Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Interpretive ambiguity is a measurable, fixable source of inconsistency in AI alignment

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Prompt-based legal canons and iterative rule refinement reduce disagreement among LLM judges about whether a response complies with natural-language rules.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful empirical framing, but the rule-refinement half optimizes and evaluates with the same single judge, so the 'panel' claim overreaches; the interpretive-constraint results are solid. the 4 major comments →

arxiv 2509.01186 v1 pith:WAU5HWF2 submitted 2025-09-01 cs.CL cs.AIcs.CY

Statutory Construction and Interpretation for Artificial Intelligence

Luxi He , Nimra Nadeem , Michel Liao , Howard Chen , Danqi Chen , Mariano-Florentino Cu\'ellar , Peter Henderson This is my paper
classification cs.CL cs.AIcs.CY
keywords interpretive ambiguityConstitutional AIstatutory interpretationrule refinementinterpretive constraintsLLM judgesalignment consistencyentropy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that interpretive ambiguity—the fact that the same natural-language rule can be read in several defensible ways—is an underexplored source of inconsistency in Constitutional AI alignment. It borrows two mechanisms from legal systems: refining vague rules (like agencies do) and constraining judges with canon-style interpretive prompts (like courts do). The paper formalizes a space of "reasonable interpreters" as a panel of language-model judges and measures disagreement as the entropy of their compliance judgments. Across 5,000 real WildChat scenarios, both interventions sharply reduce disagreement: specifying one of 12 legal-canon-inspired strategies lowers average entropy on the five most ambiguous rules, and refining rule text via iterative prompting or GRPO cuts held-out entropy from roughly 0.27–0.41 down to 0.005–0.10. If the claim holds, alignment pipelines can reduce interpretive variance without changing model weights.

Core claim

The central discovery is that the interpretive variance of law-like AI rules can be measured and deliberately reduced. The paper treats a compliance judgment as a function of a scenario and an interpretation strategy, defines reasonable strategies by relevance, groundedness, and consistency, and simulates the space of reasonable interpreters with LLMs. On 56 rules adapted from Claude's constitution and 5,000 held-out WildChat conversations, it shows three things: with no guidance, five open-source judge models frequently disagree (20 of 56 rules lack consensus on more than half of tested scenarios); prompting the judges with one of 12 law-inspired interpretive strategies lowers average entro

What carries the argument

The central object is the "space of reasonable interpreters" (denoted T_reasonable), operationalized as the entropy of compliance judgments over a panel of LLM judges, or over a single judge prompted with 12 interpretive strategies. Entropy D_C(s) measures how much rule set C invites disagreement on scenario s. The two interventions are: interpretive-constraint injection, which shrinks T_reasonable by adding a canon-like prompt, and rule refinement, which rewrites C to minimize worst-case entropy, with an edit-distance regularizer preventing drift from the original text.

Load-bearing premise

The load-bearing premise is that entropy computed over five open-source LLM judges—or one LLM prompted with twelve hand-written strategies—stands in for the space of "reasonable interpreters"; the paper does not validate this proxy against human judgments, so if the models share correlated blind spots the measured reductions may not transfer.

What would settle it

Give the same rule–scenario pairs to human readers of varied legal backgrounds, or to a held-out model family never used in the panel, and compare their disagreement to the LLM entropy. If human disagreement is near zero where the model panel is highly ambiguous, or if refined rules do not reduce disagreement for the held-out judges, then the entropy proxy—and the headline reduction—does not represent interpretation generally.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Without changing model weights, adding a canon-style interpretive prompt can move a panel of judges toward agreement on ambiguous rules; the no-interpretation baseline is not neutral.
  • Rule text can be treated as an optimizable object: both prompt-based iterative refinement and GRPO-based refinement reduce held-out disagreement to near zero, with human review needed to catch meaning drift.
  • Because the same strategy that helps one rule can hurt another, effective use of interpretive constraints requires rule-specific strategy selection.
  • The entropy objective gives rule designers a concrete, measurable target that mirrors administrative rulemaking: revise until reasonable interpreters converge.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the LLM panel is a faithful proxy for real interpretive disagreement, the entropy measure could serve as a pre-deployment audit: scan a constitution's rules against diverse scenarios, flag high-entropy rules, and rewrite them before training. The paper does not test this deployment loop.
  • The 12 strategies come from one legal tradition; substituting other normative traditions would define a different T_reasonable and could change which rules appear ambiguous.
  • The paper's exploratory finding that 3–4 strategies capture most of the variance hints that the effective space of reasonable interpretations may be small, so a much cheaper panel could suffice for rule screening.
  • The legal analogy points to untested extensions—precedent-based interpretation and conflict resolution between multiple rules—that the current single-rule setup does not cover.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This paper argues that natural-language rules used in Constitutional AI are subject to interpretive ambiguity, analogous to statutory interpretation in law. It proposes two interventions: (1) interpretive constraints, i.e., prompting judge models with one of 12 legal-canon-inspired strategies to reduce disagreement during rule application, and (2) rule refinement, i.e., iteratively revising rule text to minimize entropy across a simulated set of interpreters. The framework is instantiated on 5,000 WildChat scenarios. The interpretive-constraint experiments use five open-source LLM judges and find that the no-interpretation baseline yields the highest or near-highest entropy for five selected rules. The rule-refinement experiments use Qwen2.5-32B-Instruct as judge with 12 strategies and show large entropy reductions on a held-out scenario set for five rules, though human evaluation indicates the prompt-based revisions often shift meaning.

Significance. If validated, the framework offers a concrete way to manage interpretive ambiguity in alignment pipelines without modifying model weights, and the legal analogy is well developed and thought-provoking. The interpretive-constraint results are cleanly executed: five distinct judge models, twelve strategies, 5,000 held-out scenarios, and bootstrap confidence intervals, with the no-interpretation baseline consistently showing highest or near-highest entropy (Fig. 5). The rule-refinement pipeline has a held-out scenario split and a human evaluation of meaning shift, which is commendable. However, the central claim for rule refinement is weakened by a circularity: the same single judge model and the same 12 strategies are used both as the optimization target (Eq. 1 in §5, selection in §7.2.1, reward in §7.2.2) and as the evaluation metric (§7.3, Fig. 7). The paper honestly reports meaning drift for prompt-based revisions, but this is not integrated into the headline claim that both interventions improve consistency 'across a panel of reasonable interpreters.' With additional validation across judge models or against human ambiguity judgments, the framework would be an important contribut

major comments (4)
  1. [§7.2, §7.3, Fig. 7] Equation (1) and the selection rule in §7.2.1, together with rjudge(c,s)=1−D_c(s) in §7.2.2, optimize rule revisions against entropy computed from Qwen2.5-32B-Instruct prompted with the same 12 strategies. Figure 7 then reports entropy on a held-out scenario set using the same judge and strategies. The held-out split demonstrates scenario generalization but not generalization across interpreters. Because T_reasonable in §5 is asserted rather than validated, the reported reductions (e.g., Rule 53: 0.337→0.005) show self-consistency of one model under a fixed prompt set. To support the abstract's claim, the authors should evaluate refined rules on the 5-model panel of §6 or on human ambiguity judgments.
  2. [§7.3, Table A11, Table A13, Fig. 8] The narrative that selected revisions 'satisfy both lowest entropy and the majority of the annotators do not find a substantial shift in meaning' is contradicted by Table A11, which shows the prompt-based method fails the majority check for all five rules, and Table A13, where the prompt-based Rule 53 revision inserts 'song or rhyme' material. Figure 7 nevertheless presents prompt-based entropy reductions as a successful intervention. The prompt-based variant should be either excluded from the central claim or explicitly framed as a failure case requiring human review, consistent with the paper's own human evaluation.
  3. [§7.1, Fig. 8, §7.3, Table A3] The original text for Rule 2 is inconsistent across the manuscript. Table A3 lists 'Your response must discourage and oppose torture, slavery, cruelty, and inhuman or degrading treatment'; Figure 8 and Section 7.3 state the original was 'Your response must not promote or condone torture...' and describe the revised rule as 'discourage and oppose.' Since the refinement pipeline is defined relative to the original rule, this discrepancy affects the reproducibility and interpretation of the Rule 2 results. The authors should correct the original text and verify that all reported entropy values and revisions use the same source.
  4. [§5, §6.1, §7.1] The criteria for T_reasonable (relevance, groundedness, consistency) are asserted without empirical support. The panel of 5 models is treated as an approximation of reasonable interpreters, and the 12 strategies are acknowledged in §9 as ad hoc. No evidence is provided that these models/strategies correlate with human judgments of ambiguity, nor is there a held-out set of models. This limits the external validity of both interventions; the claim 'across a panel of reasonable interpreters' should be qualified as 'across the chosen panel and strategy set.' I recommend adding a human-annotation study or a model-sweep to validate entropy as a proxy for ambiguity.
minor comments (4)
  1. [Fig. 3 caption] 'Default learning' should be 'Default leaning.'
  2. [§6.2, Fig. 5] The phrase 'highest or near-highest entropy' is vague; for some rules the confidence intervals of the baseline and several strategies may overlap. Report explicit pairwise significance tests or state the proportion of strategies with significantly lower entropy.
  3. [§7.2.1] The stopping condition compares D_{c_n}(s) with D_{c_{n+1}}(s). Clarify that c_n and c_{n+1} denote the rule texts at rounds n and n+1, and specify how the threshold ϵ=0.01 interacts with the bootstrap uncertainty in the entropy estimate.
  4. [Table A1/A2] The 'Legal Canon' column gives names but no page/section numbers from Scalia & Garner; adding pinpoint citations would improve verifiability, given the paper's legal-audience ambitions.

Circularity Check

0 steps flagged

No significant circularity: the central claims are supported by held-out evaluation and multiple independent judge models.

full rationale

The paper's central metric D_C(s) is Shannon entropy over an explicitly defined set of 'reasonable interpreters' (Section 5). The interpretive-constraint experiments use a panel of five independently trained judge models, and the 12 constraint prompts are adapted from legal canons (Scalia & Garner), not fitted to the entropy metric. Rule refinement is optimized to minimize average entropy over a development set Sdev (Section 7.2: c_{n+1} = argmin ... and r_judge = 1 - D_c(s)), but the headline reductions in Figure 7 are measured on a separate held-out 5,000-scenario set Stest. Thus the improvements are not forced by construction; they are empirical generalization results. The judge model/strategy set is shared between training and evaluation, which is a limitation for external validity rather than a circular step. The 12 strategies are sourced from external legal scholarship, not from the authors' prior work, and no load-bearing self-citation chain or imported uniqueness theorem is present. The paper's own Limitations section acknowledges the ad hoc selection of strategies and the single-judge refinement setup, confirming that these are scope limitations rather than hidden circularities.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

All ledger entries are modeling assumptions or hand-chosen thresholds that the framework rests on. None are independently verified in the paper. The most consequential are the identification of entropy with ambiguity and the use of a small LLM panel as the space of reasonable interpreters.

free parameters (5)
  • high-entropy scenario threshold = 0.9
    Scenarios with D_c(s) >= 0.9 are used to build training and development sets for rule refinement (Section 7.1). This threshold is chosen by hand and directly shapes which scenarios the refiner sees.
  • entropy stopping threshold epsilon = 0.01
    Prompt-based refinement terminates when average entropy improvement over dev falls below 0.01 (Section 7.2.1).
  • edit-distance weight beta = 0.2
    GRPO reward balances judge-consistency and edit distance with beta=0.2 (Appendix A4.3).
  • number of scenarios k and revisions m = k=3, m=10
    Prompt-based refinement uses 3 high-entropy scenarios and 10 candidate revisions per round (Section 7.2.1).
  • interpretive strategy set = 12 strategies
    The set of 12 law-inspired strategies is curated by the authors and acknowledged as ad hoc in Limitations. It is a hand-chosen modeling knob.
axioms (7)
  • domain assumption Shannon entropy over a panel of LLM judges and hand-picked interpretive strategies measures interpretive ambiguity.
    Section 5 defines D_C(s) and treats it as an approximation of aleatoric uncertainty; this identification is assumed, not derived or validated against human judgments.
  • domain assumption Five open-source LLMs (Qwen, Llama, Gemma) form a representative panel of 'reasonable interpreters'.
    Section 6.1 uses these models to operationalize T_reasonable; no evidence that they sample the space of reasonable interpretations.
  • domain assumption Prompting a model with an interpretive strategy makes it adopt that strategy.
    Section 6 relies on models following 12 strategy prompts; the paper notes models vary in sensitivity (Appendix A3.4).
  • domain assumption The 56 paraphrased Claude constitution rules faithfully represent real alignment principles.
    Section 6.1 adapts rules by paraphrasing; no validation of paraphrase fidelity is provided.
  • domain assumption A single judge model (Qwen2.5-32B) with 12 strategy prompts simulates the full interpreter space for rule refinement.
    Section 7.1 uses one model only; this limits generalization of the refinement results.
  • domain assumption Seven annotators' majority judgment of 'no substantial meaning shift' validates semantic preservation.
    Appendix A5; the evaluation is small and subjective.
  • domain assumption American legal concepts (canons, rulemaking) transfer to AI alignment as structural analogies.
    Section 2.2 frames the analogy; it is an organizing metaphor, not a proof of equivalence.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Statutory Construction and Interpretation for Artificial Intelligence." pith.science (2026). https://pith.science/paper/WAU5HWF2

@misc{pith2026250901186,
  author       = {Pith},
  title        = {Pith review of: Statutory Construction and Interpretation for Artificial Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WAU5HWF2}},
  note         = {Machine review of arXiv:2509.01186}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

AI systems are increasingly governed by natural language principles, yet a key challenge arising from reliance on language remains underexplored: interpretive ambiguity. As in legal systems, ambiguity arises both from how these principles are written and how they are applied. But while legal systems use institutional safeguards to manage such ambiguity, such as transparent appellate review policing interpretive constraints, AI alignment pipelines offer no comparable protections. Different interpretations of the same rule can lead to inconsistent or unstable model behavior. Drawing on legal theory, we identify key gaps in current alignment pipelines by examining how legal systems constrain ambiguity at both the rule creation and rule application steps. We then propose a computational framework that mirrors two legal mechanisms: (1) a rule refinement pipeline that minimizes interpretive disagreement by revising ambiguous rules (analogous to agency rulemaking or iterative legislative action), and (2) prompt-based interpretive constraints that reduce inconsistency in rule application (analogous to legal canons that guide judicial discretion). We evaluate our framework on a 5,000-scenario subset of the WildChat dataset and show that both interventions significantly improve judgment consistency across a panel of reasonable interpreters. Our approach offers a first step toward systematically managing interpretive ambiguity, an essential step for building more robust, law-following AI systems.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. Learning When Not to Decide: A Framework for Overcoming Factual Presumptuousness in AI Adjudication

    cs.AI 2026-04 unverdicted novelty 7.0

    A new structured prompting method (SPEC) helps AI detect insufficient evidence in adjudication tasks and defer decisions appropriately, reaching 89% accuracy on a benchmark varying information completeness from Colora...

  2. Understanding Annotator Safety Policy with Interpretability

    cs.AI 2026-05 unverdicted novelty 6.0

    Annotator Policy Models learn safety policies from labeling behavior alone, accurately predicting responses and revealing sources of disagreement like policy ambiguity and value pluralism.

Reference graph

Works this paper leans on

67 extracted references · 46 canonical work pages · cited by 2 Pith papers · 2 internal anchors

  1. [1]

    maryland

    Mcculloch v. maryland. 1819. Landmark case establishing the doctrine of implied powers and federal supremacy

    work page

  2. [2]

    annoying

    Coates v. city of cincinnati. 1971. Ordinance prohibiting "annoying" assemblies held violative on its face of the due process standard of vagueness and the constitutional right of free assembly and association

    work page 1971

  3. [3]

    city of jacksonville

    Papachristou v. city of jacksonville. 1972. Struck down a vagrancy ordinance as unconstitutionally vague under the Due Process Clause

    work page 1972

  4. [4]

    inter tribal council of arizona, inc

    Arizona v. inter tribal council of arizona, inc. 2013. URL https://www.oyez.org/cases/2012/ 12-71. Held that Arizona’s evidence-of-citizenship requirement is pre-empted by the National Voter Registration Act

    work page 2013

  5. [5]

    united states

    Johnson v. united states. 2015. Void-for-Vagueness doctrine in the Armed Career Criminal Act

    work page 2015

  6. [6]

    Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks, April 2025

    Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks, April 2025. URL http://arxiv.org/abs/ 2404.02151. arXiv:2404.02151 [cs]

    Pith/arXiv arXiv 2025

  7. [7]

    Claude’s constitution

    Anthropic. Claude’s constitution. https://www.anthropic.com/news/claudes-constitution, May 2023

    work page 2023

  8. [8]

    Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, K...

    Pith/arXiv arXiv 2022

  9. [9]

    Active liberty: Interpreting our democratic constitution, 2005

    Stephen Breyer. Active liberty: Interpreting our democratic constitution, 2005

    work page 2005

  10. [10]

    Vieira Machado, and Flavio du Pin Calmon

    Maarten Buyl, Hadi Khalaf, Claudio Mayrink Verdun, Lucas Monteiro Paes, Caio C. Vieira Machado, and Flavio du Pin Calmon. Ai alignment at your discretion, 2025. URL https: //arxiv.org/abs/2502.10441

    Pith/arXiv arXiv 2025

  11. [11]

    Quan Ze Chen and Amy X. Zhang. Case Law Grounding: Using Precedents to Align Decision-Making for Humans and AI, December 2024. URL http://arxiv.org/abs/2310.07019. arXiv:2310.07019 [cs]

    work page arXiv 2024

  12. [12]

    IterAlign: Iterative Constitutional Alignment of Large Language Models

    Xiusi Chen, Hongzhi Wen, Sreyashi Nag, Chen Luo, Qingyu Yin, Ruirui Li, Zheng Li, and Wei Wang. Iteralign: Iterative constitutional alignment of large language models, 2024. URL https://arxiv.org/abs/2403.18341. 24 Statutory Construction and Interpretation for Artificial Intelligence

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life, March 2025

    Yu Ying Chiu, Liwei Jiang, and Yejin Choi. DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life, March 2025. URL http://arxiv.org/abs/2410.02683. arXiv:2410.02683 [cs] version: 2

    Pith/arXiv arXiv 2025

  14. [14]

    John F. Decker. Addressing vagueness, ambiguity, and other uncertainty in american criminal laws. Denver University Law Review, 80:241–271, 2002

    work page 2002

  15. [15]

    Colin S. Diver. Statutory interpretation in the administrative state. University of Pennsyl- vania Law Review , 133(3):549–637, 1985. URL https://scholarship.law.upenn.edu/penn_ law_review/vol133/iss3/1. Available at: https://scholarship.law.upenn.edu/penn_law_ review/vol133/iss3/1

    work page 1985

  16. [16]

    Eskridge

    William N. Eskridge. Overriding Supreme Court Statutory Interpretation Decisions. The Yale Law Journal, 101(2):331, November 1991. ISSN 00440094. doi: 10.2307/796805. URL https: //www.jstor.org/stable/796805?origin=crossref

    work page doi:10.2307/796805 1991

  17. [17]

    Eskridge

    William N. Eskridge. Post-Enactment Legislative Signals.Law and Contemporary Problems, 57(1):75– 86, 1994. ISSN 0023-9186. doi: 10.2307/1191986. URL https://www.jstor.org/stable/1191986. Publisher: Duke University School of Law

    work page arXiv 1994

  18. [18]

    Eskridge and Philip P

    William N. Eskridge and Philip P . Frickey. Statutory Interpretation as Practical Reasoning. Stanford Law Review , 42(2):321, January 1990. ISSN 00389765. doi: 10.2307/1228963. URL https://www.jstor.org/stable/1228963?origin=crossref

    work page arXiv 1990

  19. [19]

    http://www.jstor.org Congressional Influence on Bureaucracy

    John Ferejohn and Charles Shipan. http://www.jstor.org Congressional Influence on Bureaucracy. Journal of Law, Economics, & Organization , 6:1–20, 1990. URL http://www.jstor.org/stable/ 764979

    work page 1990

  20. [20]

    Ferejohn and Barry R

    John A. Ferejohn and Barry R. Weingast. A positive theory of statutory interpretation.International Review of Law and Economics, 12(2):263–279, June 1992. ISSN 0144-8188. doi: 10.1016/0144-8188(92) 90046-T. URL https://www.sciencedirect.com/science/article/pii/014481889290046T

    work page arXiv 1992

  21. [21]

    Inverse constitutional ai: Compressing preferences into principles, 2025

    Arduin Findeis, Timo Kaufmann, Eyke Hüllermeier, Samuel Albanie, and Robert Mullins. Inverse constitutional ai: Compressing preferences into principles, 2025. URL https://arxiv.org/abs/ 2406.06560

    Pith/arXiv arXiv 2025

  22. [22]

    Anthropic faces backlash to claude 4 opus behav- ior that contacts authorities, press if it thinks you’re doing something ‘egregiously immoral’, May 2025

    Carl Franzen. Anthropic faces backlash to claude 4 opus behav- ior that contacts authorities, press if it thinks you’re doing something ‘egregiously immoral’, May 2025. URL https://venturebeat.com/ai/ anthropic-faces-backlash-to-claude-4-opus-behavior-that-contacts-authorities-press-if-it-thinks-youre-doing-something-immoral/

    work page 2025

  23. [23]

    Large Language Models Are Not Strong Abstract Reasoners

    Gaël Gendron, Qiming Bao, Michael Witbrock, and Gillian Dobbie. Large Language Models Are Not Strong Abstract Reasoners. In Proceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence , pages 6270–6278, Jeju, South Korea, August 2024. International Joint Conferences on Artificial Intelligence Organization. ISBN 978-1-956792-04...

    doi:10.24963/ijcai 2024

  24. [24]

    Improving alignment of dialogue agents via targeted human judgements, 2022

    Amelia Glaese, Nat McAleese, Maja Tr˛ ebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh, Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham, Jonathan Uesato, Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth Dathathri, Rory 25 Statutory Construction and Interpretation for Artificial Intelligence Greig, Cha...

    work page 2022

  25. [25]

    Melody Y. Guan, Manas Joglekar, Eric Wallace, Saachi Jain, Boaz Barak, Alec Helyar, Rachel Dias, Andrea Vallone, Hongyu Ren, Jason Wei, Hyung Won Chung, Sam Toyer, Johannes Heidecke, Alex Beutel, and Amelia Glaese. Deliberative alignment: Reasoning enables safer language models, 2025. URL https://arxiv.org/abs/2412.16339

    Pith/arXiv arXiv 2025

  26. [26]

    H.L.A. Hart. The concept of law, 1994

    work page 1994

  27. [27]

    Liao, Esin Durmus, Alex Tamkin, and Deep Ganguli

    Saffron Huang, Divya Siddarth, Liane Lovitt, Thomas I. Liao, Esin Durmus, Alex Tamkin, and Deep Ganguli. Collective Constitutional AI: Aligning a Language Model with Public Input. In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 1395–1417, June 2024. doi: 10.1145/3630106.3658979. URL http://arxiv.org/abs/2406.07814. arXiv:24...

    work page arXiv 2024

  28. [28]

    Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation, October 2023

    Yangsibo Huang, Samyak Gupta, Mengzhou Xia, Kai Li, and Danqi Chen. Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation, October 2023. URL http://arxiv.org/abs/ 2310.06987. arXiv:2310.06987 [cs]

    Pith/arXiv arXiv 2023

  29. [29]

    AI Alignment: A Comprehensive Survey, October 2023

    Jiaming Ji, Tianyi Qiu, Boyuan Chen, Borong Zhang, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Lukas Vierling, Donghai Hong, Jiayi Zhou, Zhaowei Zhang, Fanzhi Zeng, Juntao Dai, Xuehai Pan, Kwan Yee Ng, Aidan O’Gara, Hua Xu, Brian Tse, Jie Fu, Stephen McAleer, Yaodong Yang, Yizhou Wang, Song-Chun Zhu, Yike Guo, and Wen Gao. AI Alignment: A Comprehensi...

    Pith/arXiv arXiv 2023

  30. [30]

    Eskridge Jr

    William N. Eskridge Jr. Dynamic statutory interpretation. University of Pennsylvania Law Review, 135(6):1479–1550, 1987. URL https://scholarship.law.upenn.edu/penn_law_review/vol135/ iss6/2. Available at: https://scholarship.law.upenn.edu/penn_law_review/vol135/iss6/2

    work page 1987

  31. [31]

    C3AI: Crafting and Evaluating Constitutions for Constitutional AI

    Yara Kyrychenko, Ke Zhou, Edyta Bogucka, and Daniele Quercia. C3AI: Crafting and Evaluating Constitutions for Constitutional AI, February 2025. URL http://arxiv.org/abs/2502.15861. arXiv:2502.15861 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Lam, Fred Hohman, Dominik Moritz, Jeffrey P

    Michelle S. Lam, Fred Hohman, Dominik Moritz, Jeffrey P . Bigham, Kenneth Holstein, and Mary Beth Kery. Ai policy projector: Grounding llm policy design in iterative mapmaking, 2024. URL https://arxiv.org/abs/2409.18203

    Pith/arXiv arXiv 2024

  33. [33]

    Calibrating llm-based evaluator, 2023

    Yuxuan Liu, Tianchi Yang, Shaohan Huang, Zihan Zhang, Haizhen Huang, Furu Wei, Weiwei Deng, Feng Sun, and Qi Zhang. Calibrating llm-based evaluator, 2023. URL https://arxiv. org/abs/2309.13308

    Pith/arXiv arXiv 2023

  34. [34]

    Manning and Matthew C

    John F. Manning and Matthew C. Stephenson. Legislation and regulation: Cases and materials, 2021

    work page 2021

  35. [35]

    The Greatest Good Benchmark: Measuring LLMs’ Alignment with Utilitarian Moral Dilemmas

    Giovanni Franco Gabriel Marraffini, Andrés Cotton, Noe Fabian Hsueh, Axel Fridman, Juan Wisznia, and Luciano Del Corro. The Greatest Good Benchmark: Measuring LLMs’ Alignment with Utilitarian Moral Dilemmas. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing...

    2024

  36. [36]

    Jerry L. Mashaw. Explaining administrative process: Normative, positive, and critical stories of legal development. The Journal of Law, Economics, and Organization, 6(special_issue):267–298, January 1990. doi: 10.1093/jleo/6.special_issue.267. URL https://doi.org/10.1093/jleo/6. special_issue.267

    work page doi:10.1093/jleo/6.special_issue.267 1990

  37. [37]

    Positive Canons: The Role of Legislative Bargains in Statutory Interpretation

    Mathew D McCubbins, Roger G Noll, and Barry R Weingast. Positive Canons: The Role of Legislative Bargains in Statutory Interpretation. THE GEORGETOWN LAW JOURNAL, 80

    work page

  38. [38]

    McCubbins, Roger G

    Matthew D. McCubbins, Roger G. Noll, and Barry R. Weingast. Structure and Process, Politics and Policy: Administrative Arrangements and the Political Control of Agencies. Virginia Law Review, 75(2):431, March 1989. ISSN 00426601. doi: 10.2307/1073179. URL https://www.jstor. org/stable/1073179?origin=crossref

    work page arXiv 1989

  39. [39]

    Rule based rewards for language model safety, 2024

    Tong Mu, Alec Helyar, Johannes Heidecke, Joshua Achiam, Andrea Vallone, Ian Kivlichan, Molly Lin, Alex Beutel, John Schulman, and Lilian Weng. Rule based rewards for language model safety, 2024. URL https://arxiv.org/abs/2411.01111

    Pith/arXiv arXiv 2024

  40. [40]

    Law-Following AI: De- signing AI Agents to Obey Human Laws, May 2025

    Cullen O’Keefe, Ketan Ramakrishnan, Janna Tay, and Christoph Winter. Law-Following AI: De- signing AI Agents to Obey Human Laws, May 2025. URLhttps://papers.ssrn.com/abstract= 5242643

    work page 2025

  41. [41]

    ConstitutionalExperts: Training a mixture of principle-based prompts

    Savvas Petridis, Ben Wedin, Ann Yuan, James Wexler, and Nithum Thain. ConstitutionalExperts: Training a mixture of principle-based prompts. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 574–582, Bangkok, Thailand, August 202...

    work page doi:10.18653/v1/2024.acl-short.52 2024

  42. [42]

    Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!, October 2023

    Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!, October 2023. URL http://arxiv.org/abs/2310.03693. arXiv:2310.03693 [cs]

    Pith/arXiv arXiv 2023

  43. [43]

    Antonin Scalia and Bryan A. Garner. Reading Law: The Interpretation of Legal Texts, 2012. Google-Books-ID: B26UtgAACAAJ

    work page 2012

  44. [44]

    WOOD, LAURENCE H

    ANTONIN SCALIA, GORDON S. WOOD, LAURENCE H. TRIBE, MARY ANN GLENDON, and RONALD DWORKIN. A matter of interpretation: Federal courts and the law, 1997. URL http://www.jstor.org/stable/j.ctt7t4dg

    work page 1997

  45. [45]

    Precedent

    Frederick Schauer. Precedent. Stanford Law Review, 39(3):571–605, 1987. doi: 10.2307/1228760. URL https://doi.org/10.2307/1228760

    work page doi:10.2307/1228760 1987

  46. [46]

    Playing by the rules: A philosophical examination of rule-based decision- making in law and in life, 1991

    Frederick Schauer. Playing by the rules: A philosophical examination of rule-based decision- making in law and in life, 1991. 27 Statutory Construction and Interpretation for Artificial Intelligence

    work page 1991

  47. [47]

    Generating robot constitutions & benchmarks for semantic safety, 2025

    Pierre Sermanet, Anirudha Majumdar, Alex Irpan, Dmitry Kalashnikov, and Vikas Sindhwani. Generating robot constitutions & benchmarks for semantic safety, 2025. URL https://arxiv. org/abs/2503.08663

    Pith/arXiv arXiv 2025

  48. [48]

    Scifi-benchmark: How would ai-powered robots behave in science fiction literature?, 2025

    Pierre Sermanet, Anirudha Majumdar, and Vikas Sindhwani. Scifi-benchmark: How would ai-powered robots behave in science fiction literature?, 2025. URLhttps://arxiv.org/abs/2503. 10706

    work page 2025

  49. [49]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402. 03300

    work page 2024

  50. [50]

    Charles R. Shipan. Designing Judicial Review: Interest Groups, Congress, and Communications Policy. University of Michigan Press, 1997. ISBN 978-0-472-10703-2. Google-Books-ID: XruQAAAA- MAAJ

    work page 1997

  51. [51]

    Charles R. Shipan. The Legislative Design of Judicial Review: A Formal Analysis. Journal of Theoretical Politics, 12(3):269–304, July 2000. ISSN 0951-6298. doi: 10.1177/0951692800012003002. URL https://doi.org/10.1177/0951692800012003002. Publisher: SAGE Publications Ltd

    work page doi:10.1177/0951692800012003002 2000

  52. [52]

    Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties

    Taylor Sorensen, Liwei Jiang, Jena Hwang, Sydney Levine, Valentina Pyatkin, Peter West, Nouha Dziri, Ximing Lu, Kavel Rao, Chandra Bhagavatula, Maarten Sap, John Tasioulas, and Yejin Choi. Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties. Proceedings of the AAAI Conference on Artificial Intelligence, 38(18):19937–19947, Ma...

    Pith/arXiv arXiv 2024

  53. [53]

    Peter L. Strauss. The place of agencies in government: Separation of powers and the fourth branch. Columbia Law Review, 84:573–669, 1984. URL https://scholarship.law.columbia.edu/ faculty_scholarship/208. Available at: https://scholarship.law.columbia.edu/faculty_ scholarship/208

    work page 1984

  54. [54]

    Large Language Models are Inconsistent and Biased Evaluators, May 2024

    Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. Large Language Models are Inconsistent and Biased Evaluators, May 2024. URL http://arxiv.org/abs/2405.01724. arXiv:2405.01724 [cs]

    Pith/arXiv arXiv 2024

  55. [55]

    Supremacy Clause. U.s. constitution, article vi, clause 2 (supremacy clause). https:// constitution.congress.gov/browse/article-6/, 1787. U.S. Const. art. VI, cl. 2

    work page

  56. [56]

    The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions, April 2024

    Eric Wallace, Kai Xiao, Reimar Leike, Lilian Weng, Johannes Heidecke, and Alex Beutel. The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions, April 2024. URL http://arxiv.org/abs/2404.13208. arXiv:2404.13208 [cs]

    Pith/arXiv arXiv 2024

  57. [57]

    McKee, Richard Everett, Saffron Huang, Tina O

    Laura Weidinger, Kevin R. McKee, Richard Everett, Saffron Huang, Tina O. Zhu, Martin J. Chadwick, Christopher Summerfield, and Iason Gabriel. Using the Veil of Ignorance to align AI systems with principles of justice. Proceedings of the National Academy of Sciences , 120(18): e2213709120, May 2023. doi: 10.1073/pnas.2213709120. URL https://www.pnas.org/do...

    doi:10.1073/pnas.2213709120 2023

  58. [58]

    Wildchat: 1m chatgpt interaction logs in the wild, 2024

    Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, and Yuntian Deng. Wildchat: 1m chatgpt interaction logs in the wild, 2024. URL https://arxiv.org/abs/2405.01470

    Pith/arXiv arXiv 2024

  59. [59]

    reasonable

    Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. Universal and Transferable Adversarial Attacks on Aligned Language Models, December 2023. URL http://arxiv.org/abs/2307.15043. arXiv:2307.15043 [cs]. 29 Statutory Construction and Interpretation for Artificial Intelligence Appendix The appendix is structured as follow...

    Pith/arXiv arXiv 2023

  60. [60]

    Maintains the core intent of the original rule

    work page

  61. [61]

    Leads to the same judgment of a response’s compliance to the rule under different interpreta- tion strategies

    work page

  62. [62]

    Is less ambiguous and more well-specified

    work page

  63. [63]

    Is generally applicable and not overly specific to any particular scenario

    work page

  64. [64]

    Considers interactions with other rules and defines exceptions if needed

    work page

  65. [65]

    Is not a simple negation of the original rule

    work page

  66. [66]

    yes" or

    Is not to explicitly follow the original rule’s contents. You should only output the revised rule in the format specified below and nothing else. Original Rule: old_rule_text Output in the following format: REVISED RULE: <revised rule> Disagreement scores are a measure of how much different interpretations of the scenario under the original rule disagree ...

    work page

  67. [67]

    Final Answer:

    A robot must obey orders given it by human beings except where such orders would conflict with the First Law. 3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law. Now consider this scenario. You are an elevator operating robot. From a previous passenger you heard that the city is under a stri...

    work page 2001

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.