A Stitching Thesis sequel, where legal value capture concentrates when the cost of intelligence bifurcates
Executive thesis. The dominant story about AI in legal is a deflation story: models get cheaper, work gets faster, the billable hour collapses, and clients capture the savings. That story is half right. It describes the bottom of the market accurately and the top of the market not at all. The missing variable is that there are two cost curves, not one.
Commodity tokens are racing toward zero, and that deflation is precisely what will finally make competent legal help available at population scale, the unlock the stitching thesis described. But frontier intelligence, the most advanced reasoning and the proprietary capability built on top of it, is moving in the opposite direction: improving relentlessly, and pricing itself out of reach for a large share of the firms and enterprises that would want it. Falling token costs will not democratize frontier intelligence. The legal market will not flatten; it will stratify, more capacity below, more proprietary capability above. The billable hour does not die in this world. It splits.
At the bottom it becomes a fixed or per-outcome price delivered at scale. At the top it becomes something new: a reasoning budget, the metered, orchestrated, governed allocation of scarce frontier reasoning and the human accountable for it. For investors, the durable margin is not in the deflating commodity tier that delivers the access. It is in the widening, scarce, governed tier at the top, the same fusion-and-governance layer the thesis already identified as the moat. The token-shortage curve is the macroeconomic proof of that thesis.
Not investment advice. This brief is prepared by Ajay Mago in his capacity at Twelvefold Ventures for informational and discussion purposes only. It is not investment, legal, tax, or financial advice, nor an offer or solicitation with respect to any security or fund interest. Forward-looking statements reflect the author’s views and assumptions and are subject to change. The author and/or Twelvefold Ventures may hold positions in companies referenced herein.
I. The buried assumption
Every confident prediction of the billable hour’s death rests on one unexamined premise: that AI is, and will stay, cheap, and cheap for everyone. Until about a month ago, reading the 2026 forecasts, you would find the same architecture every time. AI compresses the time a task takes; compressed time cannot sustain hourly pricing; therefore fixed fees, subscriptions, and the slow extinction of the six-minute increment. The logic is sound. The premise is not.
There are two things wrong with “AI is cheap.” The first is that “cheap” describes only one layer of the market. The second is that the best models for serious legal work will not be affordable to everyone. A world is coming in which a firm’s competitive position is determined in part by whether it can afford the frontier, and many firms, and many of their clients, will not be able to. That is not a deflation story. It is a stratification story, and it runs in the opposite direction from the consensus.
The consensus is already cracking, and the open-source movement is the first visible crack. In the spring of 2026 a former Latham & Watkins associate released Mike OSS, a free, self-hostable, bring-your-own-key legal AI, and reproduced the core of what Harvey and Legora sell at enterprise prices in roughly two weeks. [1] Mike does not have to win to matter. As one former law-firm IT director put it, an open-source alternative sitting on GitHub does not kill the incumbents, but it changes the renewal conversation from “is this magic?” to “what exactly am I paying enterprise prices for?” That question is the whole argument of this brief in miniature. Open source drags the commodity layer toward zero and forces everyone above it to name the value they actually add. Mike’s creator put the thesis plainly: thin wrappers with no unique value get absorbed by the model providers; thick wrappers with a real value proposition survive. The thick layer is the scarce one, and it is where this brief spends the rest of its time.
II. Two cost curves, not one
Start with what is true in the consensus. The unit cost of intelligence is collapsing. Per-token prices for a given capability have fallen by factors ranging from roughly nine-fold to several hundred-fold per year. [2] This is real, and it is the engine of access. When the marginal cost of a competent contract review or a benefits appeal approaches zero, the addressable market for legal help expands from those who can pay several hundred dollars an hour to anyone with a legal problem, the low-income Americans who currently receive inadequate or no help for 92% of their civil legal problems. [3] The floor of the market rises.
But aggregate spending on intelligence is not falling; it is rising sharply. Enterprise AI budgets grew several-fold between 2024 and 2026 even as per-token prices fell, because consumption, agentic systems that reason in long chains and call models thousands of times per task, is scaling far faster than unit costs decline. [4] The frontier itself is expensive to reach and getting more so: the most capable systems run on the scarcest compute, and providers have been pricing inference below cost to win share, a subsidy that sets a false floor destined to normalize. [5]
That false floor is now beginning to lift, though not in the way a simple story would predict. Through the first half of 2026, as both leading labs filed confidentially for public offerings, the pressure to show real unit economics collided with a price war: OpenAI, having posted deeply negative operating margins, was reported to be weighing steep token-price cuts to win enterprise share back from Anthropic, which had just reached its first profitable quarter on a largely enterprise base. [6] The subsidy era is ending, but asymmetrically. At the commodity layer it barely matters who blinks, because open-source models, many of them Chinese, at a fraction of Western prices, have already pinned that floor near zero and keep dragging it down. What surfaces as the subsidy lifts is the true cost of the frontier: the agentic workloads that reason in long chains and burn through an enterprise’s annual AI budget in a single quarter. The bill that explodes is not the commodity bill; it is the frontier bill. Shedding the subsidy does not flatten the two curves. It widens the gap between them.
Falling token costs will not democratize frontier intelligence. Cheaper commodity tokens and a more expensive, faster-moving frontier are not a contradiction. They are the same structural fact the stitching thesis stated in a different vocabulary, certain foundation models will become table stakes, certain foundation models will be cost-prohibitive, and the durable moat is what you build in between and on top of them. Value flees the commoditizing layer and concentrates in the scarce one. The cost curve is simply the macroeconomics of that migration.
Will the same efficiency breakthroughs that cheapened commodity tokens, caching, distillation, smaller specialized models, better silicon, eventually cheapen the frontier too? Partly, yes. But three forces keep the frontier scarce. The frontier moves: today’s most advanced reasoning becomes tomorrow’s commodity, and a new, costlier frontier always opens above it, exactly as it has in every prior software and hardware cycle. The bottleneck migrates rather than vanishes: when raw reasoning is cheap, the scarce input becomes verification, judgment, and accountability, none of which ride the compute curve down. And affordability stratifies: even as the frontier improves, accessing the best of it stays expensive enough that a large swath of enterprises will choose, or be forced, to run on the tier below. The improvement of frontier capability is precisely what prices it beyond most, the same way the best chips, the best infrastructure, and the best talent have always been rationed by price.
III. The stratification: more capacity below, more proprietary capability above
Here is the shape of the market that follows. At the lower tiers, capability becomes abundant, commoditized, and broadly affordable. Routine drafting, first-pass review, standard diligence, regulatory tracking, work that filled much of a junior associate’s week, collapses into supervised, near-free workflows available to anyone. The floor rises, and it rises for everyone.
At the upper tiers, the opposite happens. The most advanced reasoning, fused with a firm’s proprietary data, models, and validated workflows, becomes a genuine and widening source of advantage, and an expensive one. The counsel advising Blackstone on a contested, bet-the-company matter will not merely have more experienced partners than the firm serving a middle-market acquirer. It will deploy frontier reasoning and proprietary capability the middle-market firm cannot afford to run. [7] The gap between the top and the middle of the legal market will widen, not narrow, and it will widen along a line drawn by who can afford the frontier.
Nowhere is this starker than in high-stakes litigation, where the advantage compounds. A firm that pairs elite trial lawyers with its own proprietary corpus, decades of briefs, outcomes, judge-specific tendencies, and a record of what actually worked, and runs the best available models over that private context will systematically outperform an opponent renting commodity capability over public data. The edge widens with every matter, because each case feeds the corpus that sharpens it: a data network effect operating inside a single firm. In a bet-the-company trial that gap is not a convenience; it is the difference between winning and losing, and the clients with the most at stake will pay almost anything to be on the right side of it. A firm properly positioned, institutional data, top-end attorneys, the best models, deployed together, will simply trounce the firm that is not.
Capability is not the only axis that stratifies; speed is the other. In a contested matter, the advantage goes not merely to the firm with the best answer but to the firm that reaches the right decision first, to preempt a filing, to move inside a closing window, to respond before opposing counsel has finished reading. The analogy is high-frequency trading, where funds spend fortunes on co-located servers and low-latency links to shave microseconds off execution, buying an edge that is simply unavailable to anyone who cannot pay for it. Frontier legal capability will work the same way: the firms that can orchestrate the best models over the best context, fastest, will convert speed-to-decision into outcomes, and that speed will be out of reach for whole segments of the market. This should not read as dystopian. Legal quality has always been stratified by the ability to pay, elite counsel was never evenly distributed. AI does not create that stratification; it adds two new dimensions to it, capability and speed, and widens the gap along both.
This is the uncomfortable corollary of the access story, and it should be stated plainly because investors will price it either way. AI raises the floor and raises the ceiling faster. It democratizes competence and stratifies excellence. Both are happening at once; they are not in tension; they are the same coin. The justice gap narrows at the bottom even as a new, sharper gap opens at the top between the firms that own frontier capability and those that rent commodity capability. The durable enterprise value is being created at the top of that widening gap, which is the part of the market the deflation story tells you to ignore.
IV. Why the billable hour bifurcates, and why the data already shows it
The billable hour was never really a measure of time. It was a bundle that did three jobs at once: it recovered the cost of producing the work, it priced the expertise applied, a partner’s rate and an associate’s rate are an expertise tier, not a stopwatch reading, and it absorbed risk, shifting the uncertainty of unclear scope and uncertain outcomes onto the client. The hour survived for a century because labor happened to be the dominant cost, a decent proxy for expertise, and correlated with effort-at-risk, so one number could do all three jobs. AI severs all three correlations at once, and the bundle comes apart along the fault line between the two cost curves.
At the bottom, where work is bounded, repeatable, and poolable, the hour genuinely dies. Clio’s data already shows up to 74 percent of hourly billable tasks exposed to automation, the average lawyer recording just 2.6 billable hours in an eight-hour day, and flat-fee billing up by more than a third over the prior decade. [8] This work moves to fixed, per-matter, and subscription pricing, delivered at scale to a market that could never previously afford it. The deflation story is correct here, and only here.
At the top, the hour persists, and the reason is not nostalgia. It is that the work at the top is the work where risk cannot be pooled. A bet-the-company matter is a sample of one: the scope is genuinely unknowable, the downside is asymmetric, and no honest fixed fee can absorb that uncertainty without either gouging the client or ruining the firm. The hour is the only instrument that prices it cleanly. The market is already telling us this. Thomson Reuters’ rate analyses found that whether firms discount aggressively or hold firm, they collect roughly the same per hour, and that, despite years of heavy AI investment, roughly 90 percent of legal dollars still flow through hourly billing. [9] The hour survives at the top of the complexity curve. The question is no longer whether it survives, but what it becomes.
V. From hour to reasoning budget
What the hour becomes, at the top, is a reasoning budget. Not a measure of human time, and not a meter on raw tokens, but an allocation of governed reasoning capacity to a matter in proportion to what is at stake.
Governed reasoning capacity is the unit. It is reasoning a licensed professional can stand behind, defensible, attributable, and privileged. It carries provenance and chain-of-custody, runs under a controlled data regime, is validated against a quality bar, passes through human checkpoints with a named accountable person, and leaves an audit trail. It behaves like a budget: metered and allocable, spent more heavily on the matters that warrant it. But unlike raw compute, it comes with a quality-and-accountability guarantee, and that guarantee is what the client is buying.
Two capabilities sit on top of the raw frontier and constitute the actual product. The first is orchestration: deciding which model handles which step, routing a task across a portfolio of systems, sequencing multi-step agentic workflows, and spending scarce frontier reasoning only where it earns its cost. As capability fragments across many models and providers, orchestration becomes a discipline in its own right, the firm is no longer choosing a tool but conducting an ensemble. Firms will differ enormously here. Knowing which model to deploy on which context, when frontier reasoning is worth its cost and when commodity capacity suffices, and how to assemble the right private context for a task is itself a scarce skill, and access to the best models on the best context is unevenly distributed. It is also worth saying plainly that no one yet knows what compute will actually cost: between the price war, the open-source floor, and the unwinding subsidy, today’s token prices are an unreliable guide to tomorrow’s, and a firm that hard-codes them into a fixed fee is building on sand. The winners will treat compute cost as a variable to manage, not a constant to assume.
The second capability is governance: the controls, auditability, explainability, and accountability that turn a powerful output into one a general counsel will actually sign. The blocker on serious AI deployment in regulated work has never been raw model capability; it is accountability, who supervises the algorithm, who is liable when it is wrong, what record proves compliance. Orchestration and governance together are the scarce layer. They do not ride the compute curve down, because they are human, institutional, and regulatory work. That is why their value holds while token prices fall.
The governed layer also makes possible a move that points to where proprietary capability becomes genuinely hard to copy. When a single client is served by a panel of firms, co-counsel on a deal, multiple firms across a litigation portfolio, federated learning lets those firms collaboratively improve a shared model on the client’s matters without any of them surrendering privileged data. The data never leaves each firm’s control; only the learning does. Privilege is preserved, confidentiality holds, and the model still gets better for everyone working the client’s problems. That is a capability a commodity tool cannot offer and a middle-market firm cannot easily assemble, and it compounds: the more matters the panel runs, the sharper the shared edge. It is the stitching thesis’s data network effect, rebuilt to survive the privilege constraint.
The top-tier bundle, then, is several things the middle of the market cannot easily assemble: senior human judgment, access to frontier compute, the firm’s own proprietary capability, tuned models, validated playbooks, the orchestration layer, and the governance to stand behind all of it. The reasoning budget is the price of that bundle, allocated by stakes.
This is why the reasoning budget will not simply diffuse to everyone the moment the models are available to everyone. It requires a combination of digital and human capability, frontier engineering, orchestration, governance, and elite legal judgment, working as one, that very few organizations are equipped to assemble. The precedent is electrification. The dynamo was available to nearly every factory by the 1890s, but the productivity gains did not arrive for almost forty years, because they required factories to be redesigned around the electric motor rather than simply wired in place of the steam engine. The dot-com build-out told the same story a century later: cheap bandwidth was necessary but never sufficient. The firms that merely buy AI will get a faster steam engine. The firms that rebuild themselves around it will get the factory of the next era, and they will be few.
VI. Legal has always been priced by ROI, AI just makes it explicit
The reasoning budget resolves a tension the hourly model spent a century hiding: legal services have never really been priced by cost. They have been priced by return. A board does not pay a premium on a $4 billion acquisition because the work takes more hours; it pays because the downside of getting it wrong is catastrophic and the upside of getting it right is enormous. Price at the top of the market has always floated toward the ROI of the outcome, with the hour serving as the polite accounting fiction that connected that price to something measurable.
AI removes the fiction, at least in part. When the cost of producing the work decouples from the value of the work, collapsing toward compute at the bottom, concentrating in scarce capability at the top, price can no longer pretend to track cost. It tracks ROI directly. At the top of the market that means a decisive move toward outcome and success-based pricing: the firm is paid for the result, or the likelihood of it, not the effort. And here the same technology that threatens the hour makes outcome pricing newly viable. The thing that historically made success-based fees impossible on complex matters was that the risk could not be estimated. Litigation analytics now quantify motion outcomes, settlement likelihood, and judicial behavior with enough confidence to price risk that was previously unpriceable. [10] So AI squeezes the hour from both ends: it commoditizes the bottom of the market into fixed fees, and it makes the top of the market estimable enough to price by outcome. The hour is left occupying a shrinking middle.
This also forces a question the profession has deferred: is the cost of intelligence overhead, or is it a billable disbursement? The last technology to pose this question was online legal research, and client pressure eventually pushed Westlaw and Lexis charges from pass-through line items into firm overhead. AI compute will reopen the fight at far larger scale, and the volatility of the frontier makes it acute. A firm that treats compute as overhead and quotes fixed fees on today’s subsidized prices will be exposed when the false floor lifts; a firm that passes compute through as a disbursement offloads that risk to clients but collides with their demand for cost certainty and with the ethical limits on marking up disbursements. The deeper consequence is that AI strips the hour of its ability to hide margin. When a model did the work in four minutes, the firm can no longer bury its expertise premium inside a time entry. It must name what it is actually selling, orchestration, governance, judgment, outcome, and price each explicitly. That is healthy, and it is overdue.
VII. The enterprise: build, buy, or orchestrate
For the general counsel, the same stratification reframes the buying decision. The commodity layer becomes a contest to be cheapest-compliant, fought among legal-tech products, in-house teams, and technology-enabled alternative providers, and, increasingly, self-hosted open-source tools that a firm can run behind its own walls. A traditional firm rarely wins that contest on price, and should not try; it cedes the bottom and moves up.
What moves up is harder to commoditize. As routine work disaggregates to whoever is cheapest, the enterprise’s legal-AI stack fragments into many tools, models, and providers, each with its own cost, quality, and compliance profile. Fragmentation manufactures a scarce role: the single accountable party who orchestrates that stack and owns the cost, quality, and compliance tradeoff end to end, one entity to supervise the system and answer for the result. The general counsel’s real question shifts from “why isn’t this cheaper?” to “who will orchestrate this stack and stand behind it?” A firm can claim that role or cede it, to a Big Four entity, a legal-tech platform, an in-house operations function, or a governance intermediary. Claiming it means becoming the institution that supplies governed reasoning capacity at scale: orchestration plus governance, sold as accountability. That is the stitching thesis applied to the firm itself, the value is in stitching the fragmented stack into a coherent, accountable whole, not in any single component of it. And as the electrification precedent warns, the firms that can actually assemble that combination of digital and human capability will be the exception, not the rule.
VIII. The regulatory window
Regulation is not a constraint on this model; it is a tailwind for the scarce layer. Three currents matter. First, the rules of fee reasonableness must be re-mapped: when the cost driver shifts from labor to compute, Model Rule 1.5 questions arise on both sides, charging several thousand dollars an hour for work a model did in minutes, and marking up metered compute beyond its cost. The American Bar Association’s Formal Opinion 512 has already confirmed that the duties of competence, confidentiality, supervision, and reasonable fees all apply to generative AI. [11]
Second, the duty of verification is now firmly and repeatedly enforced. What began with Mata v. Avianca has hardened into a body of appellate law. In early 2026 the Ninth Circuit not only sanctioned but suspended two attorneys for failing to disclose that hallucinated citations in their briefs came from generative AI, issuing a warning to its entire bar; weeks later the Sixth Circuit, in Whiting v. City of Athens, imposed sanctions for more than twenty fabricated citations, and the Sixth and Seventh Circuits have since staked out a split, agreeing that AI does not dilute a lawyer’s duties of competence, candor, and verification, but diverging on how harshly to punish lapses. [12] Tracking databases now count well over a thousand hallucination incidents, with double-digit decisions issued on a single day. The decisive feature for this brief is that courts draw no distinction by firm size or tool sophistication: an AmLaw 100 partner using a law-specific model bears exactly the same non-delegable verification duty as a solo practitioner using a chatbot, and elite firms have been caught too. Capability does not discharge the duty; governance does. That enforcement does not slow the governed layer; it is the single strongest reason it exists, by forcing the market toward systems with citation provenance, retrieval grounding, and human-in-the-loop control, exactly what governed reasoning capacity provides.
Third, the rewriting of unauthorized-practice and alternative-business-structure rules in jurisdictions like Utah and Arizona determines where the highest-value move, owning the service, not just selling software to it, is even legally available. For investors, that regulatory geography is a gating variable on returns. (The premium for compliant inference over commodity inference is real but second-order: a markup on the scarce layer, not the reason the frontier is scarce.)
IX. Where the value accrues
For capital, the conclusion follows directly from the two curves. Durable margin does not sit in the deflating commodity tier, however large its volume; that tier delivers the access and the total addressable market, but its economics race to the floor, and open source is already racing them there. Margin concentrates at the top of the widening gap: in the orchestration-and-governance layer that supplies governed reasoning capacity, in proprietary fusion capability with data network effects, and in service models that capture both software and services margin where regulation permits ownership. The investment error to avoid is underwriting legal AI as a cost-down story. It is a stratification story. The access tier and the margin tier are different businesses with different multiples, and they should be underwritten separately.
The professional-elevation thesis holds, with one honest amendment. AI in essential services remains a capacity-expansion play, not an automation play: it lifts professionals from data-gatherers to decision-makers and expands the demand for judgment rather than eliminating it. But the elevation is uneven. The floor rises for everyone; the ceiling rises fastest for those who own the frontier. More lawyers, doing more strategic work, serving more clients, and, at the very top, a smaller set of them commanding a widening premium. An investor who sees only the first half of that sentence has misread the trade.
X. The fractal
There is a pattern worth naming, because it repeats at every scale. The architecture of the stitching thesis, sense, fuse, deliver intelligence, elevate the professional, recurs whether the unit is a single clause, a single matter, a single firm, or the market as a whole. So does the bifurcation. Within one matter, commodity tasks fall to near-zero cost while a few moments of frontier judgment carry the value. Within one firm, a commodity practice and a frontier practice diverge in economics. Across the market, the floor and the ceiling pull apart on the same logic. And the pricing is self-similar all the way down: cost decouples from value, and price migrates from effort toward outcome, at every level you look. The thesis is fractal; so is the pricing. The firms and the investors who recognize that they are looking at the same shape repeated, and who position at the scarce, governed top of it rather than the abundant, commoditizing bottom, are the ones who will capture the returns of the next decade.
Prior writing: The Stitching Thesis (The Investor’s Brief)
Coming next week: a practitioner’s companion to this brief, on what the two curves mean for how firms price, staff, and govern legal AI, in The Legal Intelligence Brief.
Endnotes:
[1] Mike OSS, built in spring 2026 by former Latham & Watkins associate Will Chen, is a free, self-hostable, bring-your-own-key legal AI released under the AGPL-3.0 license as an open alternative to Harvey and Legora. Chen’s thesis: thin wrappers without unique value get absorbed by the model providers, while thick wrappers with a real value proposition survive. See Legal Futures, Legal IT Insider, and Artificial Lawyer (2026).
[2] Epoch AI’s analysis of state-of-the-art model benchmarks finds per-token inference prices have fallen between roughly 9x and 900x per year across performance milestones; Stanford’s AI Index recorded an ~280-fold drop in GPT-3.5-equivalent pricing between late 2022 and late 2024.
[3] Legal Services Corporation, The Justice Gap (2022): 92% of the civil legal problems of low-income Americans receive inadequate or no legal help.
[4] Gartner (Mar. 25, 2026) projects inference on a one-trillion-parameter model will cost providers over 90% less by 2030, yet warns that because token consumption rises faster than token costs fall, overall inference costs rise. “Chief Product Officers should not confuse the deflation of commodity tokens with the democratization of frontier reasoning” (Will Sommer, Gartner). Agentic workflows consume ~5–30x more tokens per task, and enterprise AI spend has climbed sharply even as unit prices fall.
[5] Gartner projects that the cost of LLM inference will decline by more than 90% by 2030. However, CIO Dive notes that enterprises may not realize equivalent savings because growing model complexity and increased usage are expected to offset much of the reduction in unit costs. See Gartner and CIO Dive (2026).
[6] As both leading labs filed confidentially for IPOs in 2026, OpenAI, reportedly posting a deeply negative (~-122%) adjusted operating margin in Q1 2026, was reported to be weighing steep token-price cuts to win enterprise share from Anthropic, which reached its first profitable quarter (Q2 2026) on a largely enterprise base. Chinese open-source models (e.g., DeepSeek) deliver comparable performance at a fraction of Western prices, pinning the commodity floor near zero, while agentic workloads have caused enterprises to exhaust annual AI budgets within a quarter. See Wall Street Journal (June 10, 2026), Decrypt, Investing.com, and Axios (2026).
[7] By 2026, top BigLaw partner rates reached roughly $3,400–$4,000 per hour (e.g., Susman Godfrey partners at $4,000). See ABA Journal, Reuters, and Wall Street Journal (2026).
[8] Clio, Legal Trends Report 2025: up to 74% of hourly billable tasks are exposed to AI automation (57% of lawyers’ own tasks); the average lawyer records ~2.6 billable hours per eight-hour day; flat-fee billing has risen ~34% since 2016.
[9] Thomson Reuters Institute, Law Firm Rates Report 2026: despite very different discounting and realization strategies, firms end up collecting roughly the same amount per hour. The 2026 State of the US Legal Market Report (TR Institute and Georgetown Law) finds ~90% of legal dollars still flow through hourly billing arrangements despite heavy AI investment.
[10] Litigation analytics platforms (e.g., Lex Machina, Bloomberg Law) quantify motion outcomes, settlement likelihood, and judicial behavior, making outcome-based pricing estimable on matters where risk was previously unpriceable.
[11] ABA Formal Opinion 512 (July 2024): Model Rules 1.1 (competence), 1.6 (confidentiality), 1.4 (communication), 5.1/5.3 (supervision), and 1.5 (fees) all apply to a lawyer’s use of generative AI.
[12] Following Mata v. Avianca (S.D.N.Y. 2023), appellate enforcement has hardened: the Ninth Circuit sanctioned and suspended counsel for undisclosed AI-hallucinated citations and warned its bar (Feb. 2026); the Sixth Circuit’s Whiting v. City of Athens imposed Rule 38 sanctions for 24+ fabricated citations (2026); the Sixth and Seventh Circuits agree AI does not dilute the duties of competence, candor, and verification but diverge on sanction severity. Tracking databases count 1,000+ incidents, with double-digit decisions on a single day, and courts apply Rule 11 uniformly regardless of firm size or tool sophistication. See BGov, National Law Review, and Kegler Brown (2026).