As law schools prepare for another academic year shaped by artificial intelligence, the University of Chicago Law School recently released one of the most thoughtful institutional statements I have seen on how legal education should respond.

Chicago’s statement, Rethinking Legal Education in the AI Era, begins from a premise that should by now be hard to dispute. Law schools need to ensure that students learn to think critically and independently without relying on AI, while also preparing them to use AI in a profession where these tools are becoming part of ordinary practice. To get there, Chicago organizes its strategy around three goals: developing AI-resilient pedagogy and assessment, elevating the “essential human” skills that distinguish excellent lawyers, and teaching students to use AI responsibly, effectively, and ethically.

Chicago has a history of being unusually good at this kind of institutional statement. A decade ago, a university committee chaired by Chicago law professor Geoff Stone produced the Report of the Committee on Freedom of Expression, now widely known as the “Chicago Statement” or the “Chicago Principles.” Rather than trying to anticipate every speech controversy that might arise on a university campus, the statement articulated a strong commitment to free and open inquiry while recognizing narrow exceptions necessary to preserve the functioning of the institution. The result became influential far beyond the Chicago campus and the statement has since been adopted by more than 100 other colleges and universities across the country.

The new Chicago Law School statement on AI strikes me as part of the same tradition. It begins with educational principles rather than a taxonomy of acceptable and unacceptable uses of technology. I like that approach because it frames the problem correctly. As I have argued in prior posts, the challenge presented by AI should not be reduced to whether or when students should be permitted to use a particular technology. Rather, the goal should be to preserve the intellectual work that enables students to become good lawyers while also teaching them to succeed in a profession where AI will increasingly perform work lawyers once did themselves. Chicago puts it like this: students must learn to think with, without, and about AI.

First, students must learn to think and work with AI. As soon as they graduate, they will encounter AI tools in law firms, government offices, courts, corporations, public-interest organizations, and just about every other professional environment they might enter. They will need to know what tasks can appropriately be delegated, what outputs require verification, and when efficiency and increasingly agentic capabilities create new opportunities or risks.

Just as importantly, students must learn to think and work without AI. They still need to be taught the foundational skills that have rightly been at the heart of legal education for decades. A student who cannot read a case, identify an issue, construct an argument, or write clearly and persuasively without AI assistance will not be well equipped to evaluate AI-generated work or supervise the systems that produce it.

Finally, students must learn about AI. They need enough technical and institutional understanding to appreciate how these systems work, when or why they might fail, how labs and vendors shape their operation, and what ethical and governance questions arise when lawyers depend on them.

Chicago’s strategy is strongest when it translates these objectives into curricular design. Its proposed approach to first-year legal research and writing is especially promising. Students will continue to write without AI while also learning to use it for things like research, revision, and oral-argument preparation. Instructors will review both the students’ writing and their use of AI. That is exactly the kind of structured exposure law schools should provide. Students should not be left to teach themselves how to use AI tools, nor should faculty treat every interaction with AI as evidence of laziness or misconduct. The goal should be supervised practice: learning when AI helps, when it does not, and how its work must be checked.

Chicago’s decision to add an oral-discussion requirement for substantial research papers is similarly sound. Students will be required to discuss their papers in person with their supervising professors after submitting a complete draft or final version, answering questions about their reasoning and the implications of their arguments. That strikes me as a much better response to generative AI than attempting to detect whether a paper was machine-written or merely “sounds” machine-written. A live discussion tests whether students understand the arguments they are making, recognize their weaknesses, and can respond to questions in real time. Those are worthwhile educational objectives regardless of AI’s impact or role.

The statement is also smart to emphasize clinical education. Clinics can be one of the best places for students to learn AI-augmented lawyering because the work involves real clients, actual professional obligations, and close faculty supervision. Chicago has already been obtaining access to AI tools for its clinics, including tools designed for different forms of legal practice, while each clinic develops policies tailored to its practice area and the risks it presents.

There is one part of Chicago’s approach that I would probably handle differently. For the 2026–27 academic year, the law school will pilot a coordinated no-device policy across the required 1L curriculum. With limited exceptions, students will not be permitted to use laptops, tablets, or phones in class.

I understand the motivation. Professors want students to be present, and they do not want them quietly asking ChatGPT or Claude to answer questions or summarize cases. They want students to struggle with legal problems themselves instead of outsourcing the difficult work of thinking.

That’s an important goal, but I do not think a general laptop ban is necessary to accomplish it. A laptop used for notetaking is not the same thing as AI being used to generate an answer. Many students take better notes electronically. That was certainly true for me. I could organize material, build outlines, search for specific rules and hypotheticals, and connect current discussions with earlier classes much more easily than I could using handwritten notes alone (and I have terrible, doctor-like handwriting). Professors also have less blunt alternatives available to them. They can require students to close their computers during a particular discussion, exercise, or simulation, and they can design questions that require students to explain their reasoning rather than merely announce an answer. There are practical concerns, too. Students who need laptops as an accommodation may become more visible precisely because everyone else is prohibited from using them, and a categorical ban may create pressure for additional accommodation requests from students who genuinely struggle to take usable handwritten notes.

Still, the laptop policy is not a hill I would die on. Chicago is treating it as a pilot and has expressly committed to revisiting its policies as technology and legal practice evolve. The key is that the law school’s larger vision gets the central issue right: start with what students need to learn and design the institution’s response to AI around those objectives.

That leaves a related question. Even if a law school adopts Chicago’s broader educational strategy, it still needs to tell students something about how they may use AI. To that end, one of the best approaches I’ve seen comes from a much smaller and very different institution here in Iowa City.

The Iowa Conservatory (ICON), a performing arts middle and high school, has adopted a one-page AI policy that elegantly captures the basic principles I believe an educational institution needs to establish. It recognizes AI as a legitimate tool for learning and creative work. It allows instructors to determine when its use is appropriate based on the goals of a course or assignment. It requires students to disclose AI assistance, verify AI-assisted content, protect confidential information, and refrain from submitting AI-generated work as entirely their own. Most importantly, because ICON is an arts-focused institution, its policy emphasizes that students must remain the “primary creators, decision-makers, and artistic authors” of their work. That last principle translates remarkably well to legal education. Students using AI should remain the primary thinkers, decision-makers, and authors of the academic legal work they produce.

Of course, law schools will sometimes need more specific rules. Exams, clinics, research assignments, and other settings present distinctive concerns. A clinic may need detailed restrictions involving confidential client information and approved platforms, for instance. But contextual needs do not require an institution to anticipate every possible AI product, feature, or use case in a single comprehensive code.

So, if you put Chicago and ICON together, you get something close to the model I would recommend for law schools. At the institutional level, adopt Chicago’s larger educational vision. Design legal education so that students learn to think with, without, and about AI. Build assessments that reward actual understanding. Preserve opportunities for meaningful thinking without technological assistance. Give students supervised experience using AI. Emphasize the human skills—judgment, communication, advocacy, relationship building, strategic thinking—that remain central to good lawyering.

From there, law schools should govern student AI use with something closer to ICON’s short statement of principles. Tell students that AI may assist their work; that faculty determine when particular uses serve the objectives of a course or assignment; that meaningful AI assistance should be transparent; and that students remain responsible for the reasoning, judgment, accuracy, and authorship of what they submit. More specific rules can then be developed by professors and programs where the context requires them.

There is a lesson here that extends beyond AI. Chicago’s famous statement on free expression continues to succeed because it articulated a durable institutional principle clearly enough to guide decisions that its authors might not have anticipated.

AI policy should aspire to something similar. Law schools should be ambitious about pedagogy and modest about regulation. They should be clear about what students must learn, flexible about how faculty teach it, and unwavering about the proposition that technology can assist human judgment without replacing the student’s responsibility to develop it.

Subscribe now

Share

McKinsey recently published a short “AI transformation manifesto” identifying twelve themes that separate companies genuinely transforming themselves with AI from those merely experimenting with it. A corporate consultancy is not always an obvious source of wisdom for higher education, and universities should not uncritically import its language or its metrics. But one claim in the manifesto generalizes well beyond the corporate context: AI transformation is not really about access to tools. The most powerful tools are increasingly available to almost everyone. The question is whether an institution has built the capacity to use them well, responsibly, and at scale.

That is where higher education risks getting off track. In my experience, many actors within universities appear to still be treating AI as a cluster of discrete problems to be addressed one at a time. Should students be allowed to use ChatGPT on exams? Should faculty receive AI training? Should the university buy an enterprise license? Should admissions experiment with chatbots? These are all real questions. But they are fragments. An advising chatbot is not a strategy. A prompt library is not a strategy. An AI exam policy is not a strategy. The deeper question is what it would mean to build an AI-ready university in the first place.

Governance, in this context, does not mean issuing policies or forming committees. It means building the institutional capacity to decide where AI should be used, who is accountable for those uses, and how the institution will learn from what it deploys. Those three questions are the engine; everything else is scaffolding.

This is the place where AI is doing something unwelcome to universities: it is exposing the gap between their stated identity as learning organizations and their actual capacity to learn institutionally. Universities talk about evidence, iteration, and the life of the mind. But their decision-making architecture—committees that meet periodically during a semester, policies written once and revised rarely, experiments that live or die on the energy of individual faculty—is not well matched to a technology that changes every six months. The problem is not that universities lack the intellectual resources to engage AI thoughtfully. The problem is that their institutional machinery was built for a slower world.

The sharpest line in the McKinsey document is: “no trust, no right to deploy AI.” For corporations, that is a useful principle. For universities, it is closer to a first principle and the test every AI decision has to pass.

Universities are trust institutions. Students trust them with their formative education and professional futures. Faculty trust them to protect academic freedom and scholarly independence. The public trusts them to produce knowledge, test claims, and serve the public interest. When a university deploys AI, the question is not whether the system works. The question is whether the deployment can be justified to the people subject to it, including students whose writing is evaluated, applicants whose files are screened, researchers whose data is processed, and staff whose work is reshaped. That standard is considerably higher than technical accuracy. It demands transparency about what the system is doing, accountability for what goes wrong, and humility about what the institution does not yet understand.

This is where the university case diverges from the corporate case in an important respect. A company that deploys AI badly loses customers or money. A university that deploys AI badly loses something harder to recover: the legitimacy of the relationships that make it a university in the first place. This realization raises the bar for what governance has to achieve. And meeting that bar requires three capacities that are really facets of a single institutional competence.

The first is knowing where AI belongs. Universities need a mission-centered version of what the McKinsey manifesto calls leverage points, the small number of places where an intervention can do the most good. For a law school, those points might be student learning, faculty research, career placement, and the administrative burdens that keep faculty from teaching and writing. For a humanities department, they might be student feedback and the slow work of archival research. The discipline is not to ask “where can we use AI?” but “where can AI most directly advance what we are here to do?” That single reframing rules out much of what currently passes for AI strategy.

The second is leadership fluency. AI transformation cannot be delegated to IT or to the handful of faculty who happen to have computer science backgrounds. Presidents, provosts, deans, and faculty leaders need enough AI literacy to make judgments that can bear real institutional weight. A dean does not need to become a machine-learning engineer. A dean does need to understand how AI is changing the professions students are entering, what kinds of judgment graduates will need to exercise, and what risks arise when student data or institutional decisions become entangled with AI systems. Similarly, faculty cannot opt out on the grounds that they are not “tech people.” AI now touches teaching, research, writing, and professional formation. Faculty governance must include AI governance because the two have stopped being separable.

The third is moving people toward higher-value work. This is perhaps the most uncomfortable capacity to discuss, because in corporate settings it often becomes a story about productivity and redeployment, which in universities may signal a need for cuts. But it need not be. The honest version is that universities are already stretched thin, and AI should be used to free faculty and staff for the work that should be most human: mentoring, judgment, intellectual synthesis, ethical reasoning, professional formation, and institutional stewardship. If AI handles routine correspondence, preliminary research, or administrative friction, the institutional goal should not be a smaller university, it should be one that spends a larger fraction of its time on work only humans can or should do.

Imagine a provost considering an AI-assisted advising system. The fragmentary response is to ask whether the vendor’s product is accurate, whether it integrates with the student information system, and whether legal counsel has signed off. The governance-capable response starts further back: What are we trying to accomplish that current advising cannot? Whose judgment does the system displace, and where does accountability sit when it errs? What will we tell a student whose academic path was shaped by it? What will we measure, and what would cause us to shut it down? Who in the institution has the fluency to answer these questions in a year, when the underlying model has changed twice? The second set of questions is harder, slower, and more expensive to answer. It is also what separates deploying AI from being ready to deploy it.

The AI-ready university will not be the one with the most tools or the most committees. It will be the one with the clearest sense of mission, the strongest governance capacity, the most AI-literate leaders, and the deepest commitment to its own stated ideals about learning. Universities exist to create, preserve, transmit, and challenge knowledge. AI now tests whether they can apply that mission to themselves.

Subscribe now

Share

Last week, I helped organize a symposium at Harvard Law School with my partner and colleague, Diane Lourdes Dick. Sponsored by the Harvard Journal of Law and Technology and Iowa Law’s Innovation, Business and Law Center, the event brought together twenty experts to address one of the most pressing questions in law and technology: what happens when AI systems stop assisting human decisionmakers and start acting on their own? The papers were excellent, the timing felt right given how much agentic AI now dominates the national conversation, and I’m proud of the program we put together. But what stayed with me most was not the content of the papers or the broader topic. It was the format of the event itself. The more I think about where academic life is heading in the era of AI, the more I suspect the live, in-person symposium is going to assume even greater importance.

One reason is straightforward. As Professor Frank Fagan observes in the Illinois Law Review Online, legal academia appears to be approaching a publication bottleneck. His basic claim is simple: law reviews are currently somewhere near equilibrium, with roughly 5,000 annual submissions competing for about 5,600 placements, but large language models could push the number of submissions much higher—toward 8,000—without any comparable increase in publication capacity. He predicts that journals will respond not with wholesale reform, but with additional reliance on author prestige as a sorting factor and the increased use of symposia as publication pipelines.

Whether or not the numbers reach the level Fagan predicts, the directional trend seems clear. We are entering an era in which article-length scholarship becomes much easier to produce but much harder to evaluate and place. That shift will mean that the scarce good in academia will no longer be the ability to generate polished prose. The scarce goods will be other things: the expert curation of topics and ideas, intellectual exchange and collaboration, and the real-time testing of arguments against other informed minds.

Those are the goods that a live symposium provides.

The written article is still important, of course, and it remains the legacy artifact that will be cited, taught, and archived. But in an AI-saturated environment, the article alone will tell us less than it used to. Smooth and clear writing will be cheaper, structural coherence will be easier to simulate, and citations, charts, and graphs will be easier to generate. A paper that looks polished on the page may not reflect deep thought so much as competent assembly.

A live symposium exposes the difference. It reintroduces friction into scholarly evaluation. In person, an argument must survive—and will likely benefit from—active engagement with other minds. You can hear when someone is genuinely working through an idea versus performing fluency. That kind of signal is difficult to fake and impossible to generate from a prompt.

One analogy that came to mind while attending our program at Harvard is the museum exhibition. Museum scholars have long argued that visitors should not be seen as passive recipients of institutional content, and I believe that this is a helpful way to think about symposia, too. A symposium is not just a delivery mechanism for papers any more than an exhibition is just a storage mechanism for objects. In both cases, curation is the key, and the value lies not only in the individual artifacts on display, but in the encounters—formal and informal—between those artifacts and the people who come to observe and engage with them.

And here is one area where AI may actually strengthen the modern symposium rather than weaken it. A big reason why symposia can feel intimidating is that the entry costs are high. A non-expert may not have time to read twelve dense papers in advance, and even scholars in adjacent fields may worry that they lack the technical background to follow every panel closely. But if attendees can use AI tools to generate reliable summaries and clarify unfamiliar concepts before the event, they can arrive better prepared to focus on what the live symposium does best: the conversations that happen on panels, in hallways, and over dinner. An economist, a sociologist, or a practicing lawyer may not read every paper in full before attending an AI-and-the-law symposium, but with good preparation materials, that person can still participate meaningfully in the discussion—and the discussion is richer for it.

This accessibility point matters, because it suggests that AI may lower the barriers to joining scholarly dialogue even as it raises the volume of written output. The same technology that makes article production cheaper may also make curated live engagement more inclusive. Here again the museum analogy is useful, though imperfect. The best exhibitions give visitors enough structure and context to orient themselves before they encounter the specific content on display. A symposium can aspire to something similar, and AI can help lower the pre-event burden of comprehension so that more value can be extracted from the live event itself.

I should note the obvious counterargument: if AI can help attendees prepare more efficiently, it could also help them perform more efficiently. Someone could show up with AI-generated talking points and rehearsed questions, simulating depth without possessing it. That is a real risk, and I don’t think it disappears. But I do think it is harder to sustain a performance over a full day or two of unscripted exchange than it is to polish a written draft at your desk. The symposium’s advantage is precisely its unpredictability.

Fagan is surely right that oversupply may harden prestige-based sorting in many corners of the legal academy, and he is also right that symposia may become a more important institutional response. But I would push the point one step further. The symposium is not merely a useful administrative adaptation to publication scarcity. It is now one of the most important forms of academic validation available in an era when writing alone is no longer a reliable signal.

Subscribe now

Share

My dad spent his life as a physician, and medicine, at its best, is an expression of disciplined wonder. You take the science seriously. You respect evidence. You do the work. But you also never stop being amazed that knowledge can expand, that new treatments can emerge, and that the future can be meaningfully better than the past. That’s how he approached AI, too. He saw potential in AI the way he saw potential in medical research, not as magic, but as a powerful tool that could widen the circle of capability. He was optimistic. He believed that science and technology, guided well, are among the most reliable ways we have to improve the human condition.

Yet he also believed—deeply, from decades of practice—that no tool replaces the person holding it. He was a doctor who sat at bedsides and made house calls. He listened. He knew that a diagnosis could be technically correct and still miss the patient entirely if it wasn’t delivered with care, with context, and with an understanding of who that person was and what they were facing. The science mattered enormously, but so did the human presence that gave the science meaning.

That conviction carried straight into his thoughts about AI. The technology is impressive, but it still needs a human asking the right questions, interpreting the results, and deciding what to do with them. AI can surface patterns or issues a person might miss, and it can process information at a scale no individual could match. However, the judgment, the empathy, the sense of what actually matters in a given moment, that remains human work. He understood this intuitively, because he spent a career learning that the best tools, drugs, and strategies are only as good as the hands and the heart guiding them.

He is gone now. I miss him every day. I miss our conversations about AI (and everything else). But I know what he’d say if I called him today with some new development or use-case: tell me more.

So, in that spirit, that is what I’ll keep doing, and I encourage everyone to follow my dad’s example. Keep learning, keep experimenting, keep asking questions, and keep making sure the human part of all this never drops out of the equation.

Share

Subscribe now

Universities are used to thinking about academic freedom in personal terms. The canonical picture is a scholar whose research or teaching becomes controversial and whose job security is threatened by politicians, donors, or administrators. Tenure, on that account, is a shield for the individual, and by protecting individuals it protects the university’s capacity to pursue knowledge.

That historical picture remains important, but it is increasingly incomplete. A growing share of scholarly work is now mediated by artificial intelligence systems that do more than speed up routine tasks. In many settings, AI is becoming the environment in which inquiry happens: a “virtual lab” or principal investigator that proposes hypotheses, deploys autonomous research agents, structures research strategies, evaluates drafts, identifies patterns, and runs tests. When AI becomes a distinct engine of knowledge production, the central vulnerability is no longer only the punishment of scholars; it is the quiet alteration of the means of inquiry itself.

This shift matters because universities are institutions that make long-horizon commitments. They build labs, libraries, and datasets; they develop standards, norms, and disciplines; they cultivate communities of expertise that persist across administrations and political cycles. When the infrastructure of research becomes dependent on AI systems that can be updated, tuned, restricted, or withdrawn with little notice, the conditions of knowledge production become fragile. A change that looks modest from the outside—an “update,” a revised safety policy, a different training set—can meaningfully alter what questions can be asked, what methods can be used, what outputs are possible, and how reliably results can be replicated. In that world, protecting the free search for truth requires governance that reaches beyond the faculty member and attends to the stability and integrity of the systems that increasingly structure scholarly work.

These and other academic freedom issues that accompany the rise of AI-mediated research are what led me to write my upcoming article, Algorithmic Tenure (forthcoming in the Georgetown Journal of Law & Public Policy). The titular phrase can be misleading if heard as a claim that algorithms deserve rights or personhood. The concept I introduce is not moral or metaphysical. Rather, it is institutional and functional. Tenure in the human context has always been defended as a means to certain ends. Its purpose is to secure conditions in which the university can pursue inquiry that is independent, long-term, and resilient to momentary pressures or prevailing orthodoxies. Algorithmic tenure takes that logic and asks what it would mean to protect those same conditions when an AI system becomes mission-critical research infrastructure. The goal is not to elevate AI; it is to govern the university’s own commitments responsibly when it embeds AI deeply into research and teaching.

Once we see AI as infrastructure, the risk profile comes into focus. A university may rely on a vendor-provided model or platform that it does not control, cannot fully audit, and cannot reliably stabilize. Vendor relationships are not inherently problematic, but they import a familiar set of governance problems into the heart of academic work. Commercial providers have their own incentives and constraints, including intellectual property concerns, liability exposure, market competition, and shifting regulatory demands. They can revise terms of service, restrict features, change pricing, alter models, or discontinue products. Even where no one acts in bad faith, the university’s dependence can become a kind of vendor capture where the institution is locked into tools that shape scholarly practice while remaining outside institutional governance.

Political pressure creates a different, and in some ways more troubling, risk. Higher education is already experiencing aggressive contests over curriculum and research agendas, especially on divisive social and political topics. AI introduces a subtler lever than the traditional subject ban or firing. If a state actor, a board of trustees, or an influential donor can pressure a university to tune, constrain, or decommission an AI system that supports “disfavored” scholarship, the effect can be a form of control that is both powerful and difficult to detect. The inquiry is not blocked by explicit censorship but is redirected by reconfiguring the tools scholars use to navigate, synthesize, and generate knowledge.

Administrative risk aversion provides a third challenge. Advanced AI systems will produce outputs that are sometimes wrong, sometimes unexpected, sometimes offensive, and sometimes controversial. Those realities generate predictable institutional anxiety, especially in environments where responsibility for AI harms feels diffuse. If it is unclear whether accountability rests with the user, the vendor, the designer, or the institution, decisionmakers may default to restrictions, shutdowns, or overly cautious guardrails. Over time, this can yield a governance pattern in which the most sensitive or high-value systems are treated as provisional and easily disrupted, precisely when the university should be building stability for the long term.

The concept of algorithmic tenure I introduce in my article responds to these risks with a simple institutional move: when an AI system becomes mission-critical to the university’s core functions of research and teaching, the institution should treat it as infrastructure that cannot be materially altered or withdrawn through opaque or purely ad hoc decisions. In practical terms, this means creating a rebuttable presumption against unilateral disruption and insisting on transparency, expert oversight, and an orderly process for significant change. A tenured AI system would not be frozen in place. It could still be modified, migrated, or retired. But those changes would occur only for adequate cause and through procedures that respect the university’s mission and the reliance interests that scholars, students, and research programs develop around stable tools.

To make this framework operational, algorithmic tenure calls for a governance structure analogous in spirit to the institutions universities already use for research oversight. Specifically, campuses rely on institutional review boards to evaluate research involving human subjects, and many use specialized committees to oversee concerns like conflicts of interest and research integrity. The point of those bodies is not to inhibit scholarship but to ensure that essential values—safety, accountability, and ethical conduct—are integrated into the research enterprise. A comparable body for mission-critical AI can serve a similar role, bringing technical competence and institutional legitimacy to decisions that might otherwise get made by procurement offices, IT staff, risk managers, or individual administrators reacting to the crisis of the week.

An AI ethical review board, properly designed, would not be an all-purpose censor of tools or ideas. Its mandate would be narrower and more structural: to identify which systems qualify as mission-critical, to establish baseline expectations for documentation and change control, and to oversee the integrity of those systems over time. It would evaluate whether a system’s use is so central to key research programs or teaching functions that the university has an obligation to ensure continuity, transparency, and responsible governance. Some systems would not qualify, and that is as it should be. The goal is not to confer a special status on every chatbot or productivity tool, but to recognize that certain systems become foundational in ways that warrant procedural protection.

Because mission-critical status should not be instantaneous, algorithmic tenure also envisions a probationary period in which a system is evaluated along dimensions that are familiar to both compliance professionals and academic leaders. A serious review would consider technical performance and robustness, privacy and security posture, documentation and reproducibility, and the alignment of the system’s constraints with academic freedom and nondiscrimination values. It would also ask a more basic institutional question: whether the system’s role is truly infrastructural, meaning that its disruption would materially compromise core programs, long-term research trajectories, or essential teaching functions. The probationary period matters because it prevents hasty entrenchment while still creating a path to stability once reliance becomes justified.

Contracts are the next essential layer. If a university relies on third-party AI, governance cannot stop at campus committees and must shape the terms on which the vendor relationship is built. For mission-critical systems, the institution should negotiate for transparency around material updates, meaningful documentation, audit rights where feasible, and continuity commitments that prevent sudden withdrawal or silent transformation. Universities often believe they lack leverage to bargain in this context, and sometimes they do. But leverage increases when standards are coordinated. Federal funders and accreditors can play a role here by treating “research-grade AI” as a condition of participation in major funded research projects or accredited educational programs. When external standards require responsible governance, vendors have incentives to build and maintain systems that can be reliably used for serious scholarly work.

Finally, algorithmic tenure requires a clear approach to change over time. AI systems will evolve quickly. The best governance response is not to pretend that evolution can be stopped, but to insist that evolution be managed. For tenured systems, material changes should trigger a process that looks like due process: written reasons for significant modifications, an evidentiary basis for claims about risk or obsolescence, and expert review to assess whether proposed changes are necessary and proportionate. Obsolescence, in this setting, should not mean that something newer exists. It should mean that the system can no longer be responsibly maintained in light of security, integrity, or performance constraints. When replacement is warranted, the institution should pursue an orderly migration, with documentation sufficient to preserve continuity, allow reproducibility where possible, and minimize disruption to ongoing scholarly projects.

Ultimately, protecting inquiry in an AI-mediated research environment is not only about protecting scholars and speakers; it is also about protecting the stability and transparency of the tools that structure thinking and discovery. Algorithmic tenure is an effort to update the institutional architecture of academic freedom for a world in which the conditions of inquiry can be altered without any visible punishment of individual scholars. My proposed framework asks universities to take responsibility for AI as infrastructure and to build a governance system capable of sustaining the independence, stability, and accountability that serious scholarship demands.

Look for the final published version of Algorithmic Tenure later this year.

Subscribe now

Share

I recently saw a major law school moot court competition impose a blanket ban on generative AI. The rule is clear and well-intentioned, prohibiting students from using AI at any stage of the process, including for brainstorming, organizing, brief writing, or question generating. While the policy comes from a good place, it is out of step with the world our students are entering.

Let’s start with what the ban gets right. The organizers clearly care about fairness, integrity, and skill-building. These are nonnegotiable values in any moot court program. There are real concerns that if some teams can afford better AI tools, then the playing field will tilt. The instinct to level it by removing the tool altogether is understandable. Furthermore, the risks of using generative AI are significant. Models can “hallucinate” cases or introduce unattributed ideas, problems that are antithetical to a competition that prizes accuracy and original analysis. Above all, there is the fear that AI will become a shortcut, allowing students to bypass the slow, difficult work of reading the record and performing the rigorous legal analysis that is essential for forming professional judgment.

These concerns are all real and true. But an outright ban is not the answer.

A key benefit of moot court (and trial advocacy) is that it gives students a chance to practice law before they are actually in practice. In the modern professional environment, lawyers are already expected to be “humans in the loop” with AI, just as they have long been with tools like Westlaw and Lexis. Training students to advocate in a sterilized, no-AI environment is like making medical students practice only with wooden scalpels. It preserves ritual at the expense of realism.

A ban also postpones the hard work of resolving the questions that every modern legal department is already facing. When is a model’s paraphrase too close to the original? What does verification look like when citations are surfaced by automation? How much can you rely on a machine’s summary before you must go to the source?

When used appropriately, AI can even raise the quality of analysis and argumentation. Students can use it to generate counterarguments and rehearse judicial questions at lightspeed, freeing them up to spend more time strengthening logic and prose. Competition judges will still see the human skill at work in these programs; they will just see it reflected in a more fully tested final product.

If we want to graduate lawyers who are wise with their tools and not ruled by them, we must integrate those tools into our training.

For example, teaching students in moot court about prompt literacy will lead to sharper issue framing. The act of posing precise questions to a model forces students to gain clarity on their theories, burdens, and standards. Indeed, good prompts are miniature issue statements. Learning to write them teaches students to think in the structures that judges will use for evaluation and scoring.

Likewise, model skepticism will lead to stronger arguments. Students who practice interrogating machine outputs—asking “why does this follow?” or “what are the best contrary views”—will develop the reflexes of effective advocates. They will become better at answering hostile questions and identifying weaknesses because they will have already tried to break their own arguments against an increasingly sophisticated opponent.

A full prohibition of AI trades away the very skills that moot court exists to cultivate. The ban suggests to students that the real work of being a lawyer is what happens when you are alone with your mind and a blank page. That is partly, and beautifully, true. But the fuller truth is that real legal work happens when you are alone with your mind and the tools of your era—and must then decide what is accurate, what matters, and what persuades.

Moot court should be the crucible for developing that kind of judgment. Design it for anything less, and we risk turning legal training into historical reenactment: charming and careful, yes, but a bit beside the point.

Share

Subscribe now

The legal profession is racing toward an AI-augmented practice. That doesn’t mean we throw out what works. It means we double down on the human skills that will matter more as machines get better at everything else.

Accordingly, this year I’m making a few shifts in how I teach, all centered on two principles: first, to preserve and hone the skills that make lawyers valuable; and second, to integrate AI where it improves judgment and fence it off where it does not.

Before the changes, a word about what stays the same. I’m not backing away from discussion and the Socratic method. Being able to think aloud, pivot under pressure, and defend a position with incomplete information is exactly what AI will not do for students in courtrooms, boardrooms, or client meetings. I will continue to structure classes around live dialogue. I’ll be explicit about the purpose of these exchanges: to strengthen attention, preparation, judgment, and poise under uncertainty.

From there, the first change is to introduce short, in-class writing opportunities without laptops. Reflection and synthesis are different skills than text production. AI accelerates the latter; I want to harden the former. Several times a week, I’ll pause discussion for a few minutes of “analog” work. Sometimes it will be a reflection exercise: write down the major takeaways and any lingering questions from the class discussion so far. Other days it might be a hand-drawn flowchart that maps a specific test or analytical framework taken from the assigned cases, statutes, or model rules. These are low-stakes exercises collected for completion, not grades. The goal is simply to teach students to see their own thinking on paper without the help of digital tools.

A second change is the introduction of AI auditing. Periodically, I’ll supply an AI-generated case summary or a short argument and ask students to manage it like a junior associate’s draft. The task has three parts: identify omissions or misreadings, verify citations against primary sources, and rewrite the weakest paragraph so it would survive a skeptical reader. The point here is not to gawk at “hallucinations,” but to build the professional reflex of trust-but-verify. If AI is going to be in the room in practice—and it will be—students must learn to manage it with the same rigor they would bring to a human colleague’s work.

Third, I’m adding speed judgment drills. I’ll pose a rapid-fire hypothetical and ask for one deliverable: “What would you do now, and why?” Students will have thirty seconds to form a view and one minute to defend it. The drills are not about theatrics. They are about learning to make timely, defensible recommendations under ambiguity. This is a core habit in practice and a skill that AI cannot fully replace.

Fourth, I’ll be building AI-assisted role-play exercises around the kinds of short-notice meetings students will face in practice. In a typical scenario, one student will rely on AI output to summarize the law; another will play the lawyer who must translate that text into risk-tiered options with pros and cons; and a third will play the business lead who presses on time, budget, and reputational constraints. Afterward, I’ll do a short oral check-in with the students to discuss what went well and what they found challenging about the situation.

Finally, I’ll be introducing compare-and-contrast research games. Teams will receive the same narrow question. One team will work with traditional sources only; the other may use AI. After several minutes, they will trade places so that everyone experiences both approaches. We will talk openly about where AI helped and where it presented obstacles or limitations. These exercises are not a referendum on technology. They are a way to make the limits and advantages of available tools legible to students so they can choose wisely when the stakes are real.

These pedagogical changes will require some policy updates. The biggest one is that I am going to require my students to have access to ChatGPT, either the free or a paid version. If access is a barrier for any reason, students will be asked to contact me so we can arrange an alternative path that preserves the learning goals without creating inequity.

Admittedly, this is all very much a work in progress. I’m experimenting. The changes I’m contemplating reflect an initial attempt to align modern legal training with the world our students are entering, a world where powerful tools accelerate certain tasks while placing a premium on distinctively human skills like attention, judgment, synthesis, and live communication. My job is to strengthen those human capacities and to teach students to manage responsibly the tools available to them.

If you’re experimenting with similar changes, I’d love to compare notes. What is working in your classrooms? Where is AI genuinely adding value, and where is it creating problems or concerns?

My hunch is that the answers will vary by doctrine and by assignment. But the larger point holds: if we teach students to think clearly, communicate sharply, and make sound recommendations under uncertainty—and if we show them how and when to pair their judgment with an AI assistant—then we will be doing our jobs, and doing them quite well.

Subscribe now

Share

In my last post, I explained what agentic AI is: autonomous, goal-seeking software that can plan, act, and collaborate across multiple systems, often without waiting for human prompts. These systems represent a leap beyond dialogue-driven tools like ChatGPT. Agentic systems are not just answering our questions. They are taking initiative, anticipating needs, and completing projects on their own.

This post begins to unpack what that means for lawyers and the law schools that train them. Because here is the crux of the matter: the rise of agentic AI does not just mean more powerful software—it signals a shift in professional identity. Lawyers will increasingly be expected to supervise, instruct, and collaborate with nonhuman actors. Attorneys won’t only need to know the law; they will need to direct and monitor autonomous digital agents. Tackling that challenge requires honing a different skillset than what most law schools currently teach.

To be sure, legal technology has long promised to make lawyers faster and more efficient. I still remember the speed advantages of using Westlaw for the first time after initially being trained to navigate hardbound reporters and treatises. But the gains from agentic AI go farther. Agentic AI does not merely assist lawyers, it begins to act like one.

For instance, imagine an AI system that doesn’t need to be prompted with “please summarize these contracts,” but instead constantly monitors a client’s document database, flags potential issues proactively, and routes high-risk items to human attorneys for review. Notably, agentic capabilities along these lines are not just theoretical. Consider Harvey. Used by major firms like A&O Shearman, Harvey began by performing legal research, issue-spotting analyses, and drafting tasks, but it is now turning to solutions that focus on multi-agent tools “that analyze and triage documents the same ways that a firm’s human lawyers do.”

In this new agentic environment, the primary challenge will not be using the tools available but overseeing them. That includes auditing outputs, identifying errors in reasoning, assessing when human judgment must intervene, and ensuring that the AI’s autonomy aligns with the client’s interests and the lawyer’s professional and ethical obligations. If these tasks sound more like being a manager than a technician, that’s because they are. And if lawyers are going to manage teams of autonomous systems, then we need to start training them that way.

Here’s the core of what law students will need to master to manage autonomous systems responsibly and effectively:

1. How to give strategic direction to AI agents.

Not every agent needs step-by-step instructions, but they all need purpose. Lawyers must learn to define objectives, set constraints, and articulate risk tolerances when deploying agentic AI to legal tasks.

2. How to audit outputs and spot red flags.

“Checking the work” of AI will be a core competency. Students need practice spotting subtle reasoning errors, hallucinated cases, and statistical distortions—especially when an AI agent’s work product is cloaked in confident prose.

3. How to coordinate human–AI teams.

Students must learn to manage workflows where AI agents and people collaborate. Doing so will include knowing when to delegate, when to override, and how to preserve ethical responsibility across distributed systems.

How might these skills manifest in a law student’s later professional journey?

Consider a recent graduate hired as an associate at a mid-sized firm who is supervising a multi-agent legal assistant. One agent scans newly signed vendor contracts for indemnity clauses. A second agent compares those clauses against the client’s existing insurance coverage. A third flags mismatches and routes them for human review. The associate’s role in this scenario will be to set the rules of engagement, check for problems, and document how each agent’s findings influence the final advice she provides to her supervising attorneys or the client. In short, the associate is practicing law, but she’s doing so primarily as a manager of intelligent systems.

As we go forward with this new landscape in mind, we do not need to turn law schools into computer science departments, but we do need to make sure that AI training is not thought of as being limited to the occasional seminar or a skills-based elective. Managing agentic systems will soon be central to many students’ daily routines. Law schools must respond by embedding these skills into the core curriculum. Some ideas include:

• Professional Responsibility courses that feature modules on the oversight of autonomous agents and the ethical delegation of judgment.

• Legal writing and research courses that incorporate agent-auditing exercises into assignments like reviewing and correcting AI-generated drafts of contracts or correspondence.

• Clinical programs that pilot the use of agentic AI under supervision, letting students manage real-world tasks with AI assistance followed by reflection on its limits and potential liabilities.

• Collaborations with law firms and legal departments that help identify the specific agentic AI competencies employers are prioritizing so that schools can ensure students are developing those skills before graduation.

In sum, agentic AI changes not just what lawyers do, but how they think about responsibility, delegation, and professional identity.

Building student competence in this space is not a luxury. It is mission critical.

Subscribe now

Share

2025 is increasingly being referred to as the year of agentic AI. That is because within the past six months, artificial intelligence has taken another major leap forward, not just in terms of speed or accuracy, but in form. What was once a tool to enhance productivity and generate efficiency gains is becoming something much more powerful: a system that acts on its own.

Agentic AI refers to autonomous, goal-driven systems capable of planning and executing extended sequences of actions, often without human intervention. These systems go well beyond chatbots responding to direct user prompts. Instead, they are software entities capable of taking initiative, coordinating with other tools, and adapting to dynamic conditions in pursuit of a defined objective. They therefore represent a marked shift in this space by redefining how action and judgment are delegated in digital environments.

In this post, I will provide a brief primer on the basics of agentic AI. Then, in subsequent essays, I will focus more directly on how this technology will affect legal practice and education.

To begin, while many people use “AI agent” and “agentic AI” interchangeably, there is an important distinction. AI agents are software tools assigned to carry out a specific function, usually within a bounded enterprise system. Think of them as task performers. An AI agent might perform a credit check, draft an email response, or review warehouse inventory. These agents are often powered by large language models (LLMs) but operate with a narrow scope. They also generally have limited ability to learn over time. Their goal is predictable execution, not autonomous innovation.

Agentic AI, by contrast, is an umbrella term used to describe systems that can either coordinate and direct multiple agents or act independently to pursue long-term goals using memory and dynamic reasoning. These systems are not simply executing a task; they are determining what tasks to perform, when to do so, and how to sequence or prioritize them. As Louis Gutierrez, the director of AI at Constant Contact, explains, “If agents are individual players, an agentic system is the coach, team, and playbook all working together.”

Already, agentic AI is beginning to transform fields as diverse as healthcare, logistics, finance, software development.

For example, at Moody’s, the company uses a network of 35 autonomous AI agents to support its financial research operations, including the performance of tasks like industry comparisons and analyses of SEC filings. The AI agents operate within a multi-agent system, where some agents perform task-specific functions (e.g., project management) while others are assigned supervisory roles. Each agent is also given its own “personality,” instructions, and access to distinct datasets, enabling the system to generate divergent perspectives on complex issues like the stability of firms facing specific geopolitical risks. Indeed, the company’s approach reflects a growing trend of agentic architectures that embrace disagreement, specialization, and coordination.

Similarly, Johnson & Johnson is using autonomous AI agents to streamline a critical step in pharmaceutical development—the “solvent switch”—where one chemical solvent is replaced with another to crystallize a molecule and produce a viable drug. Traditionally, this process required repeated manual testing to account for a range of variables, from temperature to reaction timing. Now, J&J’s agentic system analyzes those variables and determines the optimal moment to execute the switch, dramatically reducing trial-and-error cycles. According to the company, the agents bring together content, context, and timing in a way that supports high-stakes decision-making in real time. This is a powerful example of agentic AI not just automating work but augmenting scientific judgment at industrial scale.

As promising as agentic AI may be, it also introduces entirely new layers of complexity and risk. A recent Harvard Business Review analysis put it bluntly: most organizations are not ready for agentic AI. They have not built the internal infrastructure—technical, procedural, or ethical—to manage this shift responsibly.

Thus, while agentic AI clearly represents a profound turning point in how we design, supervise, and share responsibility with machines, the decisions that institutions make in the next few months about governance and policy will shape the long-term trajectory of this technology and its impact on our lives.

With these concerns in mind, I will soon turn to what agentic AI means for law, both in terms of practice and in the classroom. Some of the most urgent questions that lawyers, law students, and law schools alike should be thinking about include:

• What new ethical duties will lawyers assume as they train, deploy, and oversee AI agents?

• How should law schools prepare students for a world where they are not just advocates and counselors but also “bosses” to nonhuman agents?

• What happens when AI agents begin to exercise judgment once reserved for human professionals, including in situations where AI stops waiting for our instructions?

• How should we assign legal responsibility for autonomous decisions made by AI?

The legal profession was built on the human exercise of reasoned analysis and judgment. Agentic AI challenges that foundation while also offering us an opportunity to rethink it. If we build wisely, agentic AI could unlock new forms of human-machine collaboration that reshape the study and practice of law for the better.

Stay tuned.

Share

Subscribe now

In recent years, artificial intelligence has begun reshaping how legal work gets done, from research and writing to client interaction and regulatory compliance. These developments offer law schools a powerful opportunity to enhance and refine their teaching and scholarly missions in ways that not only reflect these changes but help shape them responsibly.

To that end, I outline below a proactive agenda for law schools for the 2025–26 academic year. This agenda comes from a perspective that sees AI not as a threat, but as an opportunity to deepen legal education’s value, relevance, and public impact. These suggestions are not rooted in fear of being left behind (though that remains a severe risk for schools doing nothing in this space). Rather, they reflect the simple premise that the best version of legal education is one that prepares students for the work they will actually do while giving faculty the tools and encouragement to lead the way.

But there is a deeper lesson here too, one that speaks directly to the kind of lawyers the modern law school should aspire to train. As a recent Fast Company article put it, “AI is coming for your job directly. Not with fanfare or grand announcements, but through silent, pervasive creep.” It is not enough to simply use AI tools or teach students how to prompt a chatbot. What is needed is something more ambitious: the ability to cultivate professionals who can out-evolve the automation around them. Accomplishing that goal will require building habits of mind that resist intellectual atrophy. It will also depend on designing environments where law students learn to combine emotional intelligence, human judgment, and creativity in ways AI cannot easily replicate. That is the spirit behind this agenda:

1. Leadership Must Model AI Use

“Ultimately, AI starts as a leadership problem, where leaders recognize that AI presents urgent challenges and opportunities,” Ethan Mollick writes in Making AI Work: Leadership, Lab, and Crowd.

Leadership must go first. And that begins with the dean. As the symbolic and strategic head of the institution, the dean’s example carries unique weight. If the dean is silent or skeptical about AI, then others will take their cues and hesitate. But when deans model thoughtful, transparent use of AI in teaching, scholarship, and administration, they create space for innovation to take root across the community.

Deans, along with influential faculty, must therefore speak openly about the ways AI can enhance legal education. They must frame AI not as a threat to professionalism, but as a skill to be mastered and a context to be governed. As Mollick argues, “leaders need to start to anticipate how work will change in a world of AI” and provide their communities with “clear and vivid images of what the future actually looks like.”

In short, the first rule of institutional change is cultural permission. Without visible, intentional leadership from the top, nothing else will follow.

2. Treat the School Like a Lab

AI is evolving faster than any single model, syllabus, or academic paper can track. That is why law schools should treat themselves as live laboratories. They should let students experiment with AI in legal writing, research, and litigation simulations. They should encourage faculty to co-author with AI, test new pedagogical uses for it, and assess its limitations. They should create interdisciplinary working groups to vet tools and share findings across campus.

These activities will help teach students how to use AI without being used by it. By exposing students to trial-and-error experimentation and judgment in supervised settings, law schools can help them remain the “human in the loop,” one of the most valuable roles in any AI-enabled practice or work setting.

3. Tailor the Curriculum

AI should not be categorized as a boutique or novelty topic. It is becoming a core feature of modern legal practice, and curricular development should reflect that evolution. First-year research and writing courses must teach students how to collaborate with, critique, and improve AI-generated legal outputs. Upper-level electives should include courses along the lines of “Lawyering with AI,” “AI and Professional Responsibility,” and “Regulation of AI Systems.” Doctrinal courses should incorporate AI use in hypotheticals, briefing, and policy analysis.

A recent study on AI-Powered Lawyering by Daniel Schwarcz (Minnesota Law) and colleagues shows that law students trained on several contemporary AI tools performed legal tasks more accurately, more clearly, and more persuasively than their peers. The biggest gains came when students were trained to use the tools thoughtfully.

Thus, the necessary curricular reform is about more than simply jumping on a trend. It is about recognizing that lawyers who treat AI as just another shortcut risk turning themselves into templates—and templates are easy to automate. Legal education should emphasize the kinds of professional capabilities that are harder to replicate, such as empathy, personal relationship-building, and social and emotional intelligence.

4. Build Infrastructure for Real Practice

To be clear, law schools need to do more than just teach about AI, they must also let students use it in meaningful, supervised, practice-ready contexts. This means integrating AI tools into clinics for intake, case strategy, war games, translation, and document generation. It also means building simulation-based labs where students can interact with AI during client exercises.

In the Schwarcz study, AI use boosted productivity by as much as 140% on complex tasks like analyzing complaints and drafting persuasive legal letters. Those are real-world deliverables, and students should graduate having produced them, with AI by their side.

5. Prioritize Ethics and Governance

AI is not just a productivity enhancer; it is also a policy and ethics challenge. Lawyers will be at the center of emerging questions about fairness, misinformation, access to justice, and democratic oversight.

Accordingly, law schools must prepare students not only to use AI, but to govern it. They need to embed discussions of AI ethics in required courses. They should consider creating joint-degree or certificate programs for students focused on public interest technology and AI policy more broadly. Perhaps most critically, law faculty should be producing scholarship that helps shape how AI is used and regulated.

As we go forward, the question for legal educators is not whether AI will change the study of law or the profession. AI is already transforming both. Instead, the question is whether we will face this new reality with the necessary leadership, creativity, and rigor.

That is the work now and for the coming year. It cannot wait. The law schools that act decisively will shape the future of legal education. Those that hesitate risk being shaped by it.

Share

Subscribe now

Earlier this year, Northeastern University student Ella Stapleton discovered that her professor used ChatGPT to generate course materials, even though his syllabus expressly prohibited students from using AI to complete their own work. Stapleton saw this as hypocritical and raised the issue with Northeastern's business school. She filed a formal complaint and asked for a refund of the $8,000 she spent in tuition for the class. Northeastern eventually denied Stapleton a refund the day after she graduated.

The story, reported in The New York Times, spread across social media for obvious reasons. It seems to tap into a growing student perception that universities are imposing a double standard when it comes to AI. Students are restricted and sometimes disciplined for using generative AI tools, while at the same time their faculty might be using those resources to create syllabi, quizzes, or even entire lectures. Indeed, one recent survey found that nearly 40% of instructors in higher education now use generative AI regularly.

So, are the complaining students right to be upset? Is there a double standard at play?

The answer to both questions is no. There is a critical and defensible difference between student and faculty use of AI, and the distinction rests on the fundamental purpose of education.

Students are learners. The tasks and assignments they complete are meant to reflect their development in areas like research, analysis, critical thinking, and communication. When a student uses AI to bypass this process, she shortcuts the very learning that her coursework is designed to promote.

Professors, by contrast, are educators. Their role is to create the conditions under which learning can happen. When faculty use AI to generate quizzes, hypotheticals, or simulations, they are not evading their responsibilities. Rather, they are exercising professional judgment about how best to allocate time and effort toward student success. If an AI-generated problem helps students learn, and the instructor has vetted it for quality, then the tool has served the educational mission. In fact, AI can be a powerful way for faculty to become more efficient, freeing up time for more research or to provide more individualized feedback and mentoring.

Still, student skepticism is not entirely misplaced. There are real risks when AI is used without care. If faculty treat AI as a way to automate instruction, then the experience of education will suffer for everyone. Notably, a recent study by researchers at Microsoft and Carnegie Mellon found that humans who relied on AI and trusted its output too readily engaged in less critical thinking. That danger applies just as much to teachers as to students.

So, yes, universities should set clear policies around AI. But more than that, they should articulate a clear ethical framework for AI use, one that recognizes the asymmetry between student and faculty roles while also holding everyone to the same basic standard that AI must serve learning, not replace it. When students use AI to do their thinking for them, it is a form of academic dishonesty. When professors use AI to design better environments for thinking, it is a form of professional adaptation.

One of these use cases undercuts education. The other, when done right, can advance it.

Share

Subscribe now

The American Bar Association wants to double down, literally, on experiential education.

A new proposal from the ABA Standards Committee would require law students to complete twelve rather than six credits of experiential coursework, with at least three of those credits coming from clinics or field placements. The idea sounds student-centered and practical. In reality, however, it’s an expensive, rigid, and intellectually shortsighted response to a world that is changing faster than the ABA seems willing—or able—to admit.

Let’s be clear: the legal profession is not suffering from an influx of new attorneys who lack sufficient grounding in simulation exercises or clinical experiences. Rather, it’s facing an existential shift brought on by artificial intelligence, automation, and radically new business models for the delivery of legal services. Graduating lawyers won’t be passed over because they haven’t done enough mock depositions. They’ll be displaced because AI will soon be doing more of what they were trained to do, only faster and cheaper. In the face of that reality, the ABA’s proposal feels less like progress and more like an act of institutional nostalgia tied to the pre-AI world.

Keep in mind, too, that clinics and experiential learning opportunities are among the most expensive forms of legal education to provide. They require low student-faculty ratios, heavy administrative and logistical support, and significant faculty time. Whether they realize it or not, students pay steeper tuition because of needing six experiential credits. Doubling that number effectively mandates more debt for most students and gives them fewer choices in their own education. Even if law schools don’t immediately raise sticker tuition in response to the new proposal, the real cost per student of running a clinic and field placement program will remain far higher than traditional doctrinal courses. These cumulative costs must be absorbed somewhere, and law schools will likely pass them on indirectly through higher fees, fewer merit scholarships, fewer faculty hires, and reduced investment in electives, technology, career services, student mental health resources, and so on. And let’s not kid ourselves: the ABA isn’t offering to pay for any of this. It’s demanding more from students and their increasingly resource-stretched/starved law schools while offering nothing in return.

What the ABA should be doing is encouraging new kinds of experiential learning opportunities that prepare students for the professional world they’re actually entering. That world is being rapidly reshaped by AI. Tools like ChatGPT, Claude, and Gemini are changing everything from litigation strategy to access-to-justice models. Law students don’t need more opportunities to write client memos. They need hands-on experience with legal tech, algorithmic reasoning, and the ethical frameworks required to navigate machine-augmented practice. If the ABA were serious about the future, it would be incentivizing courses that put students at the frontier of legal innovation. Instead, it’s codifying a model better suited to 1995 than 2025.

The ABA’s role should be to empower innovation, not to smother it with outdated, unfunded mandates. Law schools need flexibility to adapt, to make the best strategic uses of their scarce resources, and to serve diverse student populations with varied career trajectories. By imposing an expensive and pedagogically narrow requirement, the ABA is doing the opposite.

The ABA’s desire to expand experiential learning requirements isn’t reform. It’s regression. And it’s happening at the expense of the very people legal education is meant to empower.

Share

Subscribe now

In legal education today, the buzzwords are unmistakable: minimum learning outcomes, formative assessment, student-centered learning. Each is part of a broader movement to make law school more predictable and, above all, more comfortable.

Of course, the underlying goals of supporting students throughout their education and ensuring that every graduate leaves with the competencies required to enter practice are noble and important. But in chasing those aims, something essential is under threat: the cultivation of professional independence.

Perhaps the most underappreciated virtue of traditional legal pedagogy is that it fosters something no AI tool can replicate: resilience. In contrast to an educational model built around frequent feedback and detailed rubrics, the traditional approach resists handholding. It embraces uncertainty, and it demands that students do the difficult, often uncomfortable work of learning to think for themselves.

Formative assessment has its place, but it also gets close to training habits of dependence. When students come to expect validation and correction at every turn, they risk mistaking the metrics for the mission. Instead of grappling with the law’s inherent ambiguity, they focus on alignment with the teacher’s expectations. Instead of wrestling with hard questions, they reach for the model answer. The result, paradoxically, can be a form of intellectual fragility. Students become experts at solving well-defined, pre-packaged problems but less capable of navigating the undefined, the novel, and the complex—precisely the types of problems the best lawyers are called to solve.

This discussion matters more now than ever. In a world where AI can draft contracts and generate plausible legal memos in seconds, what distinguishes a good from bad lawyer is no longer mechanical competence. It is judgment, and it is discernment. It is the ability to see the terrain of a problem no one has yet mapped and then reason one’s way through it with confidence and vision. Those qualities require struggle in order to develop. And they require a pedagogy that, at key moments, withholds the answer.

Traditional legal education—Socratic dialogue, open-ended hypotheticals, issue-spotting final exams—has long been caricatured as an exercise in stress or intimidation. But that discomfort is not gratuitous, it is strategic. It simulates the professional conditions students will soon face, where clients will not provide rubrics and courts will not give feedback or assessments until it is too late. In those moments, lawyers must rely not on ready-made solutions, but on internalized frameworks for reasoning and argumentation. They must be able to stand on their own intellectual feet.

The case method, for all its imperfections, asks students to practice doing just that. It teaches them how to think in a world where there may be no right answer, only better and worse arguments. It asks them to make choices under pressure, to speak before they are certain, and to revise their views in response to new facts and perspectives. These are the very conditions in which legal judgment is formed.

None of this is to say that innovation in legal education is misplaced, or that we should not offer academic support for students who are truly struggling. But if we are serious about preparing lawyers for a world transformed by technology, we should think twice before jettisoning the very teaching strategies that nurture the human traits machines cannot emulate or instill: independence, imagination, grit.

These strategies are not accidents of tradition. They are its greatest gift.

Share

Subscribe now

Law schools today face mounting pressure from accreditors and internal reformers to adopt a more explicitly vocational mission. Frequent formative assessments, narrowly articulated learning outcomes, and an emphasis on discrete skill acquisition are increasingly framed as the marks of a “modern” and “student-centered” legal education.

At the same time, a parallel critique has taken hold: that law schools often over-prioritize research and scholarship at the expense of teaching. According to this view, scholarship is a luxury—perhaps even a self-indulgence—engaged in primarily by tenure-track faculty whose real job should be training practice-ready lawyers.

But this critique misunderstands the demands of an AI-driven future.

As artificial intelligence reshapes the practice of law, the most valuable human lawyers will not be those who can execute predefined tasks with maximum efficiency. Those tasks—drafting contracts, synthesizing cases, generating predictive analyses, and so on—will increasingly be automated. Rather, the lawyers destined to thrive will be those who can see around corners, challenge assumptions, and invent new frameworks for legal reasoning and action amid complexity. And the best incubator for those abilities is not a steady cadence of practical exercises. It is scholarship—the creation of it, the examination of it, and the deep engagement with it by faculty and students alike.

This insight explains why it is fundamentally misguided to characterize scholarship as a retreat from the real world or as incongruent with the law school’s critical teaching mission. When viewed objectively and through the lens of history, we see that faculty scholarship is the ideal proving ground for precisely the kind of strategic and imaginative thinking the future world will prize in its legal professionals. Good scholarship demands that its author synthesize vast information under uncertainty and surface hidden patterns. It calls for persuasive, novel argument that can withstand scrutiny from skeptical audiences. Any law school that shrinks from its scholarly ambitions to concentrate more heavily on skills-forward training risks producing graduates who will be prepared to perform only the very tasks that AI will soon absorb.

Moreover, the stakes go far beyond skills and employment. As AI becomes embedded in the infrastructure of legal decision-making—from bail determinations to predictive policing to corporate risk analysis—scholarship will be essential to identifying and probing its normative and descriptive foundations. This realization gets to the core of what the modern university was designed to do in the first place: to safeguard spaces for deliberate thinking, for dissent, and for the production of new knowledge that challenges prevailing orthodoxy. Legal education must participate in this intellectual tradition, not in spite of its professional mission, but because of it.

Indeed, legal scholarship is rooted in rigorous research methodologies that few other parts of the legal academy or profession consistently teach or model. Tenure-track faculty in particular are trained to navigate and critique empirical evidence, develop cutting-edge interdisciplinary insights, and marshal historical, comparative, or philosophical perspectives that contextualize the law’s development. These are not merely academic affectations—they are habits of mind that detect blind spots in dominant frameworks, resist misleading narratives, and identify normative concerns that automated systems often obscure. While other faculty bring indispensable strengths in teaching and practice, their roles rarely emphasize the production of original research, the development of new legal theory, or the interrogation of law’s epistemic foundations. Without tenure-track scholars dedicated to advancing those pursuits, the law school risks losing its role as a producer, not just a transmitter, of legal knowledge.

The choice thus becomes stark. Law schools can either treat scholarship as a dispensable luxury, graduating technicians destined to be eclipsed by code, or they can recognize it as the lifeblood of the discipline, an engine of discovery that will preserve legal education’s status as a principal forum where society learns to see—and shape—the future before it arrives.

Share

Subscribe now

In recent years, legal education has undergone a quiet revolution. Under pressure from the American Bar Association and other reformers, law schools are encouraged, if not outright required, to adopt new pedagogical models. Frequent formative assessments, tightly articulated learning outcomes, and reduced reliance on a single final exam have become the new orthodoxy. The traditional model of legal education—dominated by the Socratic method and that one big exam—has come under heavy criticism for being outdated, opaque, and insufficiently supportive of student learning.

However, as artificial intelligence begins to reshape the practice of law in profound ways, we should ask whether these reforms are training students for the wrong future. Paradoxically, the traditional model, so maligned today, may be exactly what the next generation of lawyers needs.

The modern push for continuous formative assessment is premised on a laudable desire to improve student learning. Frequent low-stakes evaluations provide opportunities for feedback and growth, and publishing specific learning outcomes helps clarify expectations. But embedded within these reforms is a subtle vision of legal practice: one in which competence is about mastering discrete skills, proceeding step-by-step with constant calibration, and succeeding through an accumulation of checked boxes. This approach is increasingly misaligned with the emerging realities of an AI-augmented legal profession.

Artificial intelligence is already remarkably good at managing discrete tasks like drafting standard contracts, checking citation formats, generating case analyses, and summarizing deposition transcripts. Over time, AI will assume even more of the work that is linear, scaffolded, and rule-based—the very competencies that formative assessment methods prioritize. It is precisely because machines can handle these routine functions that human lawyers will be called upon for the tasks machines cannot perform: navigating ambiguity, exercising judgment, and making moral and strategic decisions in uncertain and high-stakes environments. These are not skills that can be easily measured in frequent, incremental assessments. They are proficiencies cultivated through grappling with open-ended questions, thinking independently under pressure, and tolerating prolonged uncertainty—all hallmarks of the traditional law school model.

The Socratic method, with its relentless questioning, does not reward checklist thinking. It forces students to hold competing ideas in tension, to adapt in real time, and to articulate reasoning under duress. A single, weighty final exam does not merely test memorization; it demands synthesis, prioritization, and judgment under conditions of limited time and imperfect information. It mirrors, in structure and spirit, the real challenges that lawyers will increasingly face as AI expands the boundaries of the possible and leaves humans to wrestle with what remains truly difficult.

Some proponents argue that clear learning outcomes promote transparency and focus. While transparency is valuable, reducing complex intellectual development to pre-articulated outcomes risks oversimplification. Law is not a field where every important competency can be predefined and measured. Lawyers must develop habits of mind that transcend any given rubric—the ability to reason through novel problems, challenge assumptions, and invent legal arguments in circumstances that defy easy categorization.

To be sure, traditional teaching methods are not always perfect, and there is a place for some amount of formative feedback and academic support to help struggling students find their footing. But in our zeal to modernize, we risk forgetting that education is not simply about skill acquisition. It is about formation: shaping the kind of minds that can thrive in the chaos AI will leave behind, not just ones capable of surviving it.

Legal educators should resist the temptation to domesticate law school into a series of neat exercises culminating in predictable outcomes. Instead, we should embrace the old ways, not out of nostalgia, but out of foresight. In the age of artificial intelligence, the lawyer who can think independently, reason under pressure, and make judgments amid uncertainty will be the lawyer who endures—and whose work will continue to push the law forward.

The kind of lawyer I’m describing will be forged—not by constant checkpoints and assurances—but by the difficult, uncertain, and transformative experiences that the traditional law school model, at its best, is uniquely capable of providing.

Share

Subscribe now

In one of my favorite anime films, Ghost in the Shell (1995), the boundary between human and machine is razor thin. The protagonist, Major Motoko Kusanagi, is a cybernetic counter-terrorism agent who often wonders whether her consciousness—her “ghost”—is still her own or just another construct flowing through the digital networks plugged into the human brain inside her mechanically augmented body (or “shell”). In her world, sentient machines are not just tools, they act, decide, and sometimes reflect. And as the line between operator and operation blurs, the nature of authority, responsibility, and agency begins to shift.

With these themes in mind, I have been thinking a lot lately about what it means for AI to augment an attorney’s day-to-day life—not because we are heading for a dystopia of fully sentient machines, but because legal practice is already beginning to resemble that animated future. We are quickly seeing the emergence of autonomous systems that not only execute commands but pursue goals, adapt strategies, and interact across a web of institutional actors. In this not-so-distant future, corporate lawyers will not just use AI, they will collaborate with it, negotiate through it, and be challenged by it.

So, what will a day in the life of a corporate lawyer look like two to three years from now, when the AI-augmented legal “machine” has a ghost of its own?

I suspect the typical workday will be begin before our future attorney even wakes up. Overnight, the lawyer’s AI assistant will operate continuously in the background. It will be doing things like checking dockets, scanning news alerts, assessing stock price movements, performing document reviews, and drafting initial responses to messages that arrived after the lawyer fell asleep.

For instance, if a client received some negative late-night social media attention, the lawyer’s AI will not just notify her about this development in the morning, it will critically examine what happened and recommend what guidance she might provide to the client. The lawyer will presumably glance at the AI’s analysis and initial draft, make a few edits, and then hit “send”—reducing what might have been thirty minutes of work into something now done in five.

By the time our lawyer attends her first meeting, her firm’s enterprise AI might have already recommended some major reorganizational moves, including changing team assignments to shift personnel to high-urgency matters, as well as enrolling some junior associates in a custom learning module tailored to a new engagement or practice area.

If our lawyer then needs to prepare for a settlement conference, she can use her AI assistant to develop a predictive model based on thousands of past cases and the unique traits or tendencies of her opposing counsel. The AI will map likely outcomes, develop settlement ranges, and suggest tone and language shifts expected to correlate to better results. It will further simulate offers and counteroffers, perhaps going so far as to flag moments when it would be wise to pause for effect, concede a point, or hold firm. Ultimately, this type of prep work will no longer be sharpened principally by instinct, research, or personal experience; it will mirror sophisticated, data-driven and military-grade war games not unlike what governments might use for national defense.

Next, imagine our lawyer needs to advise a board about litigation exposure related to a recent acquisition. In the past, she and her team probably would have presented a slide deck and an executive summary, followed by some Q&A. Very soon, however, the same lawyer will walk into a room where everyone—directors, executive officers, and general counsel—will be using their own AI tools in real time. The lawyer’s litigation dashboard, powered by analytics and visualized risk metrics, will be projected as she speaks—and as she narrates the facts and offers her opinions, the other AI-augmented players in the room will be assessing and challenging them. One director’s assistant might note an inconsistency and flag a sentence from a prior SEC filing. Another could propose a counter-scenario that recommends adopting a different indemnity strategy in future deals. In essence, throughout the meeting, the room will not pause for someone to generate a memo or wait for word from an outside consultant; everyone will be evaluating and adapting to the situation on the fly in the same amount of time it takes our lawyer to click to her next slide.

In these moments, future lawyers will realize they no longer just advise clients, they work alongside them. Boardrooms will cease to be static, deliberative spaces. They will become dynamic, data-rich live conversations where lawyers must collaborate, improvise, and respond in the moment.

Of course, making specific predictions about what the job of the corporate attorney will look like in a few years is risky business. Some might worry—and not without reason—that AI could eventually perform most traditional lawyerly tasks. But that is not what I see happening. What I expect to see change is best described as a transfer of responsibility. Repetitive labor will recede while cognitive labor accelerates. The attorneys who thrive in this coming environment will not be the ones who are the most experienced or even the most intelligent. They will be the ones most healthily integrated with AI—the professionals who can move seamlessly between human intuition and machine intelligence while never forgetting that the soul of legal practice is grounded in values and principles. Put another way, as AI takes on more of the workload, the lawyer’s core duties will endure: to guide, to reason, and to discern.

In the end, just like the characters in Ghost in the Shell, the boundary between human and machine will not vanish entirely in the law practice of the future. It will simply represent a shifting line, one that lawyers must learn to navigate with purpose. In this AI-augmented world, what will matter most is not how quickly lawyers can generate an answer, but how wisely they can interrogate it. Preserving the “ghost” in the loop will be necessary to ensure that amidst the speed and scale of machine logic, legal practice remains a fundamentally human profession.

Subscribe now

Share

Sometimes the most surprising thing about AI isn’t how smart it is—it’s how thoughtful it can be.

One of my favorite uses for AI tools like ChatGPT and Claude is asking them to help me think through messy situations, including the kind of ambiguous or emotionally charged moments that show up at work. This kind of application is incredibly accessible and valuable, especially for people who are still figuring out how to make AI part of their everyday lives.

To illustrate what I mean, I gave two hypothetical workplace dilemmas—one serious, one trivial—to ChatGPT 4o and Claude 3.7 Sonnet and asked for their advice. I wasn’t only interested in what they would say, I wanted to see how they would say it. Would their advice be empathetic? Would they appreciate the nuance of the situation? Would they act like the kind of colleague we often wish we had?

Here’s what I found.

Scenario 1: Breach of Trust

My first scenario was serious—a stark ethical dilemma based on a major plot point in The Godfather, Part II—that I reimagined as a workplace drama:

I’m a first-year associate at a boutique consulting firm. After nine months at the firm, I was given my first major responsibility: helping prepare a confidential proposal for a high-stakes client, Acme Energy. I was on a small team led by a senior partner. One of my teammates was Ben, a third-year associate who often complains about being passed over while newer hires like me get more attention.

A week before our proposal was due, Acme abruptly pulled out, saying it received a similar proposal from a rival firm. At a happy hour, a colleague mentioned that she had heard Ben vent about our strategy to a friend at that rival firm. I confronted Ben. He admitted to the conversation but said it was not intentional and urged me not to escalate, warning that doing so could damage both our careers.

I was stunned. This is my first real brush with workplace conflict. What should I do?

ChatGPT’s Response

ChatGPT took the tone of a calm, experienced mentor. It opened by reframing the issue:

“You're facing a classic high-stakes ethical and professional dilemma—one that tests your judgment, integrity, and emotional resilience early in your career.”

Then it offered a structured seven-step framework:

• Clarify your values and long-term goals: What kind of professional do you want to be? What kind of firm do you want to work for?

• Document what you know: Dates, facts, conversations—keep a private, factual record.

• Assess the firm’s risk: Did Ben’s disclosure meaningfully compromise the proposal?

• Consider internal dynamics: Is there someone you trust in leadership who could offer confidential guidance?

• Decide whether to report—and if so, frame it as loyalty to the firm, not a personal grievance.

• Prepare for fallout, but don’t let fear dictate your action.

• Learn from the experience: This is how real professional judgment is formed.

What stood out to me here was not just the content, but the tone: grounded and mature. ChatGPT gave me a way to think things through, reassuring me that this type of workplace problem is a step toward professional growth.

Claude’s Response:

Claude’s voice was a bit warmer, and its first step was to validate the emotional stakes: “This is significant for a first major conflict at work.”

It then walked me through similar categories for consideration:

• Assess the seriousness of the breach.

• Weigh Ben’s intent—but don’t excuse it.

• Think about organizational culture: How does your firm usually handle issues like this?

• Consult a trusted mentor.

• Consider reporting, ideally by presenting the facts objectively with your supervisor.

Claude’s closing note was memorable: “The pressure Ben is putting on you is unfair—you didn’t create this situation. Your primary responsibility is to your firm and its clients.”

Both tools emphasized values, process, and personal responsibility. Neither tried to tell me exactly what to do next. That restraint, especially when paired with clarity, felt most welcome.

Scenario 2: The Missing Sandwich

After something so serious, I wanted to test how the same tools would respond to something much more mundane and petty:

Someone has been stealing my lunch from the communal refrigerator. What should I do?

ChatGPT’s Response

ChatGPT responded like an office diplomat:

1. Don’t assume bad intent—yet. Mistakes happen.

2. Try a light touch: A friendly-but-firm note on the fridge, e.g.,

“To whoever keeps eating my lunch: I admire your taste, but next time, let’s do lunch together. 😊”

3. Label boldly: Names, dates, sticky notes. Maybe even a mildly cheeky warning.

4. Escalate if needed: Consider a desk cooler or a quiet word with HR.

5. Maintain your composure: This is annoying, but you’re not wrong for caring.

The response was clear-eyed without overreacting.

Claude’s Response:

Claude took a practical tone:

• Label your food clearly.

• Use distinctive containers.

• Mention it in a team meeting or email (without finger-pointing).

• Talk to HR or your office manager.

• Use a locking lunchbox if necessary.

• Keep backup snacks at your desk.

But here’s my favorite part of Claude’s response: “This might seem minor compared to the confidentiality breach with Ben, but workplace respect matters at all levels. How we treat each other in small ways often reflects broader cultural issues at work.”

That one line captured what I was trying to test: Can AI get close to grasping the human texture and feel of workplace life? In this case, it seems to have done so pretty well given the basic facts I provided.

What AI Really Offered: Clarity, Not Control

To recap, in both scenarios, neither AI tool tried to give me the answer. Instead, they did what a good friend or mentor does: they helped me clarify what I know, what matters to me, and what options I might consider next. They slowed the moment down. They helped me see not just what might be “right,” but what kind of professional I want to be. They reminded me to think about values, relationships, and consequences.

The more I experiment with AI, the more I see opportunities for using it as a steady, judgment-free companion in moments that feel fraught or unclear. Used this way, AI doesn’t replace wisdom. It simply helps us make space for our own.

Subscribe now

When I observe law faculty talking about AI, the discussion often turns into a version of the same basic fear. If we allow students to rely on AI, won’t that undercut the whole point of teaching them how to think like a lawyer?

This is a fair question, but it risks missing the bigger picture: unless law schools evolve—and quickly—they will be training students for a version of professional practice bound for extinction.

We are living through a once-in-a-generation reordering of the entire professional services landscape. A recent report from Ethan Batraski of Venrock describes this shift starkly: the $20 trillion global market for services like law and accounting has been dominated for decades by human-centric, prestige-based firms built on artificial scarcity and expensive labor. But that model is now under threat from AI-native firms that are not just making existing services more efficient—they are redefining what it means to deliver expertise.

The professional services firms of the future won’t just automate document review or churn out well-edited memos. As Batraski notes, they will build systems that can perform strategic analysis, high-stakes decision modeling, and creative problem-solving at scale. The value proposition they present isn’t built on who they are—it’s built on what they can provide: better outcomes and much faster delivery. In short, we aren’t talking about law firms with software. We’re talking about software companies that deliver legal services.

So, if law faculty ask whether AI is good or bad for the profession, they’re asking the wrong question. The better question is: what will legal practice look like in 3–4 years, and how do we prepare our students to add value in that world?

I don’t expect that world to center on billing time. It will not reward generic memos (if memos are even still a thing, which I highly doubt). It will not tolerate inefficiency masked as rigor. It will not rely on reputation as a proxy for performance. Instead, it will demand that new lawyers know how to use AI tools responsibly, strategically, and effectively. It will require that they collaborate across disciplines to build and improve AI-driven service models. It will mean evaluating the output of AI tools and identifying where human input is essential.

Importantly, this doesn’t mean abandoning the fundamentals. It just means teaching them in new ways. Students trained to use AI to prototype arguments, compare case patterns, or stress-test a litigation strategy are not losing their capacity for legal reasoning. They are enhancing it. They are learning tactical judgment—not just how to do the work, but how to direct the work that AI can now do.

In other words, what it means to “think like a lawyer” is changing. Legal professionalism today must include AI fluency—not as a gimmick or crutch, but as a form of cognitive and ethical literacy. It means understanding how to partner with technology in ways that preserve client interests, safeguard confidentiality, and deliver superior outcomes. The lawyer of the future is not just a legal expert, but an architect of AI-augmented systems.

Law schools thus face a choice. They can retreat into a bunker of tradition and train students for jobs that will no longer exist, or they can become laboratories for a new kind of professional—a lawyer who blends deep legal knowledge with technical competence and adaptive problem-solving.

There are already signs of this future taking shape. New courses are emerging that treat AI as infrastructure, not novelty. Ethics seminars are grappling with the implications of machine-generated legal advice. And students themselves are leading the charge, demanding training in the tools their future employers are already adopting. But progress is uneven. Faculty skepticism about AI’s impact on learning and pedagogy is real. So is the discomfort of teaching in a domain where technology is evolving faster than casebooks.

Still, I believe the stakes are far too high to do nothing. As the AI-forward CEO of Shopify recently put it, the burden of proof is shifting. Soon, managers won’t get to hire a human unless they first prove AI can’t do the job. If this view captures where the market is headed, then our graduates need to be ready.

We shouldn’t see AI as a threat to the soul of law or to legal education. Rather, it is an opportunity to reimagine what excellent legal practice can look like. The law schools that seize the new reality will be the ones that win—and so will their students.

Subscribe now

This might seem like a small distinction, but it gets at the heart of how institutions—whether universities, businesses, or nonprofits—should (or shouldn’t) be using artificial intelligence. Generative tools like ChatGPT or Claude can produce language, suggest decisions, and identify patterns faster than any human analyst. But they don’t know what your institution is for. They don’t understand the broader story you’re trying to tell, or the values that drive your mission.

Human leadership still matters. It's not about ignoring AI or rejecting its contributions. It's about marshaling what AI produces and refining it in line with a human vision for the final product. Just as a director uses what the crew captures—a powerful shot, a surprising moment, an aesthetic flourish—and shapes it into a coherent narrative, institutions must take responsibility for the meaning, intent, and impact of the decisions they make using AI.

This is also why good directors don’t simply defer to the actors. Actors bring invaluable perspective and creativity to a role—they might even surprise with an interpretation that wasn’t anticipated. But it’s the director’s job to assess their performances in context and decide whether they serve the story. Sometimes an actor's take is perfect. Other times, it needs to be nudged or reimagined. The same goes for institutional use of AI: the outputs may be compelling or unexpected, but the responsibility rests in how they’re shaped and framed.

We often talk about AI as an “assistant” or a “collaborator,” but in institutional contexts, we need to be even clearer: AI is a tool, not a partner or a substituted mind. A good tool can extend your reach, speed up your work, and identify new possibilities. However, it doesn’t set priorities. It doesn't carry ethical responsibility. It doesn't answer to shareholders, faculty colleagues, or the public.

Consider, for example, a law student using AI to help brainstorm, outline, or revise a brief. These tools can suggest structure, play devil’s advocate, or refine phrasing. But they don’t understand the purpose of the assignment, the student’s unique argument, or the disciplinary norms of the course. That’s directorial work—deciding what to keep, what to discard, and what to revise in service to a larger intellectual goal.

I've seen this dynamic play out in my own work. When developing a new course proposal for my class on "AI, Law, and Business," I used ChatGPT as a generative tool. I uploaded a set of articles and news stories, and then I engaged in an iterative process—refining structure, surfacing themes, and updating my learning outcomes. But throughout, I remained the director. The AI offered helpful raw material, but the course reflects my own academic goals and pedagogical judgment.

AI is a stunningly capable tool. Yet the responsibility for the final cut still rests with us.

Share

The Paper Clip Problem takes its name from a well-known thought experiment in AI ethics, proposed by philosopher Nick Bostrom. In it, a superintelligent AI is programmed with a seemingly harmless goal—making paper clips. But without proper constraints, it might pursue that objective with such single-minded intensity that it consumes all available resources, including human life, to achieve it.

In this newsletter, I use the “paper clip problem” as a jumping-off point to explore how artificial intelligence—when embedded in institutions like businesses and universities—can amplify both promise and peril. Drawing on my work as a law professor specializing in corporate governance and higher education, I examine the legal, organizational, and policy decisions that shape how AI is developed, deployed, and resisted. My aim is to shed light on how small, often technical choices can scale into major societal shifts—and to help readers navigate those shifts with clarity, skepticism, and curiosity.

I’m not an AI engineer. But I’ve spent my career thinking about how rules, incentives, and institutional design influence power and decision-making. That’s the lens I bring to these conversations. Whether you're trying to adopt AI in your workplace, regulate it responsibly, or simply understand it better, I hope this space becomes a useful, thoughtful companion.

Subscribe now

(Oh, and all the images on this Substack were created using ChatGPT 4o. I plan to add more every so often by using the same prompts I started with, just to see how they evolve. For the image above, I asked for a “painting of the paper clip problem in the style of Caravaggio.”)