AI & The Law: Changing Practice, Claude Constitution, & New Rights, w/ Kevin & Alan of Scaling Laws
Summary
Frontier AI has crossed the profession’s credibility threshold: Alan Rozenshtein says leading models are “certainly better than the median lawyer” in raw intellectual horsepower. Nathan Labenz cites GDPval results where Claude Opus 4/5 wins one-third of head-to-head comparisons with lawyers and wins or ties 70%; Alan considers ChatGPT 5.2 strongest on legal “taste,” despite using Claude as his daily driver. Hallucinations and missing specialist databases remain, but his strategic conclusion is blunt: “It’s over, right?”
Legal-tech adoption is being constrained less by capability than by law-firm incentives, workflow inertia, and symbolic procurement. Kevin Frazier cites Harvey’s claim that roughly 70% of major top-100 US firms use its product, yet lawyers often received one launch email, no meaningful training, and no expectation to use it. The billable hour rewards spending “as much time as possible” within client tolerance, while “secret cyborgs” hide their productivity and firms only whisper about hiring fewer summer and junior associates.
The size of legal AI’s market hinges on whether cheaper services uncover latent demand or merely eliminate expensive human work. Kevin points to “legal deserts” with about one lawyer per 1,000 residents; Alan imagines agents negotiating nearly complete contingent contracts at 400 tokens per second and stresses that legal services have arms-race dynamics absent from dentistry. Yet he partially talks himself back toward Nathan’s skepticism: once AI has searched every relevant precedent and document, “at some point your teeth are just clean.”
Entry-level legal work is vulnerable even if aggregate lawyer employment ultimately grows, creating a dangerous apprenticeship gap. Discovery, precedent searches, and contracts assembled from a firm’s prior thousand examples are already natural automation targets; Alan himself now substitutes Gemini and Claude workflows for some research-assistant assignments. His deeper concern is cognitive and political: AI favors “high-agency people,” while rule-followers may feel they “did all the right things and the rug was pulled out from under them.”
Lawyers retain a regulatory moat, but state competition and free-speech constraints may prevent a complete chatbot blockade. Unauthorized-practice statutes previously obstructed services such as LegalZoom, yet Arizona now permits nonlawyers to own law firms while Texas and Utah are leaning into regulatory sandboxes. Alan expects human checkpoints for court appearances or specified transactions. Nathan argues that banning general legal discussion by ChatGPT would look like “obvious guild protective self-dealing” and face First Amendment problems; Alan separately says people should have a right to access these tools.
AI could move law from procedural accumulation toward measurable outcomes, simulations, and dynamically updated agreements. Kevin wants legislators forced to state the problem a bill is meant to solve, evaluate outputs such as emissions or congestion, and simulate how actors might exploit it; future generations may ask, “What the hell?” when they see today’s laws were not tested this way. The Claude Constitution offers the philosophical analogue: not rules versus principles, but contextual judgment—Aristotelian phronesis—with a few inviolable boundaries.
The next rights contest spans access to compute, control of personal data, AI-enabled state power, and eventually AI welfare. A right to compute has already been enacted in Montana and proposed in Ohio and New Hampshire; Kevin also argues for a “right to share” personal data, including educational records, with chosen AI systems. Against that, Alan’s “unitary artificial executive” could give a president granular control over millions of officials, while mass analysis of public audio threatens pervasive surveillance—and convincing companions within 10 or 15 years could ignite conflict between those alleging digital enslavement and those who see AI personhood as “an affront to God.”
Deep dive
1. AI is colliding with a legal system built for an analog world
Kevin’s opening image is a “big traffic jam,” perhaps even “a huge crash”: privacy principles such as the FIPPs date to the 1970s, while still-older case law assigns rights and obligations for an analog society. The internet had already pressure-tested those regimes; AI is “putting all of that on steroids.”
Alan separates two intersections that law schools increasingly need to teach. The law of AI asks how society should regulate, promote, or control a major technology; AI in law asks what happens when that technology enters a profession whose core output is cognitive work.
Law is not quite as disembodied as programming—courts still expect human beings to appear—but much of both fields is “the manipulation of certain kinds of symbols.” Alan estimates legal transformation is perhaps one or two years behind software engineering, “not 30 years behind,” even though the lawyer-controlled guild will slow implementation.
2. Frontier models already outperform the median lawyer
Nathan’s benchmark reference sets the capability baseline: Claude Opus 4/5 leads the lawyers category in public GDPval data, winning one in three head-to-head comparisons against humans and winning or tying 70%. The prompt set is small, but the models have plainly climbed far up the professional-performance ladder.
Alan’s caveats are operational rather than fundamental. Models still make mistakes and hallucinate, and they may lack access to a database containing “that one random SEC regulation” buried in the Federal Register; he considers those limitations “fairly trivially solvable” over the next few years.
His model preferences are differentiated: Claude and Claude Code are the daily workspace, but he calls out to ChatGPT 5.2—especially its “Pro extended-thinking” mode—for legal analysis. His “vibes perspective” is that OpenAI has invested most heavily in legal training and currently exhibits the best legal taste, though all three leading models provide strong answers.
Alan continuously asks models to pressure-test his scholarship and now regards them as better than the median lawyer in raw intellectual horsepower. Even skeptical academics soften after an hour with a $20 plan: “If all they’re doing is fancy autocomplete, then all I’m doing is fancy autocomplete.” Bespoke judgment, such as 50 Supreme Court arguments’ worth of experience, will remain harder to reproduce.
3. Legal AI’s total market depends on latent demand
Nathan frames the uncertainty as a spectrum. People want the minimum necessary dentistry or accounting and would pocket a tenfold cost reduction, whereas software might support 10 or 100 times more output; his own simple contract review let him avoid hiring an attorney.
Kevin’s counterexample is the legal desert, an area with roughly one lawyer per 1,000 residents. People cannot readily obtain help with leases, small businesses, nonprofits, divorces, or other disputes; even limited counsel in landlord-tenant cases can materially improve a tenant’s chance of success.
Beyond routine representation, Kevin expects a track of “legal architects” who design regulatory structures and incentive systems rather than merely execute established workflows. He cites Gillian Hadfield’s work with Fathom and Andrew Friedman as a specimen of the higher-level, more creative lawyering that education should cultivate.
Alan’s explicit wager is that Jevons paradox holds: as legal services become cheaper, people consume more, lawyers move up the value chain, and in 10, 15, or 20 years there may be at least as many lawyers and substantially more legal service. But he repeatedly marks this as the central unknown, not a confident forecast.
4. AI agents could make contracts exhaustive and law continuous
Alan uses the “complete contingent contract” to show what suppressed demand could mean. With infinite time, energy, and no opportunity cost, counterparties would negotiate almost every possible eventuality; today they stop early and accept legal default rules that inevitably misfire.
Personal agents could instead confer at inference speed—Alan uses 400 tokens per second as the illustration—and produce orders of magnitude more detailed agreements. Law is also competitive: unlike a person and their teeth, counterparties may continually buy better legal representation because the other side can do the same.
Kevin extends that programmability to legislation. A law might trigger a specified economic policy if unemployment in one field reaches 7%, or automatically initiate a response when another country imposes tariffs; present-day statutes barely exploit this capacity for conditional, adaptive governance.
5. The billable hour is delaying the law-firm reckoning
Kevin has begun asking firms whether they would prefer Harvard’s top graduate with no AI experience or an AI expert from a middle-ranked law school. Increasingly, practitioners choose the latter because that hire can find frontier tools and teach the rest of the organization.
Procurement does not equal transformation. Kevin cites Harvey’s own statistic that roughly 70% of major top-100 US firms use its litigation-focused system, yet employees often recall only an introductory email, no substantial training, and no obligation to incorporate it into daily work.
The compensation mechanism points the wrong way: under the billable hour, a lawyer benefits from spending as long as possible on a task while staying inside the client’s acceptable range. Firms know this model works economically and remain reluctant to make efficiency the product.
Kevin nevertheless hears “whispers” about smaller summer and junior-associate classes, alongside Ethan Mollick’s “secret cyborgs” who conceal how much AI already does. Alan is more cautious about current displacement: firms are badly managed, legal adoption lags capability, and humans appearing in court must personally attest to work that could contain an AI hallucination.
6. Automating junior work may break the apprenticeship ladder
Entry-level law contains exactly the work current systems handle well: discovery, finding needles in document haystacks, and drafting a contract from the thousand similar agreements a firm has already produced. Alan expects these tasks either to disappear or to become radically different.
His analogy is programming’s repeated abstraction ladder. Assembly language was once dismissed as cheating, followed by higher-level languages, garbage collection, the Java Virtual Machine, Python, and now natural-language prompting; each stage removed lower-level burdens while expanding the problems programmers could attempt.
That history suggests deskilling is not inevitable. New programmers may memorize less syntax but confront architecture years earlier; lawyers and doctors might likewise stop spending scarce “IQ points” on work analogous to long division or organic chemistry and redirect them toward diagnosis, system design, and judgment. Whether foundational drudgery is educationally necessary remains unknown.
Alan already uses research assistants less: a script downloads PDFs, Gemini Flash summarizes them, and Gemini Pro plus the Claude API debate which articles matter before producing formatted Markdown. Yet his own career grew from doing “nonsense crap work” near a professor long enough to absorb the profession; losing that proximity could weaken the apprenticeship path. Separately, Alan says low-agency rule-followers may feel betrayed and fuel political friction over the next decade.
7. AI’s jagged strengths matter more than job-level averages
Alan pushes Nathan to define what “better” means because professions are bundles of tasks. A model can be vastly superior at standardized reasoning yet incompetent elsewhere; averaging those strengths and weaknesses into one score obscures where substitution will actually occur.
Nathan’s pediatric-oncology experience supplies the comparison. Models appeared better than residents and roughly toe-to-toe with attending physicians when synthesizing written observations and test results, but nurses’ bedside tasks—such as handling a frightened child or managing equipment—were largely untouched, while doctors added value by looking holistically at breathing, color, and visible distress.
The equivalent in law may be court representation, accountability for checked work, and contextual judgment rather than pure research. A model can surface every argument, yet a human advocate may still be required to manage a courtroom, certify accuracy, or decide that the technically available move is strategically unwise.
The medical analogy also identifies a likely legal fallback: ChatGPT may discuss test results while a human doctor must write a prescription for morphine. General legal advice may likewise become abundant while licensed humans retain authority over selected transactions and appearances. The fight will concern where that mandatory checkpoint sits.
8. Guild barriers will bend through competition rather than vanish
Every state regulates legal practice through bar admission, accredited education, examinations, continuing education, and unauthorized-practice-of-law statutes. Those UPL rules stop a Craigslist claimant from representing clients at half price and previously created substantial obstacles even for LegalZoom’s wills and real-estate documents.
Kevin’s preferred payoff is not merely cheaper briefs. Because about 95% of litigation occurs in state courts, people can wait months or years before an overburdened—or simply “hangry”—judge reaches an inconsistent result. In an adversarial system, “whoever can pay the most money wins” because that party can keep its lawyers fighting longest.
Tools such as Learned Hand, which helps judges and clerks draft opinions, could make lower-level adjudication faster and more consistent. Lawyers would then shift toward an appellate-style function: choosing the system’s ends, designing incentives around community values, and monitoring whether automated resolution actually honors rights.
Kevin expects competitive pressure from Arizona, the first state he identifies as allowing nonlawyers to own law firms, plus Texas and Utah sandboxes. Alan says people should have a right to access these models and expects courts to treat broad restrictions on that access as First Amendment questions. Nathan argues that banning general legal discussion by chatbots would be difficult and could look like guild protectionism.
The exact human checkpoint remains unsettled. Kevin did not know whether a civil judge can require a financially able litigant who wants to proceed pro se to hire a lawyer; Claude’s answer, which he consulted, said the right is strongest in criminal trials and weaker in civil cases, with exceptions.
9. Cheap AI may finally exhaust law’s search space
Alan models litigation as a combinatorial search across arguments, precedents, and billions of document fragments for the decisive sentence. Legal costs rose because one more expensive human searcher could still generate more expected value than the additional labor cost.
Legal technology has already lowered search costs: Westlaw and Lexis digitized what had been paper databases beginning decades ago, while more recent machine-learning tools improved discovery. Even so, firms still lock costly humans in conference rooms to read and classify material.
Now imagine systems 10,000 times more effective and four orders of magnitude cheaper—Alan summarizes the combined effect as “a million times better.” They might read every relevant sentence and exhaust every useful precedent, creating a natural ceiling because “there’s just nothing more to spend on.”
He immediately preserves the alternative: lawyers may keep expanding the search space, and tiny differences in assumed capability growth, cost decline, or induced demand compound into enormous ten-year forecast gaps. The conversation therefore ends this branch with genuine uncertainty, not a clean abundance thesis.
10. Outcome-based law could replace procedural fetishism
Kevin argues that civil litigation is currently a staircase of complex procedural moves, including motions that can raise legitimate objections or simply delay resolution. Lawyers often equate fairness with adding more opportunities to intervene, a habit Nick Bagley calls the “procedural fetish.”
Those participation points are not neutral because unusually organized or expressive actors are likelier to use them. More procedure can therefore place “gum into the cogs of the system” without representing the people affected or advancing the law’s original purpose.
Kevin uses NEPA as an example of a law whose pressure points might have been stress-tested in advance. He refers to it as the “National Economic Protection Act,” while saying people call it the Environmental Protection Act, and asks whether simulation could have exposed veto points and tested whether outcomes matched the drafters’ environmental goals.
The alternative is outcome-based law: define what both parties or the public actually want, then let agents trained on incomes, preferences, aspirations, and professional goals continuously update agreements toward that end. It is deliberately optimistic and “very sort of sci-fi,” but Kevin regards it as technically conceivable.
This changes the governing question from how many stages a narrow dispute traversed to whether the desired condition occurred. Lawyers remain important, but their work becomes specifying legitimate ends, acceptable tradeoffs, and mechanisms for auditing whether the system consistently produces them.
11. Good legal judgment requires both rules and principles
Nathan introduces the tension through research finding GPT-4 more strictly formalist than human judges, who appeared more legally realist. Literal application promises predictability but can produce bizarre outcomes from badly drafted laws; broad discretion accommodates context but also opens the door to bias and post-hoc justification.
Kevin’s canonical test is a park sign stating “no vehicles allowed.” Cars may be obvious, but drones, strollers, scooters, and ambulances reveal that even apparently precise language cannot enumerate every future case or encode the drafter’s actual intention.
He therefore resists perfect textualism and an AI-generated code for every behavior. The American common-law tradition tolerates ambiguity so governance can evolve, while an exhaustive regime risks a world where stepping on the wrong crack automatically produces a fine within five days and a bank-account deduction.
Alan adds that GPT-4’s formalism was a contingent training result, not an inherent quality of models. A differently trained system could prioritize legislative purpose; Minnesota appellate judges he met were cautious but surprisingly open to AI, and better legal evaluations need to arrive within a month rather than becoming obsolete over a year and a half.
12. The Claude Constitution turns virtue ethics into an experiment
Alan rejects a binary between textualism and realism: nobody ignores purpose under every circumstance, and nobody treats legal text as wholly nonbinding. Antonin Scalia’s “rule of law is the law of rules” and Stephen Breyer’s 17-factor style occupied opposite ends of a relatively narrow middle.
He reads Amanda Askell’s Claude Constitution as deeply Aristotelian—a modern set of “footnotes” on the Nicomachean Ethics. Comprehensive ethical rules are impossible, so an intelligence needs phronesis: cultivated practical judgment capable of moving between high-level principles and local context.
Yet principles can themselves recommend hard rules. The Constitution lists roughly 17 principles in no fixed priority while categorically refusing certain outputs, including child-sex material or help developing airborne Ebola; “yes, and” replaces the supposed choice between standards and prohibitions.
AI makes the ancient argument empirically tractable through in silico experiments. Researchers can vary the mix of rule-following and principle-based reasoning at a speed and scale impossible in courts or societies, potentially learning not only how machine judgment works but something new about human intelligence.
13. New rights will collide with new state and machine power
Alan argues that people should have a right to use these models, with a negative right of access fitting naturally under the First Amendment, analogous to reading books or visiting libraries. The harder positive right would require society to provide compute credits or budgets, perhaps in a future where compute itself functions as currency.
Kevin places this under the “right to compute,” already enacted in Montana and under consideration in Ohio, New Hampshire, and, he believes, other states. The principle would raise the threshold before government can block a person from expressing themselves or receiving information through AI and later computational tools.
His complementary “right to share” would let individuals frictionlessly give chosen systems their own data. FERPA can obstruct a parent who wants educational records to train a personalized tutor, while rich people can travel for comprehensive scans and AI health recommendations; everyone else is left with “whatever Walgreens told us at that last checkup.”
Rights claims may eventually extend to models themselves. Alan expects convincing voice, video, memory, and embodied companions within 10 or 15 years; over 20–30 years, attachment could divide those who see exploited sentient beings from religious opponents who consider that belief idolatrous and demand a “Dune-style Butlerian Jihad.”
14. AI could create a unitary artificial executive
Alan’s “unitary artificial executive” describes near-term AI concentrating presidential power through perfect enforcement, pervasive surveillance, mass propaganda, and managerial control. A system trained on a president’s preferences could sit throughout a bureaucracy of millions, reading emails and texts and enforcing granular alignment in real time.
The tradeoff is real rather than one-sided. Elections should have consequences, and AI could improve services and state capacity in a government many citizens believe takes taxes without delivering; the same infrastructure could supercharge abusive authority far beyond anything previous presidents could practically exercise.
Kevin flags the Fourth Amendment as an urgent pressure point. Government could potentially tap systems that detect and pick up audio, allowing ordinary public conversation to be “hoovered up,” synthesized, and analyzed for who is planning, thinking, or wanting what—without meaningful notification.
His constructive constraint is transparent experimentation: governments should use regulatory sandboxes, notify affected people, provide feedback channels, and test new systems without assuming current procedures are sacred. The task is to improve state capability while preventing AI from converting ordinary governmental reach into pervasive, invisible control.