Artificial Authorship and Judicial Opinions

Richard M. Re

92 GEO. WASH. L. REV (forthcoming 2024)

Generative AI is already beginning to alter legal practice. If optimistic forecasts prove warranted, how might this
technology transform judicial opinions—a genre often viewed as central to the law? This symposium essay attempts
to answer that predictive question, which sheds light on present realities. In brief, the provision of opinions will
become cheaper and, relatedly, more widely and evenly supplied. Judicial writings will often be zestier, more
diverse, and less deliberative. And as the legal system’s economy of persuasive ability is disrupted, courts will
engage in a sort of arms race with the public: judges will use artificially enhanced rhetoric to promote their own
legitimacy, and the public will become more cynical to avoid being fooled. Paradoxically, a surfeit of persuasive
rhetoric could render legal reasoning itself obsolete. In response to these developments, some courts may disallow
AI writing tools so that they can continue to claim the legitimacy that flows from authorship. Potential stakes thus
include both the fate of legal reason and the future of human participation in the legal system.

I. Framing the Inquiry 4

A. Motivating Premises 4

B. Varieties of AI Assistance 8

II. Initial Effects on Authors 10

A. Quality and Quantity 10

B. Uniformity and Diversity 12

C. Authenticity and Accountability 13

D. Deliberation and Direction 14

E. Reason and Rhetoric 16

F. Canonicity and Customization 18

III. Reactions by Readers 19

A. The Perfect Adversarial System 19

B. Artificially Balkanized Readership 21

C. Rhetoric’s Rise – and Reason’s Demise 23

IV. Seeking Equilibrium 26

A. Transitions and Tradition 26

B. The Rhetoric of Wisdom 28

Conclusion 30


Professor, UVA School of Law. Thanks to Sam Armstrong, Sridevi Dayanandan, and Mary Grace
Triplett for excellent research assistance. Thanks also to Ryan Calo, Rebecca Crootof, Minji Doh,
Landis Hagerty, Mike Livermore, Alicia Solow-Niederman, Chirag Shah, the editors of this law
review, and my co-panelists at a 2023 symposium hosted by the law review.

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

In September 2023, a British court of appeals judge attracted global media attention by
describing his enthusiastic use of generative AI tools to draft court opinions.1 Lord Justice Colin
Birss provided a specific example from his own work:

“I asked ChatGPT, ‘Can you give me a summary of this area of law,’ and it gave
me a paragraph. I know what the answer is because I was about to write a
paragraph that said that, but it did it for me and I put it in my judgment. It’s there
and it’s jolly useful.” 2

Based on these sorts of experiences, the Lord Justice reached a conclusion about generative AI
that is as intuitive as it is simple: “It is useful, and it will be used.” 3

Judicial enthusiasm for generative AI may seem premature, as these tools are currently prone to
error, bias, and other serious defects.4 Yet generative AI has already affected many aspects of life
and work, including legal practice, and optimistic assessments foretell greater advances. In time,
generative AI tools may become far more reliable and effective, even as they remain fast, easy,
and efficient. Just as important, people are learning how best to use the technology. 5

If optimistic predictions prove warranted, how would generative AI affect judicial opinions—a
genre often viewed as the heart of legal practice, if not of the law itself? 6 This essay aims to
answer that question, which calls for an exercise of speculative imagination. As with many
thought experiments, the goal is not just to make heroic predictions, but also to shed light on the
present we already inhabit. 7 For instance, legal practice generally takes for granted that
persuasive resources are finite and costly, yielding an economy of persuasiveness. Yet AI tools

1 Hibaq Farah, Court of Appeals Judge Praises ‘Jolly Useful’ ChatGPT After Asking It For Legal
Summary, GUARDIAN (Sept. 15, 2023). When referring to “generative AI” or “AI,” I generally have in
mind artificial intelligence systems based on large language models, or LLMs.
2 Id. (punctuation altered for readability).
3 Id. See also Chief Justice Roberts, 2023 Year-End Report on the Federal Judiciary at 6. For

additional examples of judges using ChatGPT, see Juan David Gutierrez, Judges and Magistrates in
Peru and Mexico Have ChatGPT Fever, TECH POLICY PRESS (April 19, 2023).
4 For instance, existing tools can generate “hallucinations.” See, e.g., Molly Bohannan, Lawyer

Used ChatGPT In Court—And Cited Fake Cases, FORBES (June 8, 2023). For a measured prognosis,
see Steve Lohr, A.I. Is Coming For Lawyers, Again, N.Y. TIMES (April 10, 2023).
5 See Rebecca Crootof, Margot E. Kaminski & W. Nicholson Price II, Humans in the Loop, 76

VAND. L. REV. 429 (2023) (discussing pitfalls in human-tech integration); see also infra notes.
6 For theories that center judicial opinions, see JAMES BOYD WHITE, HERACLES’ BOW: ESSAYS ON

THE RHETORIC AND POETICS OF THE LAW (1985); DAVID A. STRAUSS, LIVING CONSTITUTION (2010); see
also PAUL W. KAHN, MAKING THE CASE: THE ART OF THE JUDICIAL OPINION (2016); MARTHA C.
NUSSBAUM, POETIC JUSTICE: THE LITERARY IMAGINATION IN PUBLIC LIFE (1996).
7 See text accompanying infra note 30. Such efforts may even be fictional. See L. Bennett Capers,

Afrofuturism, Critical Race Theory, and Policing in the Year 2044, 94 N.Y.U. L. REV. 1, 3–4, 6–8
(2019) (discussing “futurist legal scholarship”); see, e.g., Derrick Bell, The Space Traders, in FACES AT
THE BOTTOM OF THE WELL: THE PERMANENCE OF RACISM 158 (1992).

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

will challenge that premise, thereby revealing its importance.8 So there is something to learn
here, even if technological optimism proves unwarranted.

My argument poses two interrelated paradoxes. The first is that a proliferation of legal reasoning
may lead to its abandonment. The legal system is already soaked in texts—not just statutes,
regulations, and rules, but also case law, briefs, and commentary. Because of AI authorship, we
will soon bear witness to an exponential increase in textuality on at least the same scale as the
development of text-searchable databases (and perhaps comparable to the advent of published
case reports).9 This transformation could conceivably bring about a perfect adversarial system, in
which dueling AIs efficiently generate the best possible arguments for opposing views, and the
legally right answer (or the lack of such an answer) consequently becomes clear.10 Or it could
yield an ever more balkanized discourse, as readers increasingly exalt in the flawless rhetoric of
their favorite side while disparaging the contrary opinions of their opponents.

At least as likely, however, is a quite different outcome. An explosion in the supply of effective,
cogent verbiage will render persuasive reasoning better and cheaper, regardless of its legal
correctness. Even lay readers will know that AI is mainly responsible, that almost anyone (with
AI tools) could produce similarly impressive texts, and that an AI-authored opinion could
support virtually any conclusion in a contested case. Judicial opinions, and legal reasoning in
general, will become even more demystified than they already are. Would-be readers who have
come to see the futility of reading judicial opinions will then put down the case reports and
attend to other indicia of merit or correctness, such as the identity of the author, the views of
trusted commentators, and the likely consequences of court determinations. 11 In a legal system
overflowing with persuasive reasoning, there might as well be none at all.

The second paradox is partly, but only partly, derivative of the first one: courts will find use of
AI tools almost irresistibly attractive—yet these tools threaten the courts’ institutional interests
and so will become an object of concerted judicial resistance. The appeal of these tools is largely
self-evident, as they make judicial work easier, faster, and more effective. Legal practice aside,

8 See infra Section III.C.

9 See STUART BANNER, THE DECLINE OF NATURAL LAW: HOW AMERICAN LAWYERS ONCE USED
NATURAL LAW AND WHY THEY STOPPED (2021) (describing the explosion of case reports).
10 AI is of course altering many aspects of the adversarial process, including legal research.

LexisNexis, for instance, has rolled out an AI-assisted research tool, called Lexis+ AI. One related
but distinct use of AI will be to assess empirical propositions of legal import, some of which may be
included in a judicial opinion. See, e.g., Yonathan Arbel & David A. Hoffman, Generative
Interpretation, 99 N.Y.U. L. REV. (forthcoming 2024); see also Snell v. United States Ins., No. 22–
12581 (CA11 2024) (Newsom, J., concurring) (considering “whether and how AI-powered large
language models like OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude might—might—
inform the interpretive analysis”).
11 As in the 18th Century, the law “would be what it is, not because it has been so reported, but

because it has been so decided.” M. Tiersma, The Textualization of Precedent, 82 NOTRE DAME L.
REV. 1187, 1207 (2007) (quoting the REPORT OF THE LORD CHANCELLOR’S COMMITTEE (1940)).

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

moreover, these tools are already becoming staples of society at large. Yet artificial authorship
will tend to dissolve judicial authority by democratizing the ability to both produce and
understand sophisticated legal prose. Courts might also fear that regulators will want a say in
how courts make use of transformative new technologies. So while generative AI may indeed be
“jolly useful”12 for courts, it is also quite perilous.

Those considerations, in addition to the fact that most current judges have no familiarity with
these newfangled gizmos, and much intuitive distrust of them, make it plausible to imagine that
exclusively or primarily human authorship could become part of the judicial ethic. By abjuring
AI authorship, the judiciary could avoid becoming an object of regulation and preserve the
impression that justice is a proper subject for human judgment. As AI becomes ubiquitous
elsewhere, legal culture might come to view great legal reasoners and writers as akin to chess
grandmasters. Sure, a computer could do their job just as well, or much better. But the
impressiveness of the human judges’ achievement might still remain. The resulting pride of
human authorship might delay or curtail AI authorship in some quarters.

So the stakes in the present inquiry are quite large, touching on both the fate of reasoned legal
argument and the shape of human participation in the legal system.

I. Framing the Inquiry

Let me start with how I approach the topic of AI and judicial opinions, including some
methodological and empirical assumptions.

A. Motivating Premises

The rise of artificial authorship is a widespread phenomenon, affecting private correspondence, 13

government propaganda,14 and corporate public relations (among other things).15 But courts have
a distinctive and in some ways heightened interest in this topic. Unlike other government actors,
courts tend to operate through their justificatory public statements.16 The meaning of a judgment
often depends on its accompanying opinion (“The case is remanded for proceedings consistent
with this opinion”), and a precedential rule—the proverbial “holding” of a court—derives much

12 Farah, supra note 2.

13 See, e.g., David Emelianov, AI Impact on Email Correspondence: Empowering
Communication, Trimbox (Jan. 15, 2024), available at https://www.trimbox.io/blog-posts-
writer/ai-impact-on-email-correspondence-empowering-communication.
14 Jessica Brandt, Propaganda, foreign interference, and generative AI, Brookings (Nov. 8,

2023), available at https://www.brookings.edu/articles/propaganda-foreign-interference-and-

generative-ai/.
15 See, e.g., Nicholas Berryman, AI in Public Relations: the Benefits and Risks of Change,

Prowly (May 14, 2024), available at https://prowly.com/magazine/ai-in-public-relations/.

16 See Frederick Schauer, Opinions as Rules, 62 U. CHI. L. REV. 1455 (1995).

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

if not all of its content from its surrounding justification. By comparison, legislators and
presidents tend to generate either operative directives or rhetorical soundbites—one or the other.
Administrative agencies offer a closer example, insofar as their directives may stand or fall
depending on their accompanying explanations.17 But even then, the directive and justifications
remain separated, both conceptually and practically.

Anglo-American courts’ merger of directive and justification has not just functional but ethical
implications. And by ethical I mean to capture both the idea of moral rightness as well as a
professional ethos. That is, a judicial opinion is often thought to convey full authority or
legitimacy only because (or if) its author has offered an adequate justification. Similarly, the
judges who produce judicial opinions are often thought to be fulfilling their roles—to be
instantiating the character of their occupation—only if they generate adequate justifications. The
issue, in other words, isn’t whether the decision is justified at all or by anyone. The issue, at least
in part, is whether the deciding court itself has adduced an acceptable justification. Judicial
decisions without any accompanying justification can therefore be unsettling and are usually
deemed tentative or otherwise peripheral—even though such decisions could be rationalized by
one of the parties or by outside commentators.18

But if authorship is integral to the judiciary, its role is also complex and qualified. Judicial
authors often borrow heavily from one another as well as from the filings of parties, submissions
by so-called “friends of the court,” and scholarly articles. These forms of borrowing are not
always or fully attributed. Further, judges are generally assisted by officially anonymous clerks
who “ghostwrite” the vast majority of judicial opinions.19 Judicial authorship is therefore
something of a collective and specialized undertaking. The production of a judicial “voice,”
while important to any judge’s professional reputation, may be more akin to an art studio in
which a senior virtuoso manages the apprentices’ work before signing it with his own name.20
More concretely, judicial authorship already makes extensive use of ratified rationalizations,

17 See DAVID FREEMAN ENGSTROM ET AL., GOVERNMENT BY ALGORITHM: ARTIFICIAL INTELLIGENCE

IN FEDERAL ADMINISTRATIVE AGENCIES (2020); Danielle K. Citron & Ryan Calo, The Automated
Administrative State: A Crisis of Legitimacy, 70 EMORY L. J. 797 (2021); e.g., Motor Vehicle Mfrs.
Assn. of United States, Inc. v. State Farm Mut. Automobile Ins. Co., 463 U.S. 29 (1983).
18 Any discomfort with unexplained rulings does not prevent them from being commonplace and

sometimes justifiable, see Mathilde Cohen, When Judges Have Reasons Not to Give Reasons: A
Comparative Law Approach, 72 WASH. & LEE L. REV. 483 (2015), though they often attract attention
as part of critiques. See, e.g., STEPHEN VLADECK, THE SHADOW DOCKET: HOW THE SUPREME COURT
USES STEALTH RULINGS TO AMASS POWER AND UNDERMINE THE REPUBLIC 239 (2023) (criticizing
“shadow docket” rulings in part for often being unexplained); see also infra note 36, at 591–92.
19 See, e.g., Stephen J. Choi and G. Mitu Gulati, Which Judges Write Their Opinions (And Should

We Care)?, 32 FLA. ST. U. L. REV. 1077 (2005) (“Law clerks are said to draft the majority of opinions
for judges.”).
20 See Peter Friedman, What Is a Judicial Author?, 62 M ERCER L. REV. 519 (2011); see, e.g.,

Michiel Franken and Jaap van der Veen, The Signing of Paintings by Rembrandt and His
Contemporaries, in THE LEIDEN COLLECTION CATALOGUE (Arthur K. Wheelock Jr. and Lara Yeager-
Crasselt, eds.) (2023).

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

that is, justificatory texts that are authored in the first instance by someone other than the judge
but then officially endorsed after the fact.

Against this backdrop, one might think that AI tools will simply replace the law clerk, without
otherwise disrupting legal practice. But the advent of the law clerk is itself thought to have
changed judicial practice in meaningful ways, as did the subsequent rise of email, searchable
legal databases, and word processing.21 New capabilities engender novel choices, altered
incentives, and reshuffled power relationships. At the Supreme Court, for instance, justices
armed with law clerks and laptops have tended to create opinions that are longer, more citation-
studded, and similar to one another.22 Meanwhile, lower courts with the same resources may be
more adept at either avoiding or, if they choose, creating circuit splits for the Supreme Court to
review.23 And analogous technologies have empowered courtwatchers, granting instant access to
court rulings as well as the ability to fact check them in real time. 24 In earlier eras, by contrast,
judges tended to be more personally responsible for their pronounced or published opinions,
lower courts had trouble keeping track of what their colleagues were deciding, and the public had
limited, delayed access to most court decisions.

To a great extent, artificial authorship is not just inevitable but actual. Thanks to programs like
ChatGPT, millions of people are already taking advantage of artificially intelligent editing and
drafting. So one might think that we need only look around to discover how new writing tools
affect legal practice. Yet both the technology’s development and its uses are still very much in
flux. Just as important, the law itself is a dynamic system with many opportunities for re-
equilibration and adjustment. So when a new technology generates pressure for institutional
change over here, the upshot may simply be a new counter-pressure over there, somewhere else
within the system. This systemic perspective can help us recognize the importance of
technologically driven change in the law without succumbing to either utopianism or
doomsaying.25 That is, we can anticipate the relevant changes, as well as responses,
countermoves, and interventions.

The present inquiry is more focused and modest than many recent discussions of artificial
intelligence and the law. “Robot judges” have understandably attracted a great deal of attention

21 See, e.g., RICHARD POSNER, THE FEDERAL COURTS: CRISIS AND REFORM 102 (1985).
22 See Keith Carlson, Michael A. Livermore & Daniel Rockmore, A Quantitative Analysis of
Writing Style on the U.S. Supreme Court, 93 WAS. U. L. REV. 1461 (2016).
23 See Allison Orr Larsen and Neal Devins, The Amicus Machine, 102 VA . L. REV. 1901, 1949

(2016) (“there is at least some support for the claim that circuit splits are less common in a world in
which the lower courts have greater access to one another's opinions”).
24 See, e.g., Dan Farber, More About EPA’s Victory, LEGAL PLANET (April 29, 2014).
25 See Edward A. Parson, Richard M. Re, Alicia Solow-Niederman, & Elana Zeide, Artificial

Intelligence in Strategic Context: An Introduction, SSRN (2019). Cf. ABDI AIDID & BENJAMIN ALARIE,
THE LEGAL SINGULARITY: HOW ARTIFICIAL INTELLIGENCE CAN MAKE LAW RADICALLY BETTER 1410–41
(2023) (providing an optimistic long-term perspective).

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

as it has become more plausible to imagine chatbots and other algorithmic tools dictating specific
decisions presently left to human judgment. 26 Bail decisions and the imposition of punishment—
both of which can be determined largely if not entirely based on algorithms—are among the
most developed and salient examples.27 While these applications are certainly important, they
have too much eclipsed a set of subtler and now seemingly more imminent scenarios involving
language, communication, and reason-giving.

With a human “in the loop,” moreover, many problems with AI judges diminish or disappear. 28
There is a clearer basis for political legitimacy, a possibility of empathy and interpositionality,
and an assurance of reasonableness deriving from the judge’s everyday life in a flesh and blood
society. In effect, the gradual schedule of technological change is forcing us to grapple with
cases of mixed human-and-AI decision-making, in all their complexity, before fully resolving or
understanding some of the conceptually “easier” (but technologically harder) cases of near-total
automation. Still, the complex and easier inquiries are interlinked, not only because our view of
the stark scenarios can inform the blurry ones, but also because the judiciary’s use of AI today
will influence how the technology is deployed for years to come.

Throughout, I will take for granted certain basic features regarding the technology behind AI
authorship. Perhaps most importantly, I generally assume that highly effective AI tools will be
widely available and used, including by many courts. This assumption seems quite plausible
given the current use of fairly effective AI tools at no charge by millions of people. Still, a
dramatic increase in draconian regulation or an abrupt halt to technological progress in this area
could disrupt this premise. At the same time, I assume that the returns to AI authorship are both
finite and, at some point, diminishing. That is, even the most effective AI-authored text will not
be tantamount to mind control or the song of the sirens. And while different AI tools will
doubtless exhibit varying degrees of quality, and are already being specially designed for certain

26 See Eugene Volokh, Chief Justice Robots, 68 DUKE L. J. 1135 (2018); Benjamin Minhao Chen,
Alexander Stremitzer & Kevin Tobia, Having Your Day in Robot Court, 36 HARV. J. L. & TECH. 127
(2022); Richard M. Re & Alicia Solow-Niederman, Developing Artificially Intelligent Justice, 22 STAN.
TECH. L. REV. 242 (2019); Rebecca Crootof, “Cyborg Justice” and the Risk of Technological-Legal
Lock-In, 119 COLUM. L. REV. FORUM 233 (2019); Tim Wu, Will Artificial Intelligence Eat the Law?
The Rise of Hybrid Social-Ordering Systems, 119 COLUM. L. REV. 2001 (2019).
27 See Alexis Morin-Martel, Machine Learning in Bail Decisions and Judges’ Trustworthiness, AI

& SOC’Y (2023); Jessica M. Eaglin, Racializing Algorithms, 111 CAL. L. REV. 753 (2023); Sandra G.
Mayson, Bias In, Bias Out, 128 YALE L.J. 2218 (2019).
28 See Crootof, Kaminski & Price, supra note 5; Thomas Julius Buocz, Artificial Intelligence in

Court: Legitimacy Problems of AI Assistance in the Judiciary, 2 COPENHAGEN J. OF L. STUD. 41

(2018); Aziz Z. Huq, A Right to a Human Decision. 105 VA. L. REV. 611 (2020); Kiel Brennan-
Marquez & Stephen E. Henderson, Artificial Intelligence and Role-Reversible Judgment, 109 J. CRIM.
L & CRIMINOLOGY 137, 146 (2019).

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

goals and clients,29 I assume that these comparative differences will be relatively small compared
with the effects of having an AI tool at all.

The point of exploring this topic has more to do with seeing the various possibilities, pressures,
and equilibria than making ironclad forecasts or even placing odds on specific outcomes.
Imaginative predictions of the type offered here are partly aimed at better understanding what is
happening in the present.30 Our goal is not just to get ahead of a technological wave that is
already rising above the legal profession, but also to find a fresh way of thinking about how legal
writing has long worked, what role it currently plays, and how it might imminently be improved.
AI underlines these questions, and may change how we answer them, but the questions mattered
long before anyone had heard of ChatGPT.

B. Varieties of AI Assistance

Artificial authorship is a complex category potentially involving every aspect of resolving a case.
As a rough first cut, we might begin by distinguishing two dimensions of appellate decision-
making: the components of a judicial ruling and the degree of AI assistance. Appellate decisions
generally involve the selection of precedential rules, case results, and justificatory opinions. And
human judges could act independent of AI, receive guidance from AI tools, or delegate
decisional authority to the AI.31 These dimensions of AI assistance, along with their related
points of intersection, are outlined in Figure 1 below:

29 Some firms are already touting their creation of bespoke LLMs designed specifically for legal
services. See, e.g., Patrick Smith, Sullivan & Cromwell’s Investments in AI Lead to Discovery,
Deposition 'Assistants’, Law.com (August 21, 2023), available at
https://www.law.com/americanlawyer/2023/08/21/sullivan-cromwell-investments-in-ai-lead-to-
discovery-deposition-assistants/?slreturn=20240622060835.
30 See supra note 7 (collecting sources).
31 Cf. ARTEMUS WARD & DAVID L WEIDEN, SORCERERS’ APPRENTICES : 100 YEARS OF LAW CLERKS

AT THE UNITED STATES SUPREME COURT (2007) (discussing justices’ working relationship with clerks
generally shifting from a “retention model” to a “delegation model”). The categories in Figure 1 can of
course be embellished. For example, an expanded typology would account for finding facts, a key
undertaking of trial-level courts. And the idea of guidance might be expanded to include various
aspects of deliberation, such as identifying applicable case law, generating relevant factors, and
weighing or ranking competing considerations.

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

Figure 1: Some Dimensions and Varieties of AI Assistance

Independence Guidance Delegation

Judge crafts a rule, AI proposes or assesses

Rule AI selects or crafts a rule
independent of AI a rule for human review

Judge identifies a case AI proposes or assesses

Result result, independent of a result for human AI selects or crafts a result
AI review

Judge drafts or revises a AI proposes or assesses

Opinion judicial opinion, a draft opinion for AI crafts a judicial opinion
independent of AI human review

Some scenarios fall neatly into one or another of the above categories. For example, a judge
might simply ask an AI which way to rule and then abide by the resulting recommendation
(Result Delegation). Or an AI might suggest edits to the penultimate draft of a nearly finished
opinion (Opinion Guidance).

Still, these categories are interlinked, mix-and-matchable, and present to various degrees. For
instance, the selection of an outcome or rule will naturally influence what kind of opinion is
practically available. A judge might opt for Result Independence, then seek Rule Guidance, and
finally engage in Opinion Delegation. Or a judge might independently select both the result and
the rule (Result and Rule Independence), only to change her mind on both counts during an AI-
assisting writing process (Opinion Guidance). In that last scenario, when the experience of
writing an opinion suggests a new outcome or rule, judges might say that the initially envisioned
opinion “wouldn’t AI-write.”32 An intended instance of Opinion Guidance would thus become
an occasion for both Result Guidance and Rule Guidance.

For present purposes, the most important category is Opinion Guidance. AI now seems most
immediately able to supply assistance in that domain, generating significant practical effects in
the near future. This conclusion is somewhat surprising, as most discussion of AI adjudication
has focused on tools bearing on Result Guidance, or even Result Delegation—that is, use of
algorithmic tools to furnish technical information bearing, with various degrees of
conclusiveness, on discrete case outcomes. 33 By comparison, the idea that AI would so quickly
be able to engage in the seemingly ultra-nuanced task of drafting or improving sophisticated

32 Judges sometimes conclude that “an opinion would not write.” Richard A. Posner, Judges'
Writing Styles (and Do They Matter?), 62 U. CHI. L. REV. 1421, 1448 (1995).
33 See supra text accompanying notes 22-24.

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

human expression seemed improbable just a few years ago.34 The rise of AI rhetoricians,
moreover, is virtually the opposite of the time-honored trope that robots are, well, robotic. In
fact, it seems that AI will be lucid and charming long before it can accurately or reliably make up
its own mind.

That said, Opinion Guidance can come in many forms, including every stage or aspect of a
judge’s work on an opinion:

● Argumentation: generating possible arguments and counterarguments

● Composition: drafting or proposing revisions to a draft opinion
● Commentary: analyzing a draft opinion’s style, strengths, and weaknesses
● Recommendation: identifying the best way to compose an opinion

Here, too, conceptual distinctions blur in practice, as existing AI tools invite iterative and
reflective use. True, the AI will sometimes do far more than simply act as an editor, in that it will
compose sentences, paragraphs, and entire sections for human review and further modification.
Yet the AI’s first response will often lie far from the final product that is used or published.
Generally, then, there is no clear line between the generation of options, the refining of options
already on the table, and the selection among options.

II. Initial Effects on Authors

What are the first-order or immediate effects of artificial authorship?

A. Quality and Quantity

A judicial author faces a series of tradeoffs that can be framed in economic terms. 35 Producing a
strong opinion takes time, which means less work for other opinions, for other types of judicial
work, and for leisure. There is also a maximum amount of effort that a judge can realistically
expend over any given period. So when judges expend effort on opinions, they are optimizing
along several variables.

Against that backdrop, the most immediate effect of AI tools is that the cost of producing
effective writing will decline. In principle, then, any given judge would be able to maintain the
same effort as to other forms of judicial work while improving the quality of her writing. Or,
equivalently, the judge could maintain the same quality of prose but devote more time to other

34 See Ajeya Cotra, Language Models Surprised Us, PLANNED OBSOLESCENCE (BLOG) (August 29,
2023) (“ML researchers, superforecasters, and most others were all surprised by the progress in
large language models in 2022 and 2023.”).
35 See RICHARD POSNER, HOW JUDGES THINK 35–40 (2008).

10

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

aspects of her work, such as identifying correct results. Viewed from either of these standpoints,
AI tools would be an unmitigated improvement to the legal system.

But the answer isn’t that simple. When the cost of a good declines, the optimal response is often
to consume more of that good while trading down consumption of other goods. In the context of
judicial opinions, that might mean spending additional hours making opinions more readable and
entertaining, thereby enhancing the judge’s reputation, rather than using the time freed up by AI
tools to make sure that cases are decided correctly. Writing more opinions, or more persuasive
ones, could distract judges from the task of identifying the best outcomes and rules.

In addition, demand for judicial opinions might change in light of their diminished cost. Today,
many forms of adjudication lack any opinion whatsoever or come with only short, technical
explanations.36 But AI will soon allow almost every determination—from certiorari denials to
routine appellate affirmances to trial-court minute orders—to come with an instantly generated,
artificially authored explanation. 37 And those explanations could be tailor-made for the specific,
legally unsophisticated individuals involved in many retail-level disputes.38 The result would be
an explosion of judicial prose. And increases in the total volume of judicial opinions will tend to
mitigate improvements in quality. 39

Whether the overall effect is viewed as desirable partly depends on how we imagine artificially
authored opinions that are relatively low-quality but that would not have existed at all without AI
assistance. These opinions could be viewed as a pure gain for the legal system. Something is
better than nothing, one might say. From another standpoint, however, these opinions could pull
down the average quality of judicial writing in the system, and they risk degrading the public’s
view of judicial opinions as a genre. If these opinions are lacking in quality, or are discounted as
cheap robot-talk (which in some sense they would be), then they might tarnish the overall
perception of the judiciary and its work product. 40

36 For data on summary denials of appeal in the federal system, see Merritt E. McAlister,
“Downright Indifference”: Examining Unpublished Decisions in the Federal Courts of Appeals, 118
MICH. L. REV. 533 (2020).
37 See text accompanying infra note 64. AI authorship will operate differently in different courts.

For instance, front-line adjudicators who resolve disputes at scale will likely veer toward Opinion
Delegation, whereas apex appellate tribunals will tend toward Opinion Guidance. See also text
accompanying infra note 65.
38 Generating internal court memos, such as the initial “pool” memos evaluating petitions for

certiorari at the Supreme Court, also seem like apt tasks for AI tools.
39 Cf. RICHARD SUSSKIND, ONLINE COURTS AND THE FUTURE OF JUSTICE 8–9 (2019).
40 See Elise Karinshak, Sunny Liu, Joon Park, & Jeffrey Hancock, Working with AI To Persuade:

Examining A Large Language Model’s Ability to Generate Pro-Vaccination Messages, 7 PROCEEDINGS

OF THE ACM ON HUMAN-COMPUTER INTERACTION 1, 19 (2023) (presenting a study indicating that,
while ChatGPT created effective messaging, audiences devalued those messages when they knew
that the message was created by an AI). One obvious point of comparison is the therapeutic ELIZA
chatbot from the 1960s, which people felt comfortable chatting with even though—or because—they
knew it was a machine. See JOSEPH WEIZENBAUM, COMPUTER POWER AND HUMAN REASON: FROM

11

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

B. Uniformity and Diversity

In this enhanced writing environment, weak or mediocre writers will enjoy the greatest relative
improvement.41 Almost anyone will be able to produce AI-assisted prose, after all. The writing
quality of a standard AI tool will therefore tend to establish a baseline or floor for all minimally
competent users of AI tools, even if those users are neither particularly good at using the tools
nor talented writers on their own. So while excellent and especially hard-working writers will
likely be able to eke out meaningful improvements over the AI-facilitated baseline, the marginal
returns on that effort will be both small and diminishing.

It would be tempting to conclude that prose will become more uniform and bland as AI guides
all writers to converge on the same efficient and artificial style. 42 Yet AI tools also enable
authors to express their personalities or adopt idiosyncratic writing personas. Mediocre writers,
whether judges or clerks, may be trapped in a familiar style or simply unable to conceive of a
creative way to express themselves. Few writers are poets. And judges and clerks are always
selected based on many criteria other than writing virtuosity.

But artificial authorship can already convert prose into poetry with the touch of a button. And it
can alter the tone of any text, including by assuming the voice of a desired speaker. The AIs, in
other words, will be much more versatile writers than clerks. In some ways, these tools already
are superior. To see this, copy a passage of a judicial opinion into ChatGPT and ask it to convert
the text into a haiku, sonnet, or hymn. As these capabilities increase, the result may be an
increase in rhetorical personality and diversity.

Paradoxically, AI tools may tend to promote both uniformity and panache. Capabilities are a
critical determinant of style in part because they limit what is possible. But writers produce for
audiences, so consumer demand is often the most important factor. As AI makes it easier to write
both clearly and entertainingly, writers will take advantage of those opportunities. And with
artificial authorship enabling almost anyone to write more like the most popular writers around,
professional standards will rise. Judicial writing could also become stylistically dynamic, even
faddish, as jurists instruct their AIs to match the latest trends.

JUDGMENT TO CALCULATION 188-191 (1976). Similarly, some people might prefer to be “judged” only
by a machine. “It isn’t personal,” they might think.
41 See Jonathan H. Choi & Daniel Schwarcz, AI Assistance in Legal Analysis: An Empirical

Study, 73 Journal of Legal Education (forthcoming 2024), available at

https://dx.doi.org/10.2139/ssrn.4539836 (reporting a study in which weaker scorers showed the
greatest improvements from AI use).
42 See Vishakh Padmakumar and He He, Does Writing with Language Models Reduce Content

Diversity?, arXiv preprint arXiv:2309.05196 (2023) (presenting study evidence that writing with
“InstructGPT (but not the GPT3) . . . increases the similarity between the writings of different
authors and reduces the overall lexical and content diversity”).

12

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

C. Authenticity and Accountability

Will judicial opinions tend to represent authentic expressions of a judge’s actual views and
personality, as opposed to rationalizations or an assumed persona?

Artificial authorship might seem incompatible with any kind of human authenticity, and in many
instances that assumption will be borne out. Again consider the vast number of decisions that
presently come with no explanations, such as one-line denials of appeal.43 A court might set up
an AI that reviews the record and generates a plausible explanation for the ruling, thereby
affording the parties the dignity of an understandable explanation. Yet the automated explanation
might fail to capture, or even be unrelated to, the actual basis for decision. 44 And readers unable
to tell the difference between authentic and artificial expression might become cynical about
judicial opinions in general, treating them all as cheap, fake AI talk. 45

In many other situations, however, artificial authorship will foster or enhance authenticity. Life
coaches sometimes talk about making people “the best versions of themselves.” AI tools, like
human editors and law clerks, might similarly make opinion writers realize their best selves,
achieving what might be called aspirational authenticity. When judges are motivated by a
complex legal argument, or else by an elusive, ineffable moral intuition, writing assistance might
help the judges both form and communicate their ideas.

Better writing would thus mean better access to authors and their true thought processes. Several
consequences follow. Public engagement with legal decision-making may increase.46 The
public’s better understanding of the judiciary could facilitate political efforts to hold courts
accountable. And readers who can understand what judges are talking about might end up being
persuaded, leading them to afford courts greater legitimacy. What was once a technocratic and
inaccessible profession might become open and relatable.

But what about a judge who uses writing tools to produce empathetic, entertaining, or relatable
prose simply to placate readers? Whereas the authentic judge wants an opinion whose form and
substance harmonizes with her own considered thoughts, the cynical judge seeks a particular

43 For an argument that unexplained or summary denials of appeal are objectionable, see
McAlister, supra note 36, at 585.
44 See Andrew D. Selbst & Solon Barocas, The Intuitive Appeal of Explainable Machines, 87

FORDHAM L. REV. 1085 (2018). And, again, parties may be put off by the knowledge that they are
getting only a robotic explanation. See supra note 40.
45 This tendency is already arising in many areas of social life, as emails, speeches, love letters.

And one possible response is to create rules requiring disclosure when AI tools are used, or else
prohibitions on using such tools altogether. See infra Part III.
46 In this sense, generative AI will foster “demosprudence.” See generally Lani Guinier, Foreword:

Demosprudence Through Dissent, 122 HARV. L. REV. 4 (2008).

13

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

effect on the audience, even though the result misaligns with his own personality or views. A
self-absorbed, egotistical jurist might opt to generate opinions that exhibit manufactured
empathy.47 Any resulting legitimacy would then flow from inauthenticity.

Yet inauthenticity has its benefits. Imagine an angry judge who, to promote a favorable public
reputation, always asks a clerk or AI tool to render his opinions calm and courteous and so
becomes a popular symbol of collegiality. 48 Though subject to personal criticism for his motives,
a cynic may still have engaged in beneficial conduct. The persona, though inauthentic and
therefore misleading in some respects, might even be for the best. Far from being mere
deception, the persona would have become a beacon of virtue. 49

So both authenticity and its opposite are appealing in different ways and circumstances, a truism
discernible in popular aphorisms like “do as I say, not as I do,” or “hypocrisy is the tribute vice
pays to virtue.” Whether beneficial inauthenticity is well-motivated—that is, whether it is
aspirational or cynical—is thus generally of secondary importance. Nor is the reality/perception
distinction particularly critical. The key question, which admits of no simple answer, is whether
to prioritize accountability based on the judge’s public or private self.

D. Deliberation and Direction

AI can and often will improve judicial deliberation. For example, a judge could call upon an AI
to brainstorm arguments and counterarguments or to conduct research that parties overlooked. Or
the judge could instruct the AI to point out draft prose that has certain problematic features,
much as a confident editor or intrepid clerk might “push back” on an errant passage. 50 AI tools
may thus increase both the volume and the quality of internal debate among judges. The result
would be to challenge judges’ biases, deepen their understanding of the law, and enrich their
appreciation of competing points of view.

To some extent, the AI’s deliberative efforts—like a clerk’s—will supplement the adversarial
system of litigation. AI’s ability to find missing arguments or details might prove especially
useful when parties are under-resourced or otherwise fail to advance the best litigation

47 Thanks to Larry Sager for this phrase.

48 Insincere respectfulness could be viewed as a key life skill, including in the judiciary.
49 People might come to view judicial affect as too easily constructed to place any faith in it. Fake

empathy, for instance, might be so hard to separate from the genuine article that readers assume
that no judge is truly empathetic. In essence, readers might become more cynical to avoid being
fooled. See also infra Section III.A. Judges who sincerely express virtuous ideas might then fail to get
credit for them. Cf. Robert Chesney and Danielle Keats Citron, Deep Fakes: A Looming Challenge for
Privacy, Democracy, and National Security, 107 CALIF. L. REV. 1753, 1785 (2018) (discussing the
“liar’s dividend” resulting from deep fakes).
50 See, e.g., Gil Seinfeld, The Good, the Bad, and the Ugly: Reflections of a Counterclerk, 114

MICH. L. REV. FIRST IMPRESSIONS 111 (2016).

14

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

positions.51 AI-facilitated deliberation could therefore render deliberation more fairly distributed
among claimants, mitigating the often inegalitarian distribution of attention that results from
inequalities in wealth and representative capacity.

Yet opinion-writing could become less deliberative in some respects. The experience of
composing a judicial opinion is thought to improve the final product, 52 and jurists sometimes
assert that their intended opinion “wouldn’t write,” leading them to adopt a different and
presumably better conclusion. As noted earlier, AI would sometimes help judges realize that a
planned opinion just won’t do. 53 On balance, however, that sort of experience will become rarer
as artificial authorship makes it easier to generate a plausible-sounding opinion with respect to
any given viewpoint, even if the viewpoint is substantively weak. Fewer intended dispositions or
draft opinions will seem like dead ends, and writing will be easier overall.

Deliberation also bears on the author’s character over time. For example, simply choosing to
publish respectful opinions is compatible with having a standing order that all clerks and AIs are
to write in a respectful style. But it is quite another thing to enact respectfulness. Doing so means
sitting down and actually writing one courteous opinion after another, generating considered text
while suppressing snark and snideness. 54 The intellectual labor that goes into that effort can
transform authors into their persona. Easy writing makes virtuous writing easier to display, but
also circumvents deeper processes of transformation.

A similarly formative intellectual labor currently goes into translating sophisticated legal
information for the benefit of lay audiences. Yet, as discussed earlier, an AI could help explain
technical decisions for mass consumption.55 An interminable hearing transcript riddled with
jargon and ending with “Claim denied” could thus be transformed into a compact, readable
essay.56 The adjudicator would then be freer to live and think exclusively in terms of stylized,
professional reason, rather than imagining a lay audience’s priorities or engaging in what
sometimes travels under the heading of “common sense.” Once again, style would shape
substance, transforming the author and her future rulings.

Finally, AI will reduce the deliberation that stems from interpersonal friction within the writing
process. Judges may personally employ AI tools, circumventing clerks. And even when clerks
are used, they will know that an AI tool stands ready to execute any judicial instruction. So there

51 On the related possibility of an “AI Gideon,” that is, a right to the assistance of artificially
intelligent counsel, see infra note 72 and accompanying text.
52 See, e.g., Hon. Roger Traynor, Some Open Questions on the Work of State Appellate Courts, 24

U. CHI. L. REV. 211, 218 (1957) (“I have not found a better test for the solution of a case than its
articulation in writing, which is thinking at its hardest.”).
53 See supra note 32.
54 Hence the old trope about “building character through hard work.”
55 See supra, Section II.C.
56 See text accompanying supra note 40.

15

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

will be less chance of a key clerk having her own stubborn style or unchecked point of view. 57
The opinion-writing variability that comes from rotating human clerks would dwindle, allowing
each judge’s opinions to become predictable and consistent. With clerk-based friction removed
or reduced, the judge herself will become more accustomed to the seamless execution of her
initial directives—thereby cutting off further deliberation.

Ultimately, AI’s effects on deliberation greatly depend on what judges ask AI to do. If judges ask
the AI to generate the strongest versions of competing views, especially before the judges make
up their minds, then artificial authorship might foster deliberative virtues. Cases might then seem
harder than they initially appeared. But if judges instead make quick, knee-jerk decisions before
asking the AI to implement them, then the deliberative costs might be substantial. And the judges
will often opt for the easier path, given their biases and desire for leisure. Deliberative problems
are only exacerbated when the AI itself favors certain views. Systematic or arbitrary skews in an
AI tool’s recommendations, for instance, can bias the human adjudicator.58

So regulation might be helpful.59 For example, a law governing AI tools might be designed to
require or nudge judges to engage with opposing arguments. Or, to similar effect, judges might
not be allowed to engage strong AI writing tools until they have already grappled with the
opposing arguments set forth through the adversarial system. Advocates, after all, will be coming
to court with their own AI-assisted arguments and writing. Already, some courts impose similar
rules on themselves, such as by postponing opinion drafting until after oral argument. 60 Whether
judges would welcome the imposition of such a regime is, of course, another matter.

E. Reason and Rhetoric

AI tools may be strongest when it comes to the art of rhetoric, and they are bound to get stronger
still. By rhetoric, I mean efforts at persuasion, rather than showing what is actually known based
upon facts and reason.61 Reason itself is often persuasive, but not always—or not as much as
other techniques. If you are trying to get someone to eat a particular cereal, fully explaining its

57 See supra note 50 at 123.

58 See Maurice Jakesch, Advait Bhat, Daniel Buschek, Lior Zalmanson, and Mor Naaman, Co-
writing With Opinionated Language Models Affects Users’ Views, Proceedings of the 2023 CHI
Conference on Human Factors in Computing Systems (2023). For discussion of possible corrective
steps, see Tamara N. Lewis Arredondo, Incorporating ChatGPT into Human Rights Pedagogy and
Research Practices, Opinio Juris (January 2, 2024).
59 For arguments that courts themselves are or will largely be left to deal with the challenges of

AI on their own, without new regulation, see David Freeman Engstrom, The Automated State: A
Realist View, G.W. L. REV. (forthcoming 2024) (manuscript at 30-31); Alicia Solow-Niederman, Do
Cases Generate Bad AI Law?, 5 COL. SCI & TECH. L. REV. 261 (2024).
60 Some courts, by contrast, have a practice of drafting opinions before oral argument. See Daniel

Bussell, Opinions First—Argument Afterwards, 61 UCLA L. REV. 1194 (2014).

61 See ARISTOTLE, RHETORIC 1357a23–33.

16

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

health benefits (even if true) might not get the job done. A celebrity testimonial or catchy jingle
might be far more effective, even if it lacks any rational basis whatsoever. 62

Once reason and rhetoric are teased apart, it quickly becomes clear that they frequently come
into conflict—and not just because someone with both reason and rhetoric on her side might
have to choose which one to pursue. Reason and rhetoric can point in opposite directions—a
distressingly common scenario. In the face of rhetoric, should proponents of reason persist in
arguing from reason? Ought they try to censor rhetoric? 63 Or, perhaps most alarmingly, should
they arm themselves with reason-free rhetoric of their own?

AI will facilitate rhetoric of all types. When it does so simply by making existing reasons more
understandable, the effects are salutary. Clearer explanations, after all, generally foster debate,
refinement, and accountability. But what about implicit appeals to preconceptions, prejudice,
stereotypes, and allegiances? Or pages of easy reading that gloss over critical logical flaws and
legal vulnerabilities? Human writers of course use these techniques today; but, as we have seen,
AI will spread and enhance writing capabilities.

The precise way in which AI generates rhetoric will naturally vary by context and, in some cases,
may allow for techniques that are not presently realistic. In retail or low-level adjudications, for
instance an AI could enable messages that are targeted at the specific individuals in the dispute.
Professor Alicia Solow-Niederman and I have given the following example:

Imagine an AI adjudicator whose “opinions” are leavened with personal touches

informed by instantaneous social media research. After discovering that a losing
party is a Rolling Stones fan, for instance, the AI might comment that “you can’t
always get what you want” and then play the hit song’s refrain. The song’s
aphoristic familiarity might be both emotionally comforting and cognitively
distracting ….64

As this passage illustrates, effective rhetoric can come at the cost of at least two forms of
accuracy relevant to adjudication. First is accuracy in the sense of what actually brought about
the adjudicative result—an issue most relevant to what I have referred to above as authenticity
and accountability.65 Second is accuracy in the sense of whether the result is in fact justified—an
issue I am now associating with reason.

62 See, e.g., Richard F. Yalch, Memory in a Jingle Jungle: Music as a Mnemonic Device in
Communicating Advertising Slogans, 76 J. OF APPLIED PSYCH. 76, 273 (1996).
63 See Kenji Yoshino, The City and the Poet, 114 YALE L.J. 1835, 1839 (2005) (discussing Plato’s

proposed banishment of the poet, along with contemporary implications).

64 Re & Solow-Niederman, supra note 26, at 261.
65 See supra Section II.C.

17

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

In salient or high-level adjudication, AI tools will facilitate feats of rhetoric, as judges produce
grand opinions to impress the public, mollify critics, and increase their supporters’ admiration.
The target audience here is far larger than in most retail adjudication and requires greater finesse,
as the desires of various audiences might be in competition. A great appellate opinion might have
to be legalistic, breezy, funny, distinguished, and authoritative—all at the same time.

AI is likely to be well-suited to this task. For example, I just asked ChatGPT to generate an essay
in favor of the income tax, to revise the essay to be persuasive to libertarians, and then to liven
up the essay with humor. In seconds, the AI accomplished all three of these tasks. Future tools
may be able to access supplemental information, like polling data or trending social media
memes. With such a broad base of training data to draw on, AI would be especially skillful at
playing to different audiences simultaneously. And, in special situations, the AI could tailor its
work to particular audiences, such as specific judges or swaths of the public.

F. Canonicity and Customization

Judicial opinions are hardly a set medium, much less a fixed genre. Common law rulings even
now are largely oral in some nations—and were almost exclusively so until recent centuries
made widespread printing feasible. 66 Early decisions of the U.S. Supreme Court, for instance,
were not formally published by the judiciary. 67 Instead, a private individual became designated
as “court reporter,” transcribed materials, and sold print copies for profit. 68 In recent decades,
internet posting has now essentially supplanted print publication.

Still-newer technologies will allow for dynamic, interactive judicial opinions. There is no need
for a single canonical judicial opinion, after all. Appellate decisions sometimes feature various
quite different rationales put forward in separate opinions by concurring judges. 69 In some
countries, official copies of judicial opinions are released in different languages, and different
readers effectively opt in to one or the other language version. 70 Many courts already publish
relatively concise syllabuses to ease case digestion. 71

In a similar spirit, courts might “publish” a program that interacts with its reader’s preferences,
creating a personalized version or presentation of the relevant judicial opinion. Some readers
may want a version without citations, or one that is written in plain English. Others might want a

66 See Richard J. Lazarus, The (Non)finality of Supreme Court Opinions, 128 HARV. L. REV. 540,
552 (2014); see also infra note 106 (discussing the rule of orality).
67 See Lazarus, supra note 66, at 552.
68 See id. at 582.
69 E.g., Furman v. Georgia, 408 U.S. 238 (1972).
70 For example, Canada’s federal courts sometimes publish in both English and French. Marie-

Ève Hudon, Bilingualism in Canada's Court System: The Role of the Federal Government, available
at https://lop.parl.ca/sites/PublicWebsite/default/en_CA/Research Publications/201733E#a3.1.1.
71 See United States v. Detroit Timber & Lumber Co., 200 U.S. 321, 337 (1906).

18

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

version with jokes and flair, while others prefer a “just the facts” narration. And still others might
want a dialogic version, complete with a visual avatar, that explains the decision through
questions and answers, as a human would in conversation.

In this regime of customization, one might wonder what version—if any—would be canonical,
that is, either legally authoritative or dominant in the public eye? A range of options is available,
roughly tracking existing practices. Perhaps a base text would be treated as judicially
authoritative, whereas derivative, customized versions would be legally irrelevant—much like a
syllabus today. Or different versions could be ranked so that an inferior version is trumped
whenever it came in conflict with another. Especially popular customizations might inform the
meaning or identity of the canonical version. And so on.

III. Reactions by Readers

How will the judiciary’s readers react to artificial authorship? I suggest three basic answers.
While in some sense mutually exclusive, each of these reactions is likely to take place at one
time or another.

A. The Perfect Adversarial System

If AI tools become sufficiently powerful and accessible, then virtually any party could generate a
maximally persuasive brief for any proposition. Those without access to an AI might have one
appointed by the court—effectively yielding a right to an artificial attorney.72 Or, as we have
seen, the court itself might use an AI to generate some of the arguments in favor of one side or
another, yielding a practice somewhat resembling an inquisitorial system.

AI-generated arguments could even prove to be more legally accurate than similar work by
humans. To be sure, the AI would often fall prey to its own biases, including because it could
rely on data that is itself shaped by racism, sexism, classism, and other forms of discrimination. 73
Yet human work, too, is regularly clouded by bias, as well as by self-interest, fatigue, and other
flesh-and-blood limitations. Moreover, courts could test their AIs for bias in ways that would be
infeasible with respect to human judges or clerks.74 We can even imagine parties using AI to
point out the problematic biases in one another’s AI-generated briefs.

72 We could even imagine a constitutional right along these lines, yielding an “AI Gideon.” Cf.
Gideon v. Wainwright, 372 U.S. 335, 339, 343–45 (1963) (providing certain indigent criminal
defendants with a right to appointed counsel). For predictions that AI tools will facilitate access to
justice in various ways, see, e.g., Benjamin Alarie, Anthony Niblett, & Albert H Yoon, How Artificial
Intelligence Will Affect the Practice of Law, 68 U. TORONTO L.J. 106 (2018); SUSSKIND, supra note 39.
73 See SAFIYA NOBLE, ALGORITHMS OF OPPRESSION (2018); Dorothy E. Roberts, Digitizing the

Carceral State, 132 HARV. L. REV. 1695, 1707 (2019).

74 See David Freeman Engstrom, The Automated State: A Realist View, G.W. L. REV. 18-19

(forthcoming 2024); e.g., Hadi Elzayn, Evelyn Smith, Thomas Hertz, Arun Ramesh, Robin Fisher,

19

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

The result could be viewed as a large step toward the perfection of the adversarial system. With
each side always putting its best foot forward, the stronger view would become manifest in court.
And a similar dynamic could arise in the public square. Consistent with the famous “marketplace
of ideas” metaphor,75 the public would be well positioned to assess which advocates and judges
have the better position. No longer would asymmetries in talent or resources cloud the pursuit of
truth and the exercise of reasoned judgment.

And the public, too, would gain greater access to legal reasoning. As we have seen, AI
authorship can facilitate lay readers’ ability to understand, evaluate, and, ultimately, respect the
judiciary’s work. Judges could more easily write for different audiences simultaneously, thereby
hewing to the demands of legal sophisticates without losing touch with broader society. If most
members of society can see for themselves that judges are fairly and accurately applying the law,
then AI authorship would probably enhance the judiciary’s sociological legitimacy.76

The judicial process would adapt accordingly. We might imagine a rule dictating that a certain
type of neutrally validated AI must write a dissent for every judicial opinion, whether or not any
of the actual judges wish to compose one. If the automated dissent were written and circulated
before the court’s decision, this rule would foster deliberation.77 And, if the automated AI dissent
were ultimately published, this reform would also help hold judges accountable in the event that
they are shading the facts or distorting the law.

The value of automated dissent is most evident when there is only a single adjudicator, such as in
federal district court or many administrative proceedings. But it would also be significant for
multi-member appellate courts where dissent is already possible. For one thing, not just majority
judges but dissenters, too, might want to shade or distort the truth in various ways, given their
own biases.78 The best grounds for dissent might be embarrassing for the dissenters due to their
own past writings or desire to align with present-day political trends.

Daniel E. Ho & Jacob Goldin, Measuring and Mitigating Racial Disparities in Tax Audits, 31-39
(2023).
75 See Abrams v. United States, 250 U.S. 616, 630 (1919) (Holmes, J., dissenting); JOHN STUART

MILL, ON LIBERTY 58-59 (1859).

76 For examples of scholarly work that centers the discursive role of judicial opinions, see supra

note 6 (collecting sources).

77 See, e.g., Hon. Ruth Bader Ginsburg, Remarks on Writing Separately, 65 WASH. L. REV. 133,

143 (1990). Similar logic has supported practices like the Sanhedrin’s rule against unanimous
verdicts and the “devil’s advocate” during the Catholic Church’s canonization deliberations.
78 See Patricia M. Wald, The Rhetoric of Results and the Results of Rhetoric: Judicial Writing, 62

U. CHI. L. REV. 1371, 1374 (1995) (reporting that judges compromise to avoid dissents).

20

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

Further, unanimous rulings are often more difficult or problematic than they appear, something
AI-written dissents might reveal.79 Legislatures and regulators might then be in a better position
to react to, or simply override, harmful judicial precedents. More generally, AI dissents could
help the public learn whether the judges are, for lack of a better expression, just making it up.
And that knowledge, whether viewed as desirable or not, could inform efforts at court reform,
either by legitimating or casting doubt on the judiciary’s performance.

Because legal indeterminacy is often viewed as undesirable, 80 the possibility that AI-authored
dissents could both surface and increase it may likewise be considered problematic. Judges might
not want their role in law-creation to be so evident, and onlookers might not want that role to be
so robust. To solve that perceived problem, jurists might begin to rigidify their decisional process
or promote determinate legal principles.81 A proliferation of AI assistance might therefore foster
efforts to alter the law itself, whether by judges, legislators, or others.

I have just sketched a doubly optimistic scenario: the technology not only works well, but it also
interacts favorably with social practices. It is worth pausing to note the key premise underlying
that happy outcome—namely, the assumption that optimal arguments will surface the true state
of the law, whether that be a right answer or the lack of any such answer. But a clash of optimal
arguments may leave different readers with equal and opposite confidence regarding the right
way to resolve a case. Perhaps readers would be persuaded by whichever opinion they happen to
have read first—or last. This kind of worry leads to the scenarios that follow.

B. Artificially Balkanized Readership

The foregoing section generally assumed that maximally persuasive opinions would tend to
produce homogeneous reactions among readers. That is, the relevant set of readers would more
or less all agree that one side or the other had the better of the exchange, or else that there simply
was no clear winner to choose. That prospect of consensus is what makes it possible for a perfect
adversarial system to draw out the truth, or at least the best available answer. And, in many
situations, that consensus could indeed be achieved.

But the prospect of legal indeterminacy tees up another possibility—namely, that different
groups of readers would have highly divergent reactions to the same opinion. There might be

79 See id.
80 Indeterminacy certainly has its virtues, too. See generally HRAFN ASGEIRSSON, THE NATURE
AND VALUE OF VAGUENESS IN THE LAW (2020).
81 Cf. Re & Solow-Niederman, supra note 26, at 288-89 (arguing that AI creates certain

incentives to increase legal determinacy); David Freeman Engstrom and Nora Freeman Engstrom,
Legal Tech and the Litigation Playing Field, in LEGAL TECH AND THE FUTURE OF CIVIL JUSTICE 150–
51 (2023) (Engstrom, ed.) (skeptically assessing the prospects of “recalibrating substantive law” to
solve a litigation playing tilted by technological asymmetries).

21

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

several maximally persuasive opinions available, each pitched to different constituencies. Or the
one uniquely superior opinion might be highly partisan in content and appeal. AI authors might
be especially adept at exploiting these sorts of fissures in the legal community or the public at
large, even without being specifically asked or designed to do so.

AI tools could therefore reflect or even compound reader prejudices at the expense of truth.
Biases aren’t evenly distributed across an opinion’s readership. So the most persuasive opinion
either expanding or constricting gun rights, for instance, might be designed exclusively for
conservative or liberal readers, respectively. And there may not be anything that the other side
could do to win over those readers, particularly when their priors are being stoked by writing
generated by an AI.

On this view, the rise of AI authorship, far from being a perfect adversarial system or truth
finding machine, will be a source of sophistry. This is rhetoric of the deceptive sort that Plato
warned against, the kind that brings about many false beliefs.82 Even worse, the rhetoric in
question would be balkanizing, in that it would increase the confidence of warring factions even
as it also encouraged them to adopt new partisan views.83 The shared beliefs that are traditionally
thought necessary to have a legal system might be put under strain.

The most straightforward solution is simply to require that the AI not play to partisan prejudices
or other biases. That is, the AI might be asked or trained to avoid playing to groups and instead
to write only for a legalistic audience or someone with middle-of-the-road political views.84
Because the judges have an incentive to garner whatever support is available, a requirement of
this sort might have to be imposed on the judiciary. In essence, some forms of persuasiveness
might be ruled out of bounds, at least when AI is concerned.

But that remedy leads to difficult questions about the proper means of regulating adjudication.
Shaping AI tools depends on judgments about the proper goals of judicial writing. And those
goals are highly contested and varied across judges. Just think of Justice Scalia’s norm-busting
transformation of judicial writing. 85 Moreover, forcing the AI to favor consensus will almost
necessarily come at the cost of maximal persuasiveness. Sterile rhetoric and on-the-one-hand
arguments might come at the expense of creativity, decisiveness, and zest.

82 See PLATO, SOPHIST 233. Self-interest of course plays a key role, as people and institution often
have an incentive to pretend to knowledge.
83 Cf. CASS R. SUNSTEIN, GOING TO EXTREMES: HOW LIKE MINDS UNITE AND DIVIDE (2011).
84 Relatedly, courts might try to strengthen norms against rhetoric—though the difficulty of

telling reason from rhetoric may undermine this effort. See Nina Varsava, Professional
Irresponsibility and Judicial Opinions, 5 HOUSTON L. REV. 101 (2021).
85 See, e.g., JUSTICE SCALIA: RHETORIC AND THE RULE OF LAW 1 (eds. Brian G. Slocum and Francis

J. Mootz III) (2019).

22

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

Perhaps regulating the AI tools available to courts would be legitimate for reasons akin to
existing legislation constraining who can serve as a clerk. 86 Yet existing regulations of clerks
tend not to focus on their reasoning qualities or legal views. Regulation confining the uses or
nature of AI tools might thus resemble unprecedented rules, like a ban on hiring any clerk with a
sharp wit or a record of criticizing the political branches. And a prohibition like that would
probably encroach on judicial independence and deliberation.

Moreover, efforts to regulate AI might simply be ineffectual. Various rules currently purport to
restrain judges’ use of private email or their ability to learn classified information that has been
leaked to the press and widely published.87 Yet the efficacy of those restrictions is easily called
into doubt, because the relevant technology is so pervasive. Similarly, AI writing tools may soon
be so ubiquitous as to escape government control.

C. Rhetoric’s Rise – and Reason’s Demise

There is at least one other potential response to a proliferation of reasoning and rhetoric brought
on by AI authorship: ignore it. If powerful writing becomes ubiquitous, it might stop being quite
so powerful. Unable to tell sound reasoning from persuasive rhetoric, readers might stop caring
to read at all, preferring instead to evaluate some or even all legal questions based on other
qualities. The case’s outcome, rule, or author might matter, as contrasted when the reasoning put
forward by any judge.

Several factors might conspire to bring this counterintuitive result into reality. One is the effect
of legal indeterminacy. What if it turns out that indeterminacy is nearly everywhere? 88 And
necessarily so. Consistent with legal realism, critical legal studies, and the Priest-Klein
hypothesis,89 contested cases, particularly at the appellate level, may almost always be toss ups.
Why else would they be litigated? AI authorship could surface that legal uncertainty, thereby
revealing the true, flimsy state of the law to professionals and the public alike.

86 For example, federal clerkships are limited to U.S. citizens and certain lawful residents.
United States Courts, Citizenship Requirements for Employment in the Judiciary, available at
https-//www.uscourts.gov/careers/search-judiciary-jobs/citizenship-requirements-employment-
judiciary (last visited June 21, 2024).
87 See Lauren Aratani, US Supreme Court Justices Use Personal Emails for Work, Report Says,

GUARDIAN (Feb 4, 2023).

88 As fleshed out in the main text, my basic claim here can be framed in terms of either

metaphysical indeterminacy (that is, an actual absence of a legally correct answer) or epistemic
indeterminacy (that is, a lack of ascertainable knowledge regarding the correct answer).
89 See George L. Priest & Benjamin Klein, The Selection of Disputes for Litigation, 13 LEG. S TUD.

1, 4-5 (1984); Mark Tushnet, Following the Rules Laid Down: A Critique of Interpretivism and
Neutral Principles, 96 HARV. L. REV. 781, 818 (1983); JEROME FRANK, COURTS ON TRIAL: MYTH AND
REALITY IN AMERICAN JUSTICE 74 (1949).

23

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

Even more unsettling is the prospect that AI authorship will generate so much rhetoric that it
becomes difficult or impossible to discern which side is correct. One version of this worry would
focus on writing quantity. Judges and other readers might be inundated with so much purported
reasoning—thousands of opinions, endless motions or briefs, and so forth—that there is no time
to grapple with many important matters. 90 More fundamentally, strong AI writing could
transform what presently seem like easy answers into head-scratching tossups. In many actual
cases, for instance, one side might have the benefit of common sense as well as a straightforward
legal argument. At present, that sort of “easy” case would likely generate a quick and unanimous
outcome. But, with the benefit of AI writing, the opposing side might be able to generate
sophisticated counterarguments that confound or beguile the court.

And by turning many easy cases into hard ones and right outcomes into wrong ones, AI
authorship would increase the effective scope of legal indeterminacy. Again calling sophistry to
mind, AI authorship would undermine the search for truth. We could even imagine that virtually
every contested case features two equally plausible opinions. Readers, then, would quickly learn
that it is a waste of time to look to AI opinions for guidance. What they would find there is
nothing more than persuasive pablum.

To a great extent, of course, legal and popular culture are already quite skeptical of judicial
opinions. Almost everyone these days is a legal realist to some extent, and a multitude of legal
commentators stands ready to inveigh against the courts at any moment. Perhaps oracular judges
in prior eras could credibly claim to be doing legal science or exhibiting profound sagacity, but
those days are long gone.91 Some people do withhold judgment until they “read the opinion,” as
Justice Barrett has implored. 92 But the very fact that she had to make that plea suggests that
many people, probably most, do not.

AI authorship could take legal culture several steps farther along this path. Today, judicial
dissent and media engagement both focus on a relatively small sample of all cases—especially
ones with political salience. That limited focus stems partly from resource constraints. It takes
time and talent for a human author to digest and debunk a legal argument. And, at first blush,
judicial opinions often appear plausible and well-reasoned. Legal culture accordingly operates on
the assumption that courts generally engage in sound reasoning.93

90 AI itself might offer a cure here, insofar as AI reading assistants could boil down voluminous
prose. Confronted with stacks of AI-generated briefs, the judge might ask her own AI to pick out the
facts and arguments that the judge is likely to view as most important—and even compose a draft
opinion based on them. In a rhetorical arms race, then, AI readers might effectively counteract AI
writers. This culling process would introduce even more separation between human authors and the
texts that humans read.
91 See Oliver Wendell Holmes, The Path of the Law, 10 HARV. L. REV. 457 (1897).
92 See Justice Amy Coney Barrett, Remarks at the Reagan Library (April 4, 2022).
93 Of course, most opinions find few if any readers. But the opinions that are read—whether by

the parties, lawyers arguing the next case, or students perusing their casebooks—sustain the

24

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

AI authorship would pose new challenges to that basic assumption. For one thing, automated
dissents and other forms of AI authorship could puncture the aura of authority that presently
accompanies unanimous, business-like rulings.94 Commentators often point out how many
appellate rulings, such as at the Supreme Court, are unanimous. 95 And even divided rulings often
feature partial or muted dissent. Critical writings generated or facilitated by AI could render
those statistics, and the impression of legitimacy they convey, obsolete.

Judicial rulings would also lose the authority that comes from inscrutability. Legal jargon and
technicalities can impress an audience. 96 Other times, they make it hard for the reader to engage
meaningfully with whatever the adjudicator is asserting. AI will greatly reduce both of those
effects. Not only would judicial writing instantly be translated into understandable prose, but an
AI tool could also answer follow-up questions from the reader. 97

And then there is the knowledge that an AI probably revised or shadow-authored what the judges
have published. By comparison, the public’s awareness that a judge often works with clerks
(effectively, mini-judges98) does not challenge the fundamental idea that court personnel are
expert—and special. With the AI’s help, however, most anyone might endeavor to write their
own faux judicial opinions. And those opinions, though generated by amateurs, could meet
almost any desired standard of professional competence.

Skeptics of judicial authority routinely lament that courts are making it up, rather than enforcing
predetermined legal norms.99 But those sorts of critics often struggle to convince their audiences
of how judging “really” works. In a world of AI authorship, by contrast, lay readers might
immediately understand that a judge could simply ask an AI to generate persuasive arguments
for virtually any conclusion, at least in most contested cases. After all, the lay public would itself
often be using AI in much the same way as part of their daily lives. What could more thoroughly

general assumption that courts trade in legal reason. Cf. Douglas H. Ginsburg, Remarks Upon
Receiving the Lifetime Service Award of the Georgetown Federalist Society Chapter, 10 GEO. J.L. &
PUB. POL’Y 1, 9 (April 26, 2011) (“When I was new on the court, my colleague . . . told me, tongue in
cheek, of course, ‘Remember, the only people who read these opinions are the winning lawyer, the
losing lawyer, and the winning lawyer’s mother.’”).
94 See text accompanying supra note 77.
95 See, e.g., Nora Donnelly and Ethan Leib, The Supreme Court Is Not as Politicized as You May

Think, N.Y. TIMES (Oct. 8, 2023); Devin Dwyer, Supreme Court Defies Critics With Wave of
Unanimous Decisions, ABC News (June 29, 2021).
96 See Richard A. Posner, Judges’ Writing Styles (And Do They Matter?), 62 U. CHI. L. REV. 1421,

(1995).
97 See supra Section II.F.
98 See Alvin A. Rubin, Views From the Lower Court, 23 UCLA L. REV. 448, 456 (1976) (discussing

judicial law clerks as “para-judges”).

99 This sort of accusation has long—perhaps always—been a staple of anti-court rhetoric,

whether launched from the political left or the right. See, e.g., ROBERT BORK, TEMPTING OF AMERICA
(1990); ERWIN CHEMERINSKY, THE CONSERVATIVE ASSAULT ON THE CONSTITUTION (2010).

25

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

demystify the courts? Everyone would know not merely that the judges are making it up, but that
the AIs are making it up for the judges.

What would remain are non-rhetorical proxies for desirable legal outcomes. Having been
disillusioned about the nature of judges’ work, the lay public and sophisticates alike would
generally evaluate case outcomes based on factors unrelated to the persuasive content of any
published legal analysis: who voted for the result, what interest groups applaud it, and what do
trusted authorities have to say about the ruling’s likely consequences? A judiciary operating in
that environment would resemble a legislature engaged in policymaking.

AI authorship thus draws attention to an important premise of real-world legal systems. The
existence of legal norms and elites has always depended on there being a scarcity of persuasive
resources and arguments. It takes time for lawyers to be trained, their talents honed, and their
arguments crafted for each new case. That basic reality means that it is expensive or infeasible to
litigate in the teeth of straightforward and accessible law—and costly to litigate even when the
law is unclear.100 By undermining these constraints, a surfeit of persuasiveness threatens the
effectiveness of legal norms.101 The question then arises: would (or should) legal actors try to
restore an economy of persuasive ability?

IV. Seeking Equilibrium

We have discussed the initial ways in which AI authorship will alter judicial craft and affect
readers. But the complexity of the legal system fosters dynamic change, featuring interactions
that are iterative, cross-cutting, or reinforcing. Judges will quickly anticipate or experience
reader reactions. How, at this third stage of the dialectic, will the judiciary account for reader
reactions to AI authorship?

A. Transitions and Tradition

The judiciary has a lot to lose from the long-term trends that AI authorship is setting in motion.
For reasons we have seen, judicial authority may be undermined by an endless stream of
rhetorically effective challenges to their rulings. In addition, the judiciary may become
demystified as lay persons realize that they, too, can understand, criticize, and even author
sophisticated legal opinions—all with AI assistance. Finally, AI assistance could invite novel
forms of regulation, such as restrictions on how an AI assistant is trained. 102

100 See Frederick Schauer, Easy Cases, 58 S. CAL. REV. 399, 401 (1985).
101 For different suggestions that the advent of AI might render law as we know it obsolete, see
Wu, supra note 26; Anthony J. Casey & Anthony Niblett, The Death of Rules and Standards, 92 IND.
L.J. 1401 (2016).
102 Alicia Solow-Niederman, Administering Artificial Intelligence, 93 S. CAL. L. REV. 633, 690

(2020).

26

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

Because these imagined effects are reactive, they would take longer to materialize than the initial
effects discussed earlier. We must therefore pay special attention here to issues of temporality
and to the development path that AI authorship is presently following. 103 For example, additional
technological changes—like the invention of general artificial intelligence—could overwhelm
the effects of AI authorship, rendering human judging altogether obsolete. 104 One implication of
that contingency is that we have to downgrade our predictive confidence.

Even more importantly, the existence of a development path creates an opportunity for interested
parties to resist and shape these relatively long-term effects. So even if AI authorship is bound to
become ubiquitous, for many years judges might eschew, or regulations might successfully
block, its use. And a concerted effort to swear off AI tools could generate institutional dynamics
that entrench traditional writing methods, even as AI authorship elsewhere prevails.

Imagine this: motivated largely by the anxiety that often accompanies new technologies, 105
judges disavow use of certain AI tools. And, to make their disavowal credible, judges may even
return to historical practices of judicial decision-making, such as ruling orally from the bench
after hearing public arguments with no AI involvement.106 This sort of policy would afford
human judges the legitimacy that comes from authorship. And it would insulate the courts from
regulatory interference. While the public debates how to train, limit, and monitor AI tools, judges
can remain detached. Rather than becoming exhibits in public controversy, they can pass
judgment on novel regulations created elsewhere.

Once controversies had subsided and AI tools became uncontroversial and standardized, judges
might reconsider their choice to abstain from using them. By then, new judges would have spent
much or all of their lives with these tools, both lay and sophisticate audiences wouldn’t be fazed
in the slightest by the knowledge of their use, and the case for intrusive regulations would have
subsided. Judicial avoidance of AI authorship would be temporary but consequential, essentially
allowing the least powerful branch to empower itself.

Or perhaps not. Instead of being sloughed off, the ideal of human authorship could become a
permanent, self-conscious component of judges’ professional ethic. For example, judges might
have to disclose when and how they use AI tools; or certain uses of AI tools might simply be

103 See Re & Solow-Niederman, supra note 26, at 10 (discussing development paths).
104 See supra note 18 (collecting sources).
105 See generally CALESTOUS JUMA, INNOVATION AND ITS E NEMIES: WHY PEOPLE RESIST NEW

TECHNOLOGIES (2016). Impulses toward technological panic are, paradoxically, also counteracted by
tendencies toward technological utopianism and trust in “science.”
106 This approach might be cast as a new rule of orality. See ROBERT J. MARTINEAU, APPELLATE

JUSTICE IN ENGLAND AND THE UNITED STATES: A COMPARATIVE ANALYSIS 102–03 (1990) (discussing a
traditional rule demanding orality, which arguably fostered public scrutiny of judicial work).

27

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

prohibited, whether via statute, court rules, or codes of professional conduct.107 These reforms
would create a space in which exclusively human talent and decision-making would not only
abide but remain publicly identifiable. That sort of gambit could preserve a sense of respect or
awe for human judges and the work they continue to perform.108 Likewise with many other
intellectual endeavors that have been mastered by computers. For instance, chess players know
that computers can beat any human.109 Yet we still marvel at the feats of grandmasters. A
twenty-first century Cardozo might be regarded similarly. 110

This sort of reform would recreate a degree of scarcity with respect to persuasive resources and
arguments. That is, a social norm (or set of norms) would check the unwanted productive
capabilities unleashed by rhetorical technologies.111 And that development would halt or slow
the social processes associated with both the potential demise of legal reasoning and (relatedly)
the decline in judiciary’s prestige. In this way, the interests of judges tend to align with the
preservation of traditional legal norms and practices.

B. The Rhetoric of Wisdom

But if judges are to preserve traditional legal norms and practices, they will have to invoke more
than their own self-interest. The role and status of human judges seems bound to decline, unless
there remains demand for something that AI assertedly cannot offer. 112 Human judges
determined to retain their station therefore have a strong interest in identifying the unique
qualities that they bring to the table. What can they point toward?

One option is to broaden the range of relevant considerations. So far, I have focused on just two
dimensions of opinion-writing success: rhetorical persuasiveness and legal correctness. But
judges might observe that other criteria are available. Moral rectitude, for example. AI tools that

107 Some courts have already undertaken reforms like the ones identified in the main text. See,
e.g., STATEMENT OF PRINCIPLES FOR THE NEW JERSEY JUDICIARY’S ONGOING USE OF ARTIFICIAL
INTELLIGENCE, INCLUDING GENERATIVE ARTIFICIAL INTELLIGENCE, available at
https://www.njcourts.gov/sites/default/files/ courts/supreme/statement-ai.pdf?cb=bb093263 (Jan. 23,
2024) (“Judges and their staff may use AI only for select purposes, such as for preliminary gathering
and organization of information.”).
108 See IAN M. BANKS, LOOK TO WINDWARD 319 (2000).
109 Students of chess routinely watch computers give devastating real-time assessments of

humanity’s greatest players.

110 Cardozo is often viewed as a master of judicial rhetoric, even by his critics, see generally

RICHARD POSNER, CARDOZO: A STUDY IN REPUTATION 125-43 (1990). Could a well-tuned AI tool one
day recreate Cardozo’s style—making it available to one and all?
111 See generally Ryan Calo, The Scale and the Reactor, SSRN at 5, 9 (2022), available at

https://dx.doi.org/10.2139/ssrn.4079851 (arguing for the social contingency of seemingly inevitable

technological change).
112 Cf. FRANK PASQUALE, NEW LAWS OF ROBOTICS: DEFENDING HUMAN EXPERTISE IN THE AGE OF

AI 229 (2020) (“As automation advances, we must now [adopt] a commitment to ‘a rule of persons,
not machines.’”).

28

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

aim at persuasiveness or lawfulness might tend toward popular moral views, but morality is often
thought to be independent of, and possibly contrary to, popular opinion. Pragmatic virtues are
similarly beyond an AI’s expertise. The AI tool might be great at making proposals sound
practicable. But would its prescriptions actually be sound?

These additional criteria for success offer alternative ways of criticizing AI tools, and some of
these criteria are more objectively ascertainable (at least in hindsight) than notions of morality or
pragmatism. For instance, courts are often thought to care whether their rulings will promote the
judiciary’s long-term public legitimacy.113 This criterion involves a factual prediction. Yet there
is little reason to think that imminent technologies can offer reliable guidance here. The tools
simply are not trained on these sorts of empirical prognoses.

So there is a large and multifaceted category of opinion-writing virtues other than either legal
correctness or persuasiveness, and those various ideas might be collected under the heading of
“wisdom.”114 This category represents a potential limitation on the effectiveness and appeal of
AI tools. Though these tools are already designed to appear wise, they are not in fact wise, or at
least not reliably so. 115 Any wisdom they exhibit is incidental to their pursuit of what amounts to
persuasive writing, or writing of the style that has been requested.116

Judges might also point out that the idea of persuasiveness is itself complex. Foreseeable AI
tools will be much better at writing what is regarded as a good opinion today, rather than
predicting what will be most persuasive or laudable years into the future. When wise judges
decide cases, they sometimes aim in part to change popular views and practices, rather than
simply appealing to them.117 AI tools trained on existing materials can certainly facilitate that
effort, but—unless guided by humans—they will usually neglect it.

113 See, e.g., Dobbs v. Jackson Women’s Health Org., 142 S. Ct. 2228, 2278 (2022) (“[I] is
important for the public to perceive that our decisions are based on principle, and we should make
every effort to achieve that objective by issuing opinions that carefully show how a proper
understanding of the law leads to the results we reach.”).
114 See Norman W. Spaulding, Is Human Judgment Necessary? Artificial Intelligence,

Algorithmic Governance, and the Law, in THE OXFORD HANDBOOK OF ETHICS OF AI 75, 378, 397, 399
(2020) (Markus D. Dubber, Frank Pasquale & Sunit Das, eds.); see also John Tasioulas, Ethics of
Artificial Intelligence: What It Is and Why We Need It (Oct. 2023); Cass R. Sunstein, The Use of
Algorithms in Society, REV. AUSTRIAN ECON. (2023).
115 This is almost Plato’s definition of sophistry. See supra note 82.
116 As we have seen, however, human use of AI tools can foster the attainment of wisdom by

enhancing deliberation. See text accompanying supra note 50.

117 The canonical example is Brown v. Board of Education, 347 U.S. 483 (1954), which took a side

in a matter of public controversy and was later vindicated by public opinion. Of course, malign
judges, too, can seek to change their society rather than persuade it.

29

Electronic copy available at: https://ssrn.com/abstract=4696643

 Draft

Yet wisdom’s appeal cannot supply an airtight case in favor of human judges, since humans, too,
are often unwise.118 Even expert jurists can be motivated primarily by self-interest, fads, and
bias. They may believe they are being wise when, in fact, they are self-deluded. And one might
think that using AI to appeal to the considered popular views of today is a more reliable path to
wisdom than catering to the distant judgments of history. For these reasons, the value of wisdom
may itself support greater use of AI authorship .

Ultimately, attitudes toward AI tools will be informed by a complex matrix of competing

influences. Yes, legal actors will be drawn toward persuasiveness, self-interest, and their own
biases, but they will also care about legal correctness, prudence, and morality. Judges, moreover,
will want to garner professional esteem, avoid becoming objects of regulation, and—perhaps
most of all—justify their own human involvement in the legal system. Recognizing these
competing interests, judges and other defenders of traditional legal techniques may be tempted
invoke wisdom strategically, to defend their institutional position. The result would be a rhetoric
of wisdom, one capable of displacing wisdom itself.

Conclusion

AI authorship won’t be limited to courts, and many of the tradeoffs and dynamics that arise in
connection with judges will find parallels elsewhere. If technological optimists are correct, then
rhetorical craft will crowd out reason, skill levels will quickly become flattened, and human
professionals will struggle to preserve the roles long allotted them. True, courts are special. In
other domains, authenticity may be less important, reason may more easily be separated from
rhetoric, and incumbent professionals may be less able to assert that their role is essentially
human. Even so, the present study can be viewed as a model or point of departure when
analyzing broader social trends.

118 See Volokh, supra note 18, at 1139 n.12 (“The question is never whether a proposed computer
solution is imperfect; it’s whether it’s good enough compared to the alternative.”).

30

Electronic copy available at: https://ssrn.com/abstract=4696643