AIは自動的に法律サービスを安くしない

This essay was published in Lawfare’s Research Paper Series. The official version that should be cited is linked here.

The essay is co-authored with Justin Curl, a third-year at Harvard Law School. Previously, he was a Schwarzman Scholar at Tsinghua University and earned a degree in Computer Science from Princeton, where we first collaborated. You can find more of his writing on AI and the law here.

Many AI leaders believe the technology will transform knowledge work. OpenAI CEO Sam Altman predicts AI systems that are “smarter than humans by 2030,”1 while Anthropic CEO Dario Amodei analogizes future AI models to a “country of geniuses in a data center.”2

Researchers identify legal services as especially vulnerable to disruption by AI.3 And since GPT-4 passed the bar exam,4 much of the profession seems to agree. Law schools have begun incorporating AI into their curricula5 and partnering with AI-focused legal-tech companies to prepare future lawyers for a changing profession.6 One prominent lawyer has argued AI can already replace law clerks and oral argument.7 Another predicts AI could “replace traditional lawyers by 2035.”8

This excitement about AI comes at a time when legal services are expensive. Millions of individuals are priced out of legal assistance,9 while corporate legal fees are increasing steadily, with hourly rates for partners at large law firms now exceeding $2,300.10 Unsurprisingly, many observers see the potential for AI to make legal services more accessible by delivering outcomes at lower costs.11

Our central claim is that advanced AI will not, by default, help consumers achieve their desired legal outcomes at lower costs. We examine the bottlenecks12 that stand between AI capability advances and the positive transformation of the practice of law that some envision. For AI to usher in a world of abundant legal services, the profession must address three bottlenecks: regulatory barriers, adversarial dynamics, and human involvement.

First, unauthorized practice of law (UPL) regulations may limit AI use by consumers (and to some extent lawyers). These laws prohibit nonlawyers from performing legal work.13 Individuals and organizations can face steep fines and criminal liability if courts conclude their systems cross into practicing law,14 forcing would-be providers to either limit their AI tools’ functionality in legal domains or risk enforcement actions. Entity-based regulations—which restrict who can own equity in businesses that provide legal services—restrict how legal services are offered, again limiting how AI is used by lawyers and consumers. Without reforms, if consumers cannot access AI capabilities or lawyers are not incentivized to use AI well, AI will not help people accomplish their legal goals, regardless of how advanced it becomes.

Second, even if AI is effectively and widely adopted, the American15 legal system’s adversarial structure can prevent advanced AI from lowering the cost of achieving clients’ outcomes.16 Because legal outcomes often depend on relative rather than absolute quality, when both parties become more productive, the competitive equilibrium simply shifts upward.17 In a world with advanced AI, achieving the same result—like settling favorably or prevailing at trial—would require a greater quantity and quality of legal work. So even as productivity increases and cost per legal task falls, parties are locked into an arms race of increasing amounts of legal work required to reach the same outcome.

As a historical analogy, digitization could have reduced discovery costs by making document review much easier.18 But litigators operating within litigation’s adversarial framework exploited the surge in digital documents to drive up costs for their opponents, leaving total litigation costs high.19

Though less explicitly adversarial, transactional work (like contract negotiation) can exhibit similar dynamics: Lawyers compete to control disclosures and outmaneuver opposing counsel when drafting and negotiating agreements.20 Of course, some legal outcomes (like effective estate planning) do not depend on adversarial processes, and this bottleneck would not apply to them.

The third and final bottleneck we discuss is human involvement in legal work. In a world where AI gains outpace increases in the volume of legal work, our desire for human beings to adjudicate cases and understand the contracts they sign is a final bottleneck. In litigation, if AI enables a flood of legal work, judges will likely respond by taking longer to resolve disputes (delaying outcomes) or delegate more to assistants (lowering adjudication quality).21 And with transactional work, even if AI is drafting entire contracts, human lawyers will still need time to understand what these provisions mean for an organization’s interests. The speed of human decision-makers (whether judges, lawyers, or clients) places an upper limit on how much AI can accelerate legal processes without sacrificing human involvement. See Figure 1 below for a visual description of the three bottlenecks and our argument.

Figure 1: The bottlenecks between advanced AI capabilities and the positive transformation of the practice of law

This report applies the “AI as Normal Technology” framework to a specific domain: the legal industry.22 This framework is fundamentally about agency: Rather than treating AI’s trajectory as predetermined by capability advances, it directs attention to the social and organizational bottlenecks between what AI can do and the impact it has on the world. In our analysis of the practice of law, diffusion will likely be slow. Better models have not yet translated into more reliable legal products because adapting workflows to leverage AI and teaching users takes time.23

We argue that the end state of AI diffusion can look very different depending on the institutional response. Some pathways lead to genuine improvements in access and efficiency, while others simply make producing legal work (outputs) cheaper without making it easier to achieve the results clients want (outcomes). For AI to deliver better outcomes for consumers, the legal industry must enact reforms addressing the bottlenecks. Otherwise, we risk a future in which legal work becomes more abundant, but legal outcomes remain expensive and inaccessible.

We proceed in three sections. The first section explains why legal services are so expensive. The second section aims to convince readers that AI won’t automatically deliver legal outcomes at lower costs. And the third section offers recommendations for addressing these bottlenecks based on existing proposals for legal reform and illustrates how drastically AI’s impact could differ based on the legal profession’s response.

Table of Contents

Why Legal Services Are So Expensive

Why AI Won’t Help by Default

Regulatory Barriers

Adversarial Dynamics

Human Oversight

Institutional Reforms

Reforming Professional Regulation

Reforming Adjudication

The Evolving Role of Lawyers

Conclusion

Why Legal Services Are So Expensive

Three structural factors help explain why legal services are so expensive:24 Evaluating their quality is difficult, their value is often relative, and professional regulations limit competition from alternative business models.

First, unlike a meal at a restaurant, where it’s easy to assess quality, legal services are “credence goods,” which means their quality is difficult to evaluate even with hindsight.25 The final outcome in a case reflects the cumulative effect of many smaller decisions, so it can be very hard, even for other lawyers, to evaluate whether legal services were provided effectively.26 How clear was the law on that issue? Did the client reach the desired outcome because of or in spite of the lawyer’s skill? Which decisions actually contributed to that success? This evaluation difficulty forces consumers to rely on proxies for quality (e.g., the prestige of a lawyer’s law school or judicial clerkships) when choosing between legal service providers, making it hard for traditional market mechanisms to work.

Second, the value of legal services is relative.27 Because “the American litigation system is a thoroughgoing adversarial” one, what matters to the lawsuit’s outcome often isn’t how good your lawyers are in absolute terms, but whether they’re better than the other side’s.28 Other kinds of legal work, such as drafting contracts and agreements (often called transactional work), can also be adversarial as lawyers try to “outfox” each other in terms of what is disclosed in negotiations and the language of a contract itself.29

These dynamics have kick-started an arms race for legal talent, driving up costs at the top end of the market, which serves corporate clients and is often called “BigLaw.”30 In 2024, the median partner at large law firms charged $1,050 per hour, with some commanding over $2,300.31 That’s up 5.1 percent from 2023, which was itself up 5.4 percent from 2022.32 Fortune 200 companies reported that their average litigation costs in cases exceeding $250,000 in legal fees had nearly doubled over eight years, climbing from $66 million per company in 2000 to $115 million in 2008.33 In the patent field, a 2017 survey found that patent cases worth less than $1 million typically cost $1 million to litigate ($500,000 per side).34

Third, the profession’s regulatory framework, designed with consumer protection in mind, has created its own complications. Two types of regulations are often the focus of reform: unauthorized practice of law (UPL) and law firm ownership regulations.35

UPL laws make it illegal (in some jurisdictions a felony) for unlicensed attorneys to apply legal knowledge to specific circumstances.36 An unfortunate effect is to make it more expensive to offer basic legal assistance in contexts requiring little legal expertise.

Most states have regulations limiting who may share in legal fees. Gillian Hadfield argues that these rules promote a business model that creates inefficiency for small firms.37 These firms serve individuals and small businesses and are sometimes called the “PeopleLaw” sector.38 She cites a 2017 Clio study of forty thousand customers: In an average eight-hour workday, lawyers engaged in billable work for only 2.3 hours, billed 1.9 hours, and collected payment for just 1.6 hours.39 So although clients paid an average of $260 per hour, lawyers effectively received $25–40 per hour because the rest of their time was spent finding clients, managing administrative tasks, and collecting payments. These regulations require lawyers to serve clients through partnerships fully owned and financed by lawyers. They deter alternative models that involve large-scale businesses with centralized billing, customer service, marketing, and administrative functions, which could leverage economies of scale to deliver legal services at $30–50 per hour instead of $260.40

Importantly, none of the sources of market dysfunction are intrinsic to legal services. They reflect choices about procedure, pricing, and professional governance. While reform may be politically difficult or costly, the outlook is dim without it. And contrary to what some might hope, AI will not automatically make legal services cheaper, as we discuss next.

Why AI Won’t Help by Default

Regulatory Barriers

More legal assistance would be valuable in the debt collection context. From 1993 to 2013, the number of debt collection lawsuits grew from 1.7 million to about 4 million.41 In Michigan, these lawsuits made up 37 percent of all civil district court case filings by 2019.42 The trend is similar in Texas: “Debt claims more than doubled from 2014 to 2018, accounting for 30% of the state’s civil caseload by the end of that five-year period.”43 More than 70 percent of debt collection defendants lose by default for failing to respond, even though many cases are “meritless suits” and responding is not complicated.44

New York has created a form for responding to debt collection lawsuits by checking some boxes.45 This form, however, includes questions difficult for nonlawyers to understand, such as whether someone would like to invoke the doctrine of “laches.” Recognizing this difficulty, the nonprofit Upsolve began training volunteers to offer assistance. Concerned that this might violate New York’s UPL rules, Upsolve sought an injunction declaring this basic assistance was protected by the First Amendment. A federal judge agreed.46 But New York appealed to the U.S. Court of Appeals for the Second Circuit, which invalidated the injunction, concluding that the lower court applied the wrong First Amendment test (so Upsolve was no longer protected).47

It’s easy to see how an AI system could help here. A nonprofit could provide access to a tool customized for debt collection suits. Or individuals could directly ask general-purpose tools like ChatGPT, Claude, or Gemini for relevant information. Despite this potential, organizations risk violating UPL laws whenever their AI tools complete “tasks that require legal judgment or expertise.”48 The New York Bar Association, concerned that the shortcomings of current AI models would harm consumers, has warned that “AI-powered chat bots now hover on the line of unauthorized practice of law.”49 While some legal researchers disagree because AI is not a “person” capable of exercising legal judgment50 or because AI systems simply “provide information to users, similar to paper guides about court procedure,”51 all authors cited in this paragraph agree that the status of AI tools under UPL laws is currently unclear.

LegalZoom’s history of lawsuits illustrates how UPL regulations can deter innovation in the delivery of legal services.52 LegalZoom automates rote tasks like preparing documents for trademark filings and has been plagued by UPL lawsuits for years.53 In 2011, private individuals who had purchased the company’s services sued in Missouri, alleging LegalZoom was engaged in UPL because it claimed to “take[] over once a consumer answer[ed] a few simple online questions.”54 After the court denied its motion to dismiss the case, LegalZoom agreed to compensate plaintiffs and modify its business model. In 2015, the North Carolina State Bar won a consent judgment requiring LegalZoom to conform to certain conditions.55 Trademark lawyers in California advanced similar theories in a 2017 suit against LegalZoom’s trademark-filing product.56 And in 2024, a New Jersey plaintiff brought a class action alleging UPL violations.57

While AI’s legality remains in doubt, the threat of UPL liability can inhibit its adoption. Without reform, developers risk fines and criminal liability if their AI systems provide legal advice. Organizations may simply be unwilling to provide access to users, especially those who cannot afford to compensate a developer for the risk of UPL liability. Separately, entity regulations that restrict financing for AI legal startups can deter the kinds of operational experimentation helpful for delivering legal services at lower costs. Overall, if regulatory barriers prevent consumers from effectively accessing AI capabilities, it will not translate into better legal outcomes for clients at lower costs.

That said, AI could reduce the costs of legal services for reasons unrelated to its ability to perform legal tasks. As the 2017 Clio survey mentioned above found, nonlegal work consumes a large percentage of lawyers’ time at the low end of the market. If advanced AI helps find and communicate with clients, manage administrative tasks, and handle payments, it could free up these lawyers to spend more time on legal work.

Adversarial Dynamics

Even in a world in which AI increases lawyers’ productivity and completes legal tasks, it might not lower the costs of legal services. To see why, it is crucial to distinguish inputs and outputs from outcomes. Inputs are what goes into legal work: employee talent, billable hours, and technological tools. Outputs are what legal work produces: contracts drafted, motions filed, and briefs written. Outcomes are what clients actually care about: disputes resolved, deals closed, and rights protected.

Consumers purchase legal services to achieve specific outcomes. Inputs and outputs can indirectly lead to those outcomes because more hours worked and more legal tasks help clients get the outcomes they want. But in a zero-sum context where the value of legal services is relative, if both sides increase their outputs, the advantages to either side of doing so can be limited.

Instead, AI might simply raise the inputs and outputs required to reach the same outcome, with productivity gains absorbed by greater production. The billable hours model, in which lawyers are paid and promoted based on inputs, only reinforces these dynamics: More hours worked drafting motions and reviewing documents translates into greater revenue for legal firms without necessarily improving outcomes.

Litigation

One response is that these arms races actually create value by increasing the quantity and quality of legal work. However, clients sometimes achieve their desired outcomes (like settling a case or dismissing a lawsuit) by imposing greater costs on the other side instead of improving the quality of their legal arguments or evidence. It is a “core premise of litigation economics”58 that “all things being equal, the party facing higher costs will settle on terms more favorable to the party facing lower costs.”59 And even where quality does improve, it’s uncertain that the benefits of higher quality legal work (like helping courts reach the “right” answer more often) outweigh the costs of more legal work (like overwhelming judges with cases).

Earlier technological shifts cast doubt on whether America’s “adversarial legalism” can translate productivity gains into more affordable legal services.60

Discovery is a cornerstone of American litigation that often determines whether cases settle or go to trial. In discovery, parties share information “to identify material facts that prove or disprove a claim.”61 It operates through an adversarial exchange: One party sends a discovery request; the other searches its records and decides which documents are responsive and which are protected by privilege.

Discovery was conceived of as a cooperative process during which lawyers could share information to facilitate settlement and avoid trial.62 Yet two characteristics of discovery make it vulnerable to abuse. First, the party holding the documents (the reviewing party) knows what’s in them but seeks to share as little helpful information with the requesting party as possible to minimize legal risk. Second, the reviewing party, because they bear the responsibility and costs for producing documents, must essentially act as their adversary’s agent.

Each side can leverage these features to impose costs on the other. A requesting party, through excessive requests, can compel their adversary to review more documents for confidentiality and relevance. And a reviewing party, through excessive production, can bury relevant information in mountains of extraneous material, forcing the opposing side to spend more time on review. This can create an arms race of overrequesting and oversharing, as each side drives up costs for the other to pressure them to settle on more favorable terms. The billable hours model again reinforces this behavior, with more review generating more billable hours.

These adversarial incentives can be powerful. Judge Frank Easterbrook opened a well-known article, “Discovery as Abuse,” by analogizing the process to “nuclear war.”63 Charles Yablon described how one side made life diff