There is a comforting way for lawyers to talk about artificial intelligence.
It is not intelligent. It does not think. It does not create. It consumes enormous quantities of human language, identifies patterns in what has already been said, and rearranges those patterns into a statistically plausible response. It is, in one particularly durable phrase, a “stochastic parrot.”1
Fine.
Lawyers should be extremely careful with this argument. It can sound less like an indictment of artificial intelligence than a job description for the legal profession.
Lawyers are trained by reading what other lawyers have written. We learn rules extracted from old disputes, then rearrange those rules around new ones. We borrow the language of judges, apply it to a client’s facts, and call the result a brief. We search a massive body of inherited material for the combination of words most likely to move a particular audience toward a desired conclusion. Sometimes, we even predict what a judge is likely to say next.
The comparison is not literal. Large language models generate text by estimating the probability of tokens in relation to other tokens; lawyers do not.2 A transformer and a human brain are not the same just because both can finish a sentence. But, functionally, the resemblance is difficult to ignore.
Both lawyers and generative AI take existing language, identify useful patterns within it, and repurpose those patterns for the task at hand. We call one of those things artificial intelligence and the other legal reasoning. That reality does not make lawyers useless. It does make our criticism of AI strangely useless.
The law works because the past can be repurposed for the present: words written for one controversy can acquire new force in another; an analogy can reveal that a principle extends farther than anyone previously understood; a rule designed for yesterday can be made to answer tomorrow.
Lawyers do not create from nothing. Neither do AI models. Nobody does.3 That does not erode the potential we—or they—have.
The Originality Trap
Originality carries a ridiculous amount of moral weight in the debate over generative AI. The criticism usually moves through three propositions without acknowledging the gaps between them. First, the model generates from patterns learned through existing information. Second, because the output is derived, it is not genuinely original. Third, because it is not original, it is fraudulent, empty, or undeserving of respect.
Only the first proposition follows from how the technology works.
The rest depend on a theory of originality that would disqualify most of human culture.
Human creativity begins in debt. Writers inherit vocabulary, genre, structure, and reference. Musicians work inside traditions they did not invent. Political movements revive old language for new conditions. Even rebellion depends on knowing what it is rebelling against.
Margaret Boden’s influential account of creativity expressly includes producing novel combinations from familiar ideas, alongside exploring and transforming existing conceptual spaces.4 Creation does not require virgin birth. It can be combination. It can be rediscovery. It can be taking two things everyone recognizes and placing them together in a way nobody expected.
That is not fake creativity. That is most creativity.
Copyright law, although concerned with a narrower legal question, contains a similar intuition. In Feist Publications, Inc. v. Rural Telephone Service Co., the Supreme Court held that originality requires independent creation and only a minimal degree of creativity—not novelty.5 A work does not stop being original merely because it resembles what came before it. “Original” and “without antecedent” are different ideas.
Current copyright doctrine separately requires human authorship, and the Copyright Office has concluded that purely machine-determined expression is not protectable under existing law.6 That may be a reasonable way to allocate legal ownership. It does not resolve the philosophical question of whether an output can be useful, surprising, or creative. Copyright tells us who may own something. It does not tell us everything the thing may be.
None of this erases the harder questions surrounding AI. Training practices raise legitimate concerns about consent, compensation, bias, ownership, and power.7 Those questions matter, but “it learned from existing information” is not a substitute for answering them.
Nor is it a verdict against the output.
If borrowed ingredients spoiled the meal, culture would starve. We do not usually judge human work by tracing every ingredient backward until nothing remains. We judge what the ingredients became.
Lawyers, of all people, should understand that.
Lawyers Are Prediction Machines
American legal education has long been tempted by the idea that law is a system whose correct answers can be extracted through the right technique.
Christopher Columbus Langdell treated law as a science and appellate cases as the specimens from which its governing principles could be induced. Thomas Grey later described the resulting classical orthodoxy as an aspiration toward a complete, formally ordered system: general principles at the top, predictable results at the bottom, legal reasoning as the clean line connecting them.8
Even as confidence in that system has faded, the training method persists.
Law students are still taught to read decisions, extract rules, synthesize authorities, analogize, distinguish, and apply.9 We spend years learning to sound like the cases we have read without sounding like we copied them. Here are two hundred years of people talking. Now say something new—but not too new, and preferably with citations.
Edward Levi described legal reasoning as a process in which judges and lawyers identify similarities between cases, announce a rule implicit in those similarities, and apply the rule to a new dispute.10 But similarities do not select themselves. Whether two cases are alike depends on which facts are treated as legally significant. Change the similarity and the rule changes with it.11
Lawyers do not merely recover law from precedent. We decide which parts of the precedent deserve to survive.
Holmes put the point more bluntly: “The life of the law has not been logic: it has been experience.”12 Felt necessities, public policy, moral judgment, and the accumulated intuitions of judges have always influenced the apparent certainty of doctrine. Logic may organize the explanation, but it does not necessarily make the choice.
Generative AI is unnerving to lawyers because it is becoming capable of performing many of the activities we have spent a very long time making look mystical.
The ABA now recognizes that generative systems can assist with legal research, document review, due diligence, regulatory compliance, contract analysis, and drafting.13 LegalBench, an interdisciplinary benchmark created by legal professionals and computer scientists, tests language models across 162 tasks involving several forms of legal reasoning.14 The existence of such a benchmark does not prove that machines understand law in the richest human sense, but it does prove that a meaningful amount of what we call legal reasoning can be divided into tasks that machines can attempt and researchers can measure.
The early results are not subtle. One randomized controlled trial found that access to GPT-4 produced large and consistent increases in the speed with which law students completed realistic legal assignments, though it produced only slight and inconsistent improvements in quality.15 A newer experiment involving a legal retrieval system and a reasoning model found improvements in both productivity and the quality of work across several tested assignments.16 Those studies have limits. They tested particular systems, particular tasks, and law students rather than every kind of lawyer. AI also remains perfectly capable of producing authoritative-sounding garbage.17
But the direction is clear enough. The technical production of legal work is no longer exclusively human.
That is the real professional crisis. AI does not need to become a lawyer to embarrass the lawyer’s theory of his own worth. It only needs to become competent at the work lawyers have treated as the source of that worth.
If our professional identity rests on remembering more cases, producing a first draft faster, or reciting a rule with greater polish, we have built a machine-shaped profession. We should not be shocked when a machine fits inside it.
Law Is Policy Wearing a Robe
The technical conception of law survives by treating policy as something that happens somewhere else.
Legislators make policy. Lawyers apply law. Values enter at the beginning, when a rule is enacted, and then politely disappear while trained professionals administer it.
That is a beautiful system.
That is not our system.
Policy does not disappear when a statute is passed. It hides inside application.
Every legal field depends upon words like reasonable, material, substantial, undue, fair, and necessary. Those terms do not fail because they require judgment. Requiring judgment is their function. They allow general rules to govern circumstances their authors could not fully anticipate.18
They also transfer power to the people who decide what the terms mean.
What risk is reasonable? Which burden is undue? What resemblance is substantial? Whose mistake is material? How much process is due?
There is no purely technical answer. There are better and worse answers. There are faithful and dishonest answers. There are answers more defensible in text, precedent, purpose, or principle. Regardless, the authorities do not descend from Westlaw and select the winner themselves.
Legal Process scholars responded to the Realists by acknowledging this discretion while insisting that it could be disciplined through institutional role, public justification, and “reasoned elaboration.”19 A judge should not disguise personal preference as deduction. But neither should the judge pretend that text mechanically resolves every controversy. The decision must connect text, purpose, principle, institutional competence, and consequence in a way that can be stated publicly and challenged by others.
Administrative law makes the obligation unusually obvious. Under Motor Vehicle Manufacturers Ass’n v. State Farm Mutual Automobile Insurance Co., an agency must examine relevant information, consider important aspects of the problem, and articulate a rational connection between the facts it found and the choice it made.20
The choice is still a choice. Expertise does not eliminate it. Procedure makes it governable. Explanation makes it visible. Judicial review forces the government to admit that a person or institution decided something, for reasons, rather than allowing policy to masquerade as the inevitable output of a rule.
Owen Fiss understood adjudication in similar terms. Constitutional values like liberty, equality, and due process are not self-defining. They are ambiguous. They conflict. Institutions must give them operational meaning.21 A court does not merely discover a complete constitutional command waiting patiently inside the text. It gives public values concrete form.
Technical accounts of law obscure this process. Doctrine can make a political settlement look natural, a contested classification look inevitable, and a human decision look like the impersonal operation of logic. Critical legal scholars have described this as the “freezing” of legal reality: contingent arrangements become embedded in neutral-sounding categories that conceal both their history and their alternatives.22 Policy, in this sense, does not mean partisan whim. It means deciding which purposes matter, whose experiences count, which consequences are acceptable, and which institutions should possess the authority to choose.
Those questions appear in legislation. They appear in regulation. They appear in constitutional adjudication. They appear whenever a lawyer decides which fact should become the center of a case and which should disappear into the background.
The lawyer who pretends otherwise is not avoiding policy.
He is practicing it invisibly.23
Who Gets to Choose?
Once we admit that legal interpretation is also policymaking, the relevant question is not whether lawyers or machines are capable of making those choices. The question is who gave them the right to choose.
“We already do it” is not an answer for lawyers.
Neither is “we went to law school.”
Lawyers do not acquire sovereignty by passing the bar. Artificial intelligence will not acquire it by passing us. The authority to shape law belongs neither to the people who understand its machinery nor to the machines that may eventually understand it better.
That power belongs to the people who must live beneath it. The lawyer’s role is derivative of theirs.
A statute speaks in categories. A client arrives with a story. The lawyer translates between them: identifying which parts of a life the legal system can recognize, which injuries its categories have concealed, and how a human demand can become something an institution has the authority to remedy.
That translation moves in both directions. Lawyers make people legible to institutions, but they also make institutions answerable to people. We explain what the law permits, expose what it has chosen not to see, and give those subject to public power a language through which to contest it.
The Model Rules describe the lawyer simultaneously as a client representative, an officer of the legal system, and a public citizen responsible for the quality of justice.24 That language is grandiose. Lawyers enjoy grandiose language. It nevertheless identifies something real.
Mari Matsuda’s account of “multiple consciousness” makes the role more demanding: the lawyer may need to understand law at once as an instrument of domination and as an indispensable source of protection.25 That movement is not hypocrisy—it is the work.
This position can cultivate judgment useful to policymaking, so long as we view that judgment not as an anointment, but as a tool.
Artificial intelligence can help use that tool. Moral aspiration is not self-executing. “Be fair” must become rules about coverage, conduct, proof, remedies, and institutional authority.
Someone has to turn “do better” into a rule that can actually make someone do better.
AI can help draft those rules, identify contradictions, predict their consequences, and make legal knowledge available to people whom the cost of law has traditionally excluded.26
Good.
None of those capabilities determines what fairness requires. More importantly, none determines who gets to decide.
The person living with the consequence must retain the authority to say what the injury means, what remedy is wanted, and which tradeoffs are acceptable. A lawyer is valuable when legal skill carries that judgment into an institution capable of acting upon it. A lawyer is dangerous when legal skill replaces it.
Legal skill is power, and power does not arrive with its own moral instructions. With great power comes great responsibility.27
Derrick Bell demonstrated what happens when lawyers forget where that responsibility comes from. In school-desegregation litigation, civil-rights lawyers committed to integration sometimes allowed their own ideological objectives to displace the priorities of Black parents seeking safe, adequately funded schools capable of educating their children.28 Their failure was not that they wanted to change society. These lawyers failed when they began treating their clients as raw material for that change.
That is more than an example of lawyers behaving badly. It marks the boundary beyond which a lawyer’s claim to authority loses its justification. Once the lawyer’s expertise becomes a substitute for the judgment of the people supposedly being represented, the lawyer begins to resemble any other opaque decisionmaking system: powerful, technically sophisticated, and insulated from the people forced to live with its conclusions.
Accountability matters because it can preserve that relationship—not because possessing a license makes its holder good.
Consider Mata v. Avianca. ChatGPT generated nonexistent authorities, which lawyers submitted to a federal court and continued defending after the citations were questioned.29 The machine invented the cases. The humans filed them anyway.
That story is usually offered as proof that AI cannot be trusted. It is at least equally strong evidence that human presence does not guarantee human judgment. A lawyer can be sanctioned, sued, fired, or disbarred, but knowing whom to punish after a failure does not, by itself, make the failure less likely.30 A name, conscience, and disciplinary board are not substitutes for a system in which decisions can be examined and challenged.
Accountability does not require a soul. It requires a framework.31
That framework can assign responsibility to the lawyer who relies upon an output, the firm that deploys a system, the vendor that designs it, and the institution that places it in a position to affect legal rights. The point is not that the machine itself must experience shame. The point is that power must remain traceable to people and institutions that can be confronted, corrected, and required to answer for its use.32 The same principle explains why competence is not enough. Artificial intelligence may eventually follow rules more faithfully than lawyers, identify consequences we miss, and produce something functionally indistinguishable from moral reasoning. That would make it an extraordinary instrument of governance.
It would not give it the right to govern.
Law already distinguishes expertise from authority. An agency does not become entitled to regulate merely because it understands the subject. Its power must be delegated, exercised through lawful procedures, publicly explained, and exposed to participation and challenge.33 People are entitled to speak into that process not because every person is an expert, but because they are governed.
The human role cannot therefore be reduced to ceremonial approval. A person blindly clicking “approve” on a decision produced somewhere else offers theater where judgment is needed.
Meaningful human involvement means preserving the ability of affected people to authorize, contest, revise, and resist the rules imposed upon them. Lawyers can help create that possibility. So can artificial intelligence. Either can also destroy it. The difference that matters is relational, not biological.
Law does not need lawyers because lawyers are uniquely intelligent, ethical, accountable, or empathetic. It needs lawyers when they keep legal power connected to the people in whose name it operates.
Law does not need humans because machines can never appear to care. Law needs humans because the law belongs to the people who have to live under it.34
As students of the law, our defensible policy role lies not in controlling every stage of legal reasoning, but in helping human judgment acquire institutional force. If we have any value at all, it is because of our humanity, not despite it.
Judgment
Generative AI may be the best thing to happen to the legal profession’s understanding of itself.
It is forcing a separation between legal technique and legal judgment that lawyers should have made long ago.
Technique retrieves the rule, synthesizes the cases, drafts the provision, predicts the objection, and formats the citation.
Judgment asks whether the rule deserves to survive.
Technique tells us what courts have said.
Judgment decides whether it should be said again.
This does not make technical competence unimportant. A lawyer cannot transform a system she does not understand. Bad research can destroy a good cause. Inaccurate citations, careless analysis, and imprecise drafting harm real people.
But technique is a means. When lawyers treat it as the source of their dignity, they confuse mastery of an instrument with justification for the music.
Artificial intelligence can take over some technical work, accelerate more of it, and make still more available to people who could never afford the traditional price. That should be welcomed. The purpose of legal work is not to preserve the amount of human labor once required to perform it.
Let the machine do more.
Then ask more of the lawyer.
If AI can help tell us what the law is, lawyers should spend more time asking what the law is doing. If AI can draft the ordinary argument, lawyers should become better at finding the human fact that makes the ordinary argument inadequate. If AI can follow our rules, lawyers should become more willing to decide whether those rules deserve to be followed.
Recognizing the value of humans in law is not a case against artificial intelligence. The rise of artificial intelligence does, however, make a strong case against mistaking generation for judgment—whether the generator is a machine or a lawyer.
Neither deserves praise merely for reproducing what has already been said. Both should be judged by what their acts of reproduction make possible. Both can entrench hierarchy, launder error, and give old prejudice a new voice. Both can uncover connections, widen access, clarify choices, and place inherited language in service of a better cause.
The machine is not coming for our humanity.
It is coming for the part of lawyering we mistook for humanity.
Let it.
Let machines remember. Let them retrieve. Let them synthesize. Let them draft. Let them show us everything we have said before and every plausible way we might say it again.
Then decide whether it deserves to be said.
Decide whose interests it serves. Decide which consequences we can accept. Decide what must change. Decide where we are going.
Technology may give us the best directions.
It does not get to choose the destination.
That is judgment.
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major & Shmargaret Shmitchell, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency 610, 616–18 (Ass’n for Computing Mach. 2021).
See Ashish Vaswani et al., Attention Is All You Need, in 30 Advances in Neural Information Processing Systems 5998, 5998–6001 (Curran Assocs., Inc. 2017); Chloe Autio et al., Nat’l Inst. of Standards & Tech., U.S. Dep’t of Com., NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile 4–5 (2024). The analogy developed here concerns the transformation of inherited language for new tasks, not an assertion that human cognition and transformer inference are biologically or computationally identical.
See Naomi Mezey, The (Still) Unexplored Possibilities of a Poetics of Law, 35 Yale J.L. & Human. 321, 321–25 (2024), https://scholarship.law.georgetown.edu/facpub/2646/ (considering judicial opinions as acts of imagination and meaning-making that draw upon inherited forms, conventions, and expectations).
Margaret A. Boden, Creativity and Artificial Intelligence, 103 Artificial Intelligence 347, 347–49 (1998).
Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 345–51 (1991).
U.S. Copyright Off., Copyright and Artificial Intelligence, Part 2: Copyrightability 21–40 (2025).
See Emily M. Bender et al., supra note 1, at 613–18 (discussing the reproduction of social bias and hegemonic viewpoints through large training datasets); Chloe Autio et al., supra note 2, at 4–5 (identifying harmful bias, privacy, and information-integrity risks associated with generative artificial intelligence); U.S. Copyright Off., Copyright and Artificial Intelligence, Part 3: Generative AI Training 85–107 (Pre-Publication Version 2025), https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf (discussing licensing, compensation, and control over copyrighted works used in generative-AI training).
Thomas C. Grey, Langdell’s Orthodoxy, 45 U. Pitt. L. Rev. 1, 5–12 (1983).
See, e.g., First Year Legal Practice Program, Georgetown Law, https://www.law.georgetown.edu/academics/courses-areas-study/legal-writing-and-student-scholarship/first-year-research-and-writing-program/ (last visited Aug. 22, 2026) (describing Georgetown’s first-year instruction in legal research, analysis, drafting, revision, advocacy, and the making of reasoned professional choices).
Edward H. Levi, An Introduction to Legal Reasoning, 15 U. Chi. L. Rev. 501, 501–02 (1948).
For Georgetown Law’s 2026 Write On Competition, I wrote a case comment on Beasley v. O’Reilly Auto Parts, 69 F.4th 744 (11th Cir. 2023), which involved whether an ADA failure-to-accommodate claim requires the requested accommodation to enable performance of an essential job function. My focus, however, was on what the panel did to precedent. The majority treated the Eleventh Circuit’s earlier essential-functions rule in Lucas v. W.W. Grainger, Inc., 257 F.3d 1249 (11th Cir. 2001), as dicta; Judge Luck’s concurrence argued that the rule supplied the sole ground for rejecting one of the requested accommodations in Lucas and therefore constituted a holding. Compare Beasley, 69 F.4th at 756–60, with id. at 761–62 (Luck, J., concurring). I argued that this seemingly technical reclassification consigned Lucas to doctrinal limbo—neither binding nor formally overruled. Ben Rinker, Cast Doubt, Decide Nothing: Beasley v. O’Reilly Auto Parts and the Demotion of Holding to Dicta 4–7 (2026) (unpublished manuscript) (on file with author). Even in practice, which facts and rationales we classify as decisive can determine not merely how a precedent is described, but what future courts are legally permitted to do.
Oliver Wendell Holmes, Jr., The Common Law 1 (1881).
ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 1–3 (2024).
Neel Guha et al., LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models, in 36 Advances in Neural Information Processing Systems 44,123, 44,123–24 (Curran Assocs., Inc. 2023).
Jonathan H. Choi, Amy B. Monahan & Daniel Schwarcz, Lawyering in the Age of Artificial Intelligence, 109 Minn. L. Rev. 147, 176–92 (2024).
Daniel Schwarcz, Sam Manning, J.J. Prescott, Patrick Barry, David R. Cleveland & Beverly Rich, AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice, 3 J.L. & Empirical Analysis 221, 222–24 (2026).
See Varun Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216, 216–18, 224–26 (2025), https://doi.org/10.1111/jels.12413 (reporting that three commercial retrieval-augmented legal-research systems produced hallucinated responses to between seventeen and thirty-three percent of tested queries).
See Louis Kaplow, Rules Versus Standards: An Economic Analysis, 42 Duke L.J. 557, 559–60 (1992), https://scholarship.law.duke.edu/dlj/vol42/iss3/2/ (distinguishing rules, whose content is specified before regulated conduct occurs, from standards, whose content must be supplied through application after the fact).
See Henry M. Hart, Jr. & Albert M. Sacks, The Legal Process: Basic Problems in the Making and Application of Law 4, 143–52 (William N. Eskridge, Jr. & Philip P. Frickey eds., 1994).
Motor Vehicle Mfrs. Ass’n v. State Farm Mut. Auto. Ins. Co., 463 U.S. 29, 43 (1983).
Owen M. Fiss, The Forms of Justice, 93 Harv. L. Rev. 1, 2–3, 29–31 (1979).
Robert W. Gordon, Unfreezing Legal Reality: Critical Approaches to Law, 15 Fla. St. U. L. Rev. 195, 196–200 (1987).
See Lisa Heinzerling, The Power Canons, 58 Wm. & Mary L. Rev. 1933, 1986–91, 1999–2001 (2017), https://wmlawreview.org/sites/default/files/Heinzerling.pdf (demonstrating how the framing of an interpretive question can reflect subjective preferences and redistribute authority among courts, agencies, and Congress).
Model Rules of Pro. Conduct pmbl. ¶ 1 (Am. Bar Ass’n 2024), https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/model_rules_of_professional_conduct_preamble_scope/; see also id. ¶ 6 (calling upon lawyers to improve the law, expand access to the legal system, and address economic and social barriers to adequate legal assistance).
Mari J. Matsuda, When the First Quail Calls: Multiple Consciousness as Jurisprudential Method, 11 Women’s Rts. L. Rep. 7, 8–10 (1989).
See ABA Comm. on Ethics & Pro. Resp., supra note 13, at 2–3 (identifying research, drafting, contract analysis, due diligence, and other uses of generative artificial intelligence in legal practice); Colleen V. Chien & Miriam Kim, Generative AI and Legal Aid: Results from a Field Study and 100 Use Cases to Bridge the Access to Justice Gap, 57 Loy. L.A. L. Rev. 903, 939–60 (2025), https://digitalcommons.lmu.edu/llr/vol57/iss4/2/ (documenting legal-aid uses including research, drafting, document analysis, client intake, and translation); Legal Servs. Corp., The Justice Gap: The Unmet Civil Legal Needs of Low-Income Americans 18–19 (2022), https://justicegap.lsc.gov/resource/2022-justice-gap-report/ (reporting that cost concerns frequently deter low-income Americans from seeking legal assistance and that ninety-two percent of their substantially consequential civil legal problems received no or insufficient legal help).
Stan Lee & Steve Ditko, Spider-Man!, Amazing Fantasy no. 15, at 11 (Marvel Comics Aug. 1962).
Derrick A. Bell, Jr., Serving Two Masters: Integration Ideals and Client Interests in School Desegregation Litigation, 85 Yale L.J. 470, 471–82, 512–17 (1976).
Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 448–49, 461–66 (S.D.N.Y. 2023).
See James Reason, Human Error: Models and Management, 320 BMJ 768, 768–70 (2000), https://doi.org/10.1136/bmj.320.7237.768 (contrasting an individual-blame approach to error with a systems approach that identifies recurrent conditions and constructs institutional defenses against future failures).
See Benjamin Minhao Chen, Alexander Stremitzer & Kevin Tobia, Having Your Day in Robot Court, 36 Harv. J.L. & Tech. 127, 162–65 (2022), https://jolt.law.harvard.edu/assets/articlePDFs/v36/Chen-Stremitzer-Tobia-Having-Your-Day-in-Robot-Court.pdf (finding that hearings and interpretable decisions increased perceived fairness for both human and artificial-intelligence adjudicators and challenging the proposition that human decisionmakers possess an irreducible procedural-fairness advantage).
See ABA Comm. on Ethics & Pro. Resp., supra note 13, at 2–15; Elham Tabassi, Nat’l Inst. of Standards & Tech., U.S. Dep’t of Com., NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0) 9–12, 35–37 (2023).
See 5 U.S.C. § 553(b)–(c), (e) (2018); Ryan Calo & Danielle Keats Citron, The Automated Administrative State: A Crisis of Legitimacy, 70 Emory L.J. 797, 801–11, 843–49 (2021).
See Allegra M. McLeod, Envisioning Abolition Democracy, 132 Harv. L. Rev. 1613, 1617, 1623, 1646–48 (2019), https://harvardlawreview.org/print/vol-132/envisioning-abolition-democracy/ (developing a conception of justice grounded in lived experience and positive transformation rather than abstract ideals detached from the people and practices they govern).