtable you will be well fed and caffeinated.
Organizers Matthew Sag, Jonas Robitscher Professor of Law in Artificial Intelligence, Machine Learning, and Data Science at Emory University Law School (msag@emory.edu) Charlotte Tschider, Professor of Law at Loyola Law Chicago (ctschider@luc.edu)
Today, December 11, 2025, OpenAI and Disney announced a partnership that essentially signals a marriage between generative AI and legacy media. Although some kind of deal was inevitable, the range and scope of this one are striking. Disney is sinking $1 billion into OpenAI for an equity stake and warrants, while simultaneously inking a three-year licensing deal.
The immediate result? OpenAI’s Sora and ChatGPT will legally ingest over 200 marquee characters from the Disney, Marvel, Pixar, and Star Wars vaults. We’ll see AI-generated Disney content on Disney+, and Disney employees will get enterprise-grade access to OpenAI’s tools. Notably, actor likenesses are off the table—a nod to the sensitivities of the recent labor strikes—but the direction of travel is clear. For more reporting, see the Verge.
AI companies and copyright industries are beginning to understand, and become reconciled to, the fact that neither side is going to score an absolute victory when it comes to the fair use issue for AI training. AI training that results in a model that learns from, but does not reproduce, the training data looks very likely to be upheld as fair use. Two recent cases held as much on summary judgement and this aligns with a line of precedent “nonexpressive use” cases that predate generative AI.
However, it’s becoming increasingly clear that it’s hard to train generative AI models to be really useful without some degree of memorization of the training data along the way. This is particularly problematic when it comes to copyrightable characters, because copyright protects characters more abstractly than most things. This is the well-known Snoopy problem (a term I coined in 2023).
Faced with this increasingly clear reality, it makes sense for consumer facing AI companies and entertainment Giants like Disney to think about licensing arrangements.
This deal signals a retreat from the fair use absolutism of early AI development. OpenAI and Disney have effectively priced the risk of memorization. Instead of spending the next decade in discovery arguing over pixel similarities, they are moving to a licensing regime. Disney gets paid and retains control; OpenAI gets legal certainty and the ability to serve the entertainment industry without looking over its shoulder.
Capital Crunch?
With competitors like Anthropic eyeing public listings, OpenAI’s decision to take strategic capital from a corporate giant like Disney may be telling. It suggests we are hitting a saturation point for traditional venture capital at the scale these foundation models require. It also hints that OpenAI sees more value in “smart money,” than in the volatility of the public markets. Disney isn’t just a piggy bank; it’s a hedge. By entangling itself with the world’s premier IP holder, OpenAI makes itself indispensable to the very industry that threatened to sue it out of existence. Or, I’m sure that’s the theory, whether it pans out that way remains to be seen.
The End of the Scaling Era?
Finally, this move also adds to the “Data Scarcity” thesis. The era of simply scraping the open web to make models smarter (2017–2025) might be over. The low-hanging fruit of the public internet has been picked, processed, recycled into synthetic data, and processed again, every which way you can imagine. To get better, and to stay ahead of open source rivals, companies like OpenAI are going to need access to data that no one else has. Google has YouTube; OpenAI now has the Magic Kingdom.
The Bottom Line
This is the template for the future. We are moving away from total war between AI and Content, toward a negotiated partition of the world. The tech companies provide the engine; the media giants provide the fuel. And for now, at least, both sides seem to think that’s a better outcome than leaving it up to a judge.
I wrote this blog post the morning the deal was announced, because it fits surprisingly well with a Law Review article I am writing, “The Snoopy Solution: How Fair Use and Licensing for Generative AI Can Coexist” based on a talk I gave at Yale last month.
Why content licensing cannot solve AI’s training-data problem
I have just published an article as part of the ProMarket symposium for the University of Chicago, the Booth School of Business, “The False Hope of Content Licensing at Internet Scale”
Although the article is not very long, I thought I would summarize my point even more briefly here.
AI developers have been on a shopping spree. Since mid-2023, OpenAI, Google, Anthropic and Meta have collectively spent hundreds of millions of dollars striking deals with publishers. OpenAI alone has inked agreements with everyone from the Associated Press to Condé Nast, gaining access to archives from The New Yorker, Vogue, The Wall Street Journal and dozens of other publications.
To many watching from the sidelines, these deals offer tantalizing proof that AI companies can—and should—pay for the content they consume.
However, the agreements grabbing headlines represent a tiny fraction of the data needed to train cutting-edge language models. Modern AI systems require trillions of diverse tokens scraped from across the internet—a scale and diversity that traditional licensing simply cannot reach.
