After Chevron, Who Carries the Burden?

"Deference" sounds like good manners. It is not. It is a rule about who loses when a statute runs out of words.

For forty years the answer was settled. Under Chevron U.S.A., Inc. v. Natural Resources Defense Council, Inc., 467 U.S. 837 (1984), a reviewing court first asked whether Congress had directly spoken to the precise question at issue. If it had, that was the end of the matter. If it had not, the court did not impose its own construction. It asked only whether the agency's answer rested on a permissible construction of the statute. The Institute's working draft on the decision that ended this regime names the crucial feature plainly: once a statute was deemed ambiguous, the burden fell on the challenger to show that the agency's interpretation was unreasonable or impermissible.

On June 28, 2024, in Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), the Supreme Court overruled Chevron. The burden moved.

Moved where? Not the question of whether Chevron was right; the academy has argued that at length. A narrower, colder question. When a court relocates a burden, who ends up holding the uncertainty the burden used to contain? And what does the answer demand of the machines now used to draft legal analysis?

What the Court changed

Start with the decision as the draft describes it. The case arose from a challenge to a National Marine Fisheries Service interpretation of the Magnuson-Stevens Fishery Conservation and Management Act: could the agency require fishing vessels to pay for onboard observers? The challenger said the statute did not allow it. The agency said the statute was ambiguous and its reading permissible at Chevron's second step.

The Court, by a vote of 6 to 3 in an opinion by Chief Justice Roberts, held that Chevron should be overruled. The draft summarizes three grounds. Chevron deference is inconsistent with the Administrative Procedure Act, which requires courts to exercise independent judgment. Its two-step framework conflicts with the principle that courts interpret statutes using the traditional tools of construction. And it misallocates power in a system designed for judicial review, not judicial deference. Courts must now determine the best reading of the statute. Agency views are "informative." They do not control.

The Court did not abolish all respect for agency views. It reoriented the landscape around Skidmore v. Swift & Co., 323 U.S. 134 (1944), under which an agency's interpretation earns weight by its persuasiveness: its thoroughness, its consistency, its reasoned explanation.

The draft sets out the shift in a table worth reproducing.

Read the first row again. On the challenger, before. On the agency, after. That is a burden-shifting event in the strictest sense, and it is why this essay belongs in a series about burdens.

The burden moved. Did the uncertainty?

Here is where the abstraction starts to leak.

A burden is a rule for resolving doubt: when the argument is in equipoise, the party holding the burden loses. Chevron resolved statutory doubt in one direction. Ambiguity went to the agency. Whatever one thinks of that rule, it had a property worth naming. It was predictable in direction. A regulated party reading an ambiguous statute could ask what the agency thought and expect, within the range of reasonable readings, that answer to govern.

Loper Bright removed the tiebreaker and put a judge in its place. Every ambiguous provision now has a best reading, to be found with traditional tools, and the agency's view is one input among several. That is a coherent account of the judicial role. It is also, necessarily, an account in which the meaning of an ambiguous provision is not settled until a court says what it is.

So who carries the doubt in the meantime?

Not the court. It resolves the question only when a case arrives. Not quite the agency. It still issues rules and enforces them; it simply knows each interpretation will be tested on the merits rather than for reasonableness. The doubt sits with whoever must act before the question is litigated. The regulated party deciding whether to comply. The lawyer advising it. The trial judge deciding the first case with no controlling answer. Each of them could once predict the resolution of an ambiguity by asking what the agency thought. Each must now predict what a court will call the best reading.

This is not an argument that Loper Bright was wrongly decided. It is an observation about mechanism. A rule that moves the burden from the challenger to the agency does not eliminate the uncertainty the old rule absorbed. It redistributes it, toward the parties who must act before any court has spoken.

And there is a party the doctrinal commentary rarely lists. The software.

Forty years in the training data

The draft asks what happens to legal AI when a framework that structured forty years of opinions disappears. Its answer borrows a term from machine learning: concept drift, change over time in the relationship a model was trained to predict. The draft classifies an overruling as the most severe variety, abrupt, real drift. Not a gradual change in which cases reach the courts. A discontinuous change in how the same cases are decided.

Consider what a model trained on Chevron-era administrative law opinions would have learned, on the draft's account. The linguistic and structural markers of the two-step analysis. The increased likelihood of agency victory once a statute was found ambiguous. The cues courts used to decide whether ambiguity existed. The importance of whether a case turned at Step One or Step Two. For that model, Chevron was a stable, high-frequency pattern, and the step determination predicted the outcome.

After June 28, 2024, opinions stop discussing Chevron and start emphasizing independent judgment and traditional tools. The draft lists what breaks. Label shift: similar facts now produce a different outcome. Feature importance shift: the signals that predicted agency success stop predicting it. Class imbalance: training data weighted toward agency wins meets a regime in which agencies may lose more often. Distribution shift: courts engage more heavily with interpretive canons, changing the features themselves.

The draft also offers a number. Agencies that won 70 to 80 percent of ambiguity cases might now win 40 to 50. Notice what kind of number that is. The draft labels it an empirical prediction and says outright that post-Loper Bright data will eventually show the true rate. It is a hypothesis, not a finding. That matters. A model that silently assumes the old rate and a commentator who confidently asserts a new one make the same error from opposite ends. Both turn uncertainty into a figure.

Who absorbs a stale model?

Follow the error downstream.

A research tool trained on the Chevron era does not announce that its training data describes a regime that no longer exists. It produces an answer, formatted with the same confidence as any other. Whether and when to retrain is the vendor's decision. Who pays for the period before it?

Not the vendor. The lawyer who relies on the output signs the brief. The client who relies on the lawyer absorbs the outcome. The court receives an argument framed around a standard it no longer applies. And the draft is blunt about the state of the field: most legal AI systems do not implement monitoring for drift. They are treated as static artifacts, not as systems operating in a world that changes.

This is the distributional point, and it is the doctrine's point at smaller scale. A system that assumes doctrinal stability is not neutral about the moment stability breaks. It assigns the cost of the break to whoever stands downstream of its output and cannot see inside it. The stability assumption looks like an engineering simplification. It is an allocation of risk.

The knowledge base that contradicts itself

Statistical models drift. Rule-based systems break differently.

A symbolic system that encoded Chevron as a rule, with the burden on the challenger at Step Two, must be revised when Chevron falls. The draft lays out the options. Simple deletion removes the rule, and with it the ability to analyze the cases decided under it; those opinions now appear to have no governing doctrine. Temporal scoping marks each rule with the period in which it applied, preserving history but requiring the system to know which period governs a query. Marking the rule superseded, with a pointer to what replaced it and when, leaves an audit trail.

Leave old and new rules side by side without scoping, and the knowledge base contradicts itself. One scenario defers to the agency because the statute is ambiguous. Another declines to defer despite the same ambiguity. Both rest on the same provision. An inconsistent knowledge base can derive anything.

Then the draft asks the question that shows how much depends on design: a case filed in 2019 and decided in 2024, which framework applies? For most administrative law cases, the draft suggests, the one in force when the case is decided. But a system must be built to ask before it can answer. A system never told that doctrine carries a date will not ask.

Not one overruling but a pattern

It would be comforting to treat Loper Bright as a singular event. The draft declines the comfort. In Dobbs v. Jackson Women's Health Organization, 142 S. Ct. 2228 (2022), the Court overruled Roe v. Wade, eliminating the undue burden framework that had governed abortion law for thirty years. In New York State Rifle & Pistol Association v. Bruen, 142 S. Ct. 2111 (2022), it replaced the means-ends test in Second Amendment cases with a historical tradition test. Then it overruled Chevron. These were not refinements. They changed how whole domains are analyzed.

The draft's conclusion is that legal AI cannot assume foundational doctrine is stable, and that the assumption of stability is itself the defect. The Institute has a name for the event. In Version Control for Law, an overruling is a force push: history on the main branch rewritten. A system with no representation of a force push keeps building on commits that no longer exist.

What legal AI must now do

The draft's framework for managing doctrinal disruption is concrete. Read each element as an answer to one question: how does the system make the uncertainty visible to the person who will bear it?

1. Temporal stratification. Train separate models for distinct doctrinal periods, a Chevron-era model and a post-Chevron model, and route each question to the model for its period. The cost is data and architecture. The benefit is calibration to the regime each model describes.
2. Doctrinal versioning. Encode each doctrine with explicit metadata: when it came into force, whether and when it was superseded, by what, and with what confidence. A question about 2020 and a question about 2025 should get different answers, and the system should know why.
3. Change detection. Parse new opinions for explicit doctrinal statements; an opinion that overrules Chevron should trigger a high-confidence signal. Track citation patterns, since citations to an overruled case fall away. Watch for language shifts, such as the disappearance of "Step One" and "Step Two." Calibrate to explicitness: an express overruling is not a subtle drift.
4. Dependency tracking. Know what depends on what. With Chevron gone, the draft notes, the major questions doctrine, Skidmore deference, and State Farm arbitrary-and-capricious review become primary frameworks. A system should propagate that change, not rediscover it one query at a time.
5. Meta-level monitoring. Track how often the Court overrules precedent. Track fragmentation: multiplying concurrences, dissents urging change. Track criticism in the literature and resistance in the lower courts, which the draft notes can precede an overruling. Rising instability should lower the system's confidence in its own doctrinal predictions.
6. Period-conditioned predictions. State the doctrinal period every prediction assumes, and offer scenarios rather than single points: the outcome if current doctrine holds, the outcome if it is replaced, and the likelihood of replacement. The draft's own illustrative percentages are just that, illustrative. The form is the point. A prediction that does not say which law it assumes has hidden its most important assumption.

None of this is exotic engineering. All of it is absent by default. A system that ships without these features is not simpler. It has decided that the user will carry what the system declined to represent.

Who holds the discretion now?

Discretion is never symmetrical. Ask who holds it.

Under Chevron, agencies held interpretive discretion within the range of reasonable readings. After Loper Bright, courts hold it, as independent judgment about the best reading. Agencies keep the power to persuade. Regulated parties keep the obligation to comply before anyone has told them what compliance means.

There is one more holder, and it is new. Whoever designs a legal AI system decides whether the system knows the date its doctrine carries, whether it reports the period its predictions assume, and whether it says so when it does not know. That decision is made once, upstream, by people who will never appear in the case. It determines whether uncertainty reaches the lawyer as a visible warning or as an invisible error.

The Court moved the burden of persuasion from the challenger to the agency. It did not move, and could not move, the burden of knowing which law applies. That burden sits with every lawyer who files, and now with every tool that drafts. The lawyer cannot delegate it. The tool should not be allowed to conceal it.

This essay continues the series begun in A Century of Shifting Burdens, which traces the frameworks this decision now joins. The same question, who absorbs the cost that a rule leaves unassigned, returns in the context of litigation budgets in Research the Carrier Won't Pay For.