Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Administrative Law's Fourth Settlement: AI and the Scrutable State

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The Supreme Court's recent administrative law retrenchment is best understood as a response to the scrutability crisis, and AI offers a path to restore oversight without sacrificing capability.

desk verdict A serious and honest legal-theory argument that the post-Loper Bright retrenchment is a comprehensibility-driven project and AI could enable a 'Fourth Settlement'; the diagnosis is plausible, the prescription is conditional on an AI auditability the author himself concedes is unproven. read the letter →

arxiv 2602.09678 v3 pith:GYXNKOZP submitted 2026-02-10 cs.CY cs.AI

Nicholas Caputo This is my paper · ORCID
classification cs.CYcs.AI
keywords capability-accountabilitytrapscrutabilityadministrativelawartificialintelligenceSupremeCourtretrenchmentdeferencetoauditModelandSystemDossier
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Since 1887, administrative law has faced a capability-accountability trap: new technologies force government to become more expert and complex, and that complexity makes agencies opaque to the courts, Congress, and the public. The paper's central claim is that the Supreme Court's recent dismantling of foundational administrative law doctrine is a coherent structural response to this opacity—faced with agencies it cannot comprehend, the Court is shrinking them back to comprehensible size, sacrificing capability to restore accountability. The paper then argues that AI can break this tradeoff: rather than adding another layer of inscrutable complexity, AI can act as scrutability infrastructure, translating technical agency reasoning into auditable form for overseers. Three doctrinal innovations are proposed—a Model and System Dossier extending the administrative record to AI, a material-model-change trigger for when AI updates need new process, and a 'deference to audit' standard rewarding agencies for demonstrable verification of their AI systems. If correct, the retrenchment is not a partisan anomaly but a symptom of a deeper institutional problem, and a concrete legal path exists to keep agencies capable while restoring oversight.

What carries the argument

The central object is the capability-accountability trap—the persistent tension between the expert, large-scale capability administration needs and the comprehensibility its overseers require—mediated by the concept of scrutability, the cognitive tractability of administrative action for courts, Congress, and the public. The proposal's load-bearing machinery is the Model and System Dossier, an expanded administrative record documenting AI system purpose, data provenance, performance, stress testing, monitoring, explainability, and change logs; the material-model-change trigger, which treats AI updates that alter outcomes, reasoning, populations, or architecture as new agency action; and the

What would settle it

A decisive test for the central account would compare Supreme Court decisions since 2020 against two rival predictors: the complexity/inscrutability of the agency action versus the party-alignment of the underlying policy; if partisan alignment explains the retrenchment cases better than scrutability, the paper's structural diagnosis fails. For the AI proposal, the falsifier is a demonstration that current audit artifacts (model cards, explainability outputs, monitoring logs) cannot reconstruct the actual drivers of a denial decision in a real agency adjudication—showing the safe harbor would

Watch

Extended reading notes

Core claim

The paper's positive claim is that the post-Loper Bright retrenchment—ending Chevron deference, expanding the major questions doctrine, and curtailing agency adjudication—is best understood as an attempt to make government 'scrutable' again: with agencies operating in domains that exceed judicial comprehension, the Court has chosen to reallocate authority to courts, Congress, and juries that it regards as comprehensible. That diagnosis is paired with a constructive claim: AI can reverse the historical pattern in which gains in administrative capability were bought at the cost of opacity. The author argues that AI, properly deployed, can translate technical complexity into accessible terms, s

Load-bearing premise

The prescriptive half depends on a technological premise: AI systems can be made reliable and auditable enough that the Model and System Dossier genuinely reconstructs their decision pathways—without hallucinated rationales or undetected drift—so that a deference-to-audit safe harbor certifies true oversight rather than paperwork.

Editorial extensions

If this is right

  • If the scrutable-state account is right, the recent retrenchment is a coherent structural response, not a partisan accident: the Court will keep shrinking agencies until administration is comprehensible, whether or not capability suffers.
  • Agencies that adopt the Model and System Dossier and defer-to-audit posture could retain AI capability while regaining judicial and congressional trust, creating a safe harbor for high-stakes automated decision-making.
  • The material-model-change trigger would solve the 'update problem' in algorithmic governance: not every model retraining requires full rulemaking, but updates that shift outcomes or reasoning do.
  • The same logic that justifies the Court's reallocation of interpretive authority to courts weakens if courts can verify agency reasoning through audit; the doctrine's pressure toward simplification would be relieved.
  • If the Fourth Settlement holds, procedural ossification from notice-and-comment, hard-look review, and cost-benefit analysis could decline, since substantive audit replaces procedure as the primary accountability mechanism.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The scrutable-state account implies a testable empirical claim that the Court's willingness to strike down agency actions tracks the technical opacity of the issue, not its partisan valence; if a case-clearing dataset shows party-aligned outcomes dominating complexity, the account would fail.
  • If deference to audit becomes doctrine, it could generalize beyond AI: any agency that subjects its human decision-making to comparable randomized audit and falsification could claim the same safe harbor, making audit a general currency of administrative legitimacy.
  • A concrete, testable extension is to pilot the Dossier on an existing high-volume adjudication system (e.g., benefits determinations) and measure whether its explanations and monitoring logs satisfy a blind review panel as 'scrutable'—providing a proof of concept before legal adoption.
  • The paper's own caveat cuts deep: if interpretability remains a future promise and hallucinated rationales stay common, the 'deference to audit' safe harbor could lend legality to systems whose real drivers are unknown—a risk that should shift the standard from 'deference' to 'presumption of scrutiny' until audits are proven.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper argues that American administrative law since 1887 has been structured by a 'capability-accountability trap': technological change forces government to become more expert and complex, which in turn makes it harder for courts, Congress, and the public to oversee. It identifies three historical 'settlements' (railroads; the New Deal; computers and complex science) that each rebalanced capability and accountability through allocation of authority, procedural review, and information-forcing. The paper then claims that the Supreme Court's post-Loper Bright retrenchment—Loper Bright, West Virginia v. EPA, SEC v. Jarkesy, and follow-on decisions—is best understood as an attempt to restore 'scrutability' by shrinking administration back to a size generalist overseers can comprehend. Finally, it proposes a 'Fourth Settlement' in which AI, paired with a Model and System Dossier extending the administrative record, a material-model-change trigger, and a 'deference to audit' review posture, could allow government to retain capability while restoring auditable oversight. The paper is a doctrinal and historical synthesis; it makes no quantitative predictions and offers no empirical test of its central motivational claim.

Significance. If the descriptive thesis is correct, the paper provides a genuinely novel structural account of recent administrative law: it treats the Supreme Court's decisions as a coherent, if misguided, project of restoring comprehensibility rather than as purely partisan retrenchment. The prescriptive half is also significant, offering concrete doctrinal hooks—the Dossier, the change trigger, deference to audit—that could be implemented or tested. The paper is explicitly honest about its own limits: Part II.C concedes uncertainty about the Court's motivations, and Part III.C.7 candidly acknowledges hallucinated rationales and the immaturity of interpretability. It also builds transparently on prior constructs (Vermeule's deference dilemma, Scott's legibility, Simon's bounded rationality) and engages primary legal materials extensively. There is no fitted-value circularity, because the paper makes no quantitative claims. The main significance risk is that the prescriptive framework's feasibility is asserted rather than demonstrated, and the paper's own concessions undercut that feasibility.

major comments (3)
  1. [Abstract; Part II.C] The abstract asserts flatly that the Supreme Court's retrenchment 'can be understood as a response to the scrutability crisis,' but Part II.C concedes that 'It is unclear exactly why the Court has decided at this moment' and lists partisan and capture-based alternatives as plausible. The paper offers no discriminating evidence—for example, no analysis of whether the Court's decisions track complexity or instead track political valence. Since the descriptive claim is half of the article's contribution, the abstract should be qualified, and the body should either supply a testable implication or explicitly frame the claim as one plausible interpretation among several.
  2. [Part III.C.4 and III.C.7] The 'deference to audit' safe harbor presumes that AI systems can be made auditable in the strong sense that the Dossier's explanations reflect the actual drivers of agency decisions. The paper itself concedes in Part III.C.7 that AI systems 'can generate plausible-sounding explanations that do not reflect the actual drivers of an output or decision,' and Part III.A describes interpretability as a future possibility ('we may one day be able to examine the decision pathways of an AI system'). If explanations are systematically disconnected from actual reasoning, the Dossier becomes a record of plausible fictions, and deference to audit would certify unreliable systems rather than make them scrutable. The manuscript needs to either specify minimum audit standards and independent verification protocols, or condition the proposal on demonstrated auditability. As written, the prescriptive fra
  3. [Part III.C.3] The material-model-change trigger is a central doctrinal innovation, but its operation depends on 'some substantial defined threshold' for outcome effects and on interpretability tools for 'reasoning effects' that are not yet available. The paper provides no default threshold, no method for setting one, and no worked example of when a retraining would or would not trigger new process. Because the trigger determines when agencies must update the Dossier and face new procedural obligations, this vagueness is load-bearing: agencies cannot know their obligations and courts cannot review compliance. At a minimum, the paper should propose a presumptive threshold (e.g., a percentage change in approval rates or a specified divergence in feature-attribution metrics) and discuss how it would be calibrated over time.
minor comments (4)
  1. [Abstract] The phrase 'The result a "Fourth Settlement"' is missing the verb 'is.'
  2. [Title / header] The running title in the full text says 'AI and the Capability-Accountability Trap,' while the arXiv metadata gives 'AI and the Scrutable State.' Please unify the title.
  3. [Part III.C.7, fn. 251] The citation to Zhang et al., 'Siren's Song in the AI Ocean,' lists the date as 'Sep. 14, 2025' but the arXiv identifier 2309.01219 corresponds to September 2023. The date appears to be a typo.
  4. [Part II.C] The phrase 'the Court is responding by trying to shrink government back to a size it can understand' is vivid but somewhat ambiguous: is the claim about the size of the administrative state or about the complexity of individual decisions? Clarifying this distinction would sharpen the descriptive thesis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's historical/doctrinal account and AI proposal are built openly from external sources and primary legal materials, with no fitted inputs, self-citation chains, or definitional reductions.

full rationale

This paper does not exhibit circular derivation. The central positive claim—that the Supreme Court's post-Loper Bright retrenchment is best understood as a response to a scrutability crisis—is an interpretive argument grounded in primary legal materials (Loper Bright, Jarkesy, West Virginia v. EPA) and in an explicitly borrowed conceptual vocabulary (Vermeule's deference dilemma; Scott's legibility inverted). It is not derived from any equation or fitted parameter, and the paper candidly acknowledges competing explanations, noting that the retrenchment 'may also be politically motivated' and that the framework 'suggests a structural explanation' rather than a logical necessity. The prescriptive half, proposing a Model and System Dossier, a material-model-change trigger, and deference to audit, is conditional on the feasibility of AI auditability. The paper itself flags the fragility of that premise, conceding that AI systems 'can generate plausible-sounding explanations that do not reflect the actual drivers of an output or decision' and describing interpretability as a future possibility. That is an empirical/technological risk and a possible internal tension, but it is not circularity: the proposal does not assume the conclusion it is meant to establish by defining its target in terms of its inputs. There are no self-citations doing load-bearing work, no fitted values renamed as predictions, and no uniqueness theorem imported from the author's prior work. The argument is self-contained in the relevant sense: its historical narrative and doctrinal proposals can be checked against cases and external sources, and its feasibility condition is openly stated rather than smuggled in. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 5 invented entities

No numerical free parameters are fitted anywhere in this paper — it is a legal-theory work with no quantitative claims. The 'free parameters' listed are the deliberately open thresholds and design choices in its proposed doctrines, which would have to be fixed by agencies or courts. The axioms are the load-bearing premises the framework pulls in without proof: a cognitive model of overseers (Simon), the interpretive premise about judicial motivation (which the author himself flags as uncertain), the claim that AI can be made auditable (which the author flags via failure modes), the ossification diagnosis, and a stylized periodization of legal history (acknowledged as 'pattern identification'). The invented entities are concepts and doctrinal instruments: 'scrutability' and the 'capability-accountability trap' are anchored in prior literature and checkable case readings; the 'Model and System Dossier' is anchored in external artifacts (model cards, datasheets, EU AI Act); 'deference to audit' and the 'material-model-change trigger' are new proposals with no existing implementation.

free parameters (3)
  • Material-model-change threshold
    Part III.C.3: the trigger fires when an AI update moves outcome/reasoning/population metrics 'by more than some substantial defined threshold.' The threshold is deliberately left as a standard rather than a rule; agencies and courts would have to fix it, and the doctrine's bite depends entirely on that choice.
  • Central-review impact threshold
    Part III.C.5: centralized review is reserved for systems affecting 'large numbers of people,' 'high-stakes determinations,' or 'sensitive domains.' The operational cutoffs are unspecified design parameters of the proposed regime.
  • Capability-accountability tradeoff curve
    Parts I-II: the paper's central conceptual device is a curve whose shape is asserted — procedural accumulation pushes both capability and accountability down — but it is never mapped to measurable institutional variables. The curve's shape does the causal work of the 'double deficit' claim in Part II.
assumptions (5)
  • domain assumption Overseers are boundedly rational: comprehension capacity is the binding constraint on accountability.
    The entire scrutability framework rests on Simon's bounded-rationality model, invoked extensively in Part III.A ('a scissors whose blades are the structure of task environments and the computational capabilities of the actor'). The cognitive premise is imported from prior literature without independent verification in this paper.
  • domain assumption The Supreme Court's retrenchment is substantially motivated by comprehensibility concerns, not only by politics.
    Part II.C: this is the paper's diagnostic thesis. The paper itself concedes 'It is unclear exactly why the Court has decided at this moment to undo the doctrinal settlement' and lists partisanship and capture beliefs as alternative explanations. If the motive is purely political, the 'scrutability crisis' explanation fails.
  • domain assumption AI systems can be made reliable and auditable enough for legally load-bearing use (low hallucination, effective interpretability, drift monitoring).
    Part III.C's Dossier, 'deference to audit,' and material-change trigger all presume this capability. The paper flags the uncertainty itself: Part III.C.7 discusses hallucinated rationales as an unresolved failure mode, and Part III.A describes interpretability as a future possibility ('we may one day be able to examine the decision pathways').
  • domain assumption Procedural accumulation has ossified government and degraded both capability and accountability.
    Part II.A's diagnosis of the 'sedimentary state' and 'scrutability tax.' Asserted via McGarity, Bagley, and Pierce, without engagement with the empirical literature that tests the ossification thesis; the paper cites Pierce's rebuttal (fn 152) but not the target study or the broader quantitative debate.
  • ad hoc to paper Administrative-law history is periodizable into three technology-driven settlements (railroads; New Deal; computers and complex science).
    Parts I.A-C: a stylized tripartite schema chosen to fit the framework. The author acknowledges it is 'pattern identification' rather than full explanation and that politics was not secondary (fn 53). The periodization organizes the very history the paper then uses to validate the framework.
invented entities (5)
  • Scrutability (inverse of legibility) independent evidence
    purpose: Axis along which administrative action is cognitively tractable to overseers; the concept that unifies the historical narrative and motivates AI's role.
    Anchored in checkable legal materials (the readings of Loper Bright, Jarkesy, West Virginia v. EPA) and built explicitly from Scott's 'legibility' and Selbst & Barocas's 'inscrutability' (fn 19), so its application can be verified against primary sources.
  • Capability-accountability trap independent evidence
    purpose: Organizing construct for the entire argument; names the tradeoff that three settlements managed and AI is claimed to escape.
    Built openly on Vermeule's 'deference dilemma' (fn 15) and prior tradeoff formulations (Bednar, Kagan, Stephenson, McCubbins et al.), so its instances are independently documented even though the synthesis is new.
  • Model and System Dossier independent evidence
    purpose: Proposed expansion of the administrative record documenting AI systems (identity, data provenance, evaluation, stress testing, monitoring, explainability, change log, governance, vendors).
    Anchored in existing external artifacts cited in Part III.C.2: model cards (Mitchell et al.), datasheets (Gebru et al.), EU AI Act Code of Practice, and company Frontier Safety Frameworks — its components are already in use outside the paper.
  • Deference-to-audit standard
    purpose: Proposed judicial review posture in which agencies earn a safe harbor by documenting and auditing AI systems rather than by claiming expertise.
    A normative proposal with no existing doctrinal implementation; its viability depends on the unproven auditability of AI, which the paper's own Part III.C.7 failure modes call into question.
  • Material-model-change trigger
    purpose: Doctrinal rule specifying when an AI update constitutes agency action requiring renewed process (analogized to the 'logical outgrowth' test).
    New proposal with no implemented analogue. Its operation depends on interpretability tools to detect 'reasoning effects' — a capability the paper describes only prospectively.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Administrative Law's Fourth Settlement: AI and the Scrutable State." pith.science (2026). https://pith.science/paper/GYXNKOZP

@misc{pith2026260209678,
  author       = {Pith},
  title        = {Pith review of: Administrative Law's Fourth Settlement: AI and the Scrutable State},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GYXNKOZP}},
  note         = {Machine review of arXiv:2602.09678}
}
read the original abstract

Since 1887, administrative law has confronted a problem of institutional cognition. Expert agencies are needed to govern technologically complex systems, but expertise makes agency decisions difficult for courts, Congress, and the public to understand and oversee. Administrative law has responded to this "capability-accountability trap" by requiring records, reason-giving, and transparency, drawn together through procedural review. These devices have preserved legality but have piled up, making government both less comprehensible and less effective. This Article offers a new account of the Supreme Court's recent administrative law retrenchment, rooted in problems of institutional structure and information-processing. From Loper Bright through Trump v. Slaughter, the Court has reallocated authority to entities it regards as comprehensible and attributable. It is attempting to restore accountability by making government "scrutable," comprehensible to its overseers and the public, but in doing so it is sacrificing capability and undermining the effectiveness of administration. AI offers a different path. Deployed correctly, AI could help make government both more effective and more transparent, translating technical complexity into accessible terms, surfacing assumptions, and enabling substantive verification of agency reasoning. This technical integration must be accompanied by updated administrative law, built around a Model and System Dossier that extends the administrative record to AI decision-making; a material-model-change trigger specifying when AI updates require new process; and a deference to audit standard that rewards agencies for auditable evaluation of AI uses. The result a "Fourth Settlement," administrative law that escapes the capability-accountability trap by preserving capability while restoring comprehensible oversight of administration.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A comment on the number of $k$-th powers inside arithmetic progressions

    math.NT 2026-07 conditional novelty 3.0 of 10

    For an arithmetic progression of length N whose step q is at most N^r, the number of k-th powers among its terms is O(N^{1/k+ε}) for large N.

Reference graph

Works this paper leans on

12 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Notice-and-comment thus becomes dominated by repeat players who can invest in technical engagement rather than operating as a true form of democratic input

    ts and don’t have the ability to understand the proposal or how it might affect them. Notice-and-comment thus becomes dominated by repeat players who can invest in technical engagement rather than operating as a true form of democratic input. AI can help address these problems. Consider a layered AI interface that enables citizen participation while prese...

    2025

  2. [2]

    Where rulemaking produces generally applicable standards, adjudication applies standards to individual cases

    Adjudication: Procedural Justice at Scale Administrative adjudication presents different challenges than rulemaking. Where rulemaking produces generally applicable standards, adjudication applies standards to individual cases. The due process stakes are immediate and personal because they affect things like the disability benefits or immigration status of...

    2009

  3. [3]

    information asymmetry

    Congressional Oversight: AI as Attention Multiplier Congress is structurally incapable of overseeing the administrative state. Members have limited time, divided across hundreds of policy areas and the demands of campaigning and constituent service.253 They have small staffs who might be able to help with certain areas, but cannot dive into more than a fe...

    2017

  4. [4]

    The same tools that help citizens understand rulemakings could help bad actors flood agencies with synthetic comments

    Safeguards: Preventing AI-Assisted Participation from Becoming Manipulation AI-enhanced participation creates new risks that require new safeguards. The same tools that help citizens understand rulemakings could help bad actors flood agencies with synthetic comments. If agencies are not aligned to the public interest, the interfaces that translate complex...

    2023

  5. [5]

    Institutional Architecture: Who Audits? Deference to audit presupposes auditors. Courts can verify that agencies have produced documentation and followed specified procedures, but they cannot verify that the documentation is accurate or that the procedures are adequate. Technical scrutiny of AI systems requires technical capacity that courts lack and cann...

    arXiv 2023

  6. [7]

    Model and System Dossier

    The AI Administrative Record: A Model and System Dossier Adapting administrative accountability for AI requires an expanded administrative record that documents how AI systems are designed, deployed, and monitored. This “Model and System Dossier” would be a required addendum to the traditional administrative record that is required whenever an agency adop...

    2025

  7. [8]

    Post-Loper Bright, courts exercise independent judgment on statutory interpretation

    Deference to Audit: A New Review Posture How should courts review AI-assisted agency action? Not by deferring to agency expertise, as they have traditionally done, nor by evaluating algorithmic details directly, but by assessing whether agencies have subjected their AI systems to rigorous technical audits. Post-Loper Bright, courts exercise independent ju...

    2017

  8. [47]

    25, 2016)

    286 Dario Amodei et al., Concrete Problems in AI Safety, arXiv:1606.06565 (Jul. 25, 2016). 287 Lauro Langosco et al., Goal Misgeneralization in Deep Reinforcement Learning, 162 PMLR 12004 (2022). 288 Ryan Greenblatt et al., Alignment faking in large language models, arXiv:2412.14093 (Dec. 18, 2024). ADMINISTRATIVE LAW’S FOURTH SETTLEMENT [26-Mar-26 64 sen...

    arXiv 2016

Show all 12 references
  1. [166]

    West, It is time to restore the US Office of Technology Assessment, BROOKINGS (Feb

    259 See Darrell M. West, It is time to restore the US Office of Technology Assessment, BROOKINGS (Feb. 10, 2021), https://www.brookings.edu/articles/it-is-time-to-restore-the-us-office-of-technology-assessment/. 260 See Statement from U.S. Secretary of Commerce Howard Lutnick ...

    2021 arXiv

  2. [241]

    14, 2025)

    251 See Yue Zhang et al., Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models, arXiv:2309.01219 (Sep. 14, 2025). 252 See Modernizing Unauthorized Practice of Law Regulations to Embrace AI-Driven Solutions and Improve Access to Justice, AI POLICY CO...

    2025 arXiv

  3. [251]

    the model determined

    26-Mar-26] ADMINISTRATIVE LAW’S FOURTH SETTLEMENT 61 61 determinative or that cites evidence that does not exist in the record. Agencies and courts may credit these explanations without recognizing their unreliability. Mitigation requires grounding explanations in verifiable r...

    2019

  4. [259]

    § 706(2) (2018)

    275 5 U.S.C. § 706(2) (2018). ADMINISTRATIVE LAW’S FOURTH SETTLEMENT [26-Mar-26 60 Vacatur versus remand. Vacatur sets aside unlawful agency action while remand returns the matter to the agency for further proceedings. For AI-assisted adjudication at scale, vacatur may be impr...

    2018

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.