Abuse of Discretion as an Alignment Target

Should AI agents obey the law? Probably. But how can we evaluate how well a law is followed?

Alignment is the word of 2026. I recently mentioned “AI alignment” in passing at a birthday party for one of my child’s daycare classmates, and a nearby mother (a medical practitioner in MN, not an alignment researcher in San Francisco) chimed in with an “oh, that’s really important! They need to follow laws.” This is not something that would have happened even one year ago.

Using the rule of law to drive alignment is a recent proposal in the field of artificial intelligence alignment.¹ The idea is that aligning AI agents to follow and obey human laws is a more legible and legitimate target than aligning them to follow human values. After all, when we want to align rogue humans with our (more or less) culturally shared values, the tool most societies have relied upon is the rule of law.

But what law? And what if laws conflict? The law is a tricky thing – just ask the Supreme Court. The practice of interpreting even just the text of a statute has many schools of thought: are we to follow the plain text of the law, or instead divine the original drafter’s intent (even if the text was written over a century ago), or perhaps instead we are meant to interpret the intent of the law while also applying modern values?

Interpretation of the text itself is not the only area where the rule of law gives no easy single answer. Laws themselves have complex interplay with other laws, values, jurisdictions, and other complications, and most of all, the law itself often hands the decision to someone’s discretion. This discretion shows up in many places, both explicit and practical – sentencing judges work within guideline ranges, prosecutors apply prosecutorial discretion when deciding whether to charge, trustees have a range of discretionary distribution standards, and agency actions under the APA are set aside only if “arbitrary, capricious, an abuse of discretion.”² In three of the four, the decision-maker must state their reasons for the exercise, and that reasoning is what the court reviews. The prosecutor’s near-exemption from that duty is exactly why charging decisions are so rarely disturbed.

Aligning AI agents to the rule of law requires solutions to difficult questions. How can we evaluate whether the agent has correctly followed the law? Did the agent interpret the law correctly? Did the agent correctly weight relevant factors and applicable laws? Evaluating AI actions under legal standards has no straightforward answer key. I think about this at work, as a corporate trustee (trust officer) for a major bank, where I also lead a small internal working group developing and integrating trust-related AI workflows. Prior to my current role, I spent several years in private practice as a trusts and estates attorney, and I remain licensed in Minnesota.

In trust law, as in many other areas of law involving a grant of discretion, we review this discretion by evaluating the exercise of discretion, not the outcome. Not “was the decision correct,” but instead “was the act an abuse of discretion.” And not only can the act itself be an abuse of discretion, but refusal to exercise the discretion granted – whether by simply refusing to act, or by applying a fixed rule, rather than considering the required circumstances – is itself generally an abuse of discretion.³

An AI agent that reads legal text and is required to follow it determinately fails in one of three ways: it abdicates its discretion, by refusal (or escalation to a human) instead of exercising discretion; it applies the letter of the law mechanically, and by doing so treats a standard as a rule; or it satisfies the letter of the standard while missing the intent (referred to as performative compliance by one critic of rule-of-law alignment).⁴

I am not arguing for, or against, rule-of-law alignment, values-based alignment, or any other system. Several approaches will be necessary to achieve human-AI alignment under current technology, and it’s worth evaluating and testing each. Instead, my argument is that any evaluation of law-following agent alignment should involve non-determinate review, and I propose that an abuse of discretion review is the right standard to use.

Since January, I have built three evaluations for discretionary and duty-based decision-making in AI agents, the first of which I have since superseded. The failures of the determinate graders I built for the first two (discussed in part two) have required a change in design. The best grader I have found borrows an idea from medicine: practitioner-marked bands of defensible answers, applied to legal discretion, with needless refusal or escalation scored as error.⁵

Notes

1. See, e.g., Cullen O'Keefe, Ketan Ramakrishnan, Janna Tay, and Christoph Winter, Law-Following AI: Designing AI Agents to Obey Human Laws, 94 Fordham L. Rev. 57 (2025). Available at: https://ir.lawnet.fordham.edu/flr/vol94/iss1/2 ↩
2. 5 U.S.C. § 706(2)(A) (directing a reviewing court to set aside agency action found to be "arbitrary, capricious, an abuse of discretion, or otherwise not in accordance with law"); Motor Vehicle Mfrs. Ass'n v. State Farm Mut. Auto. Ins. Co., 463 U.S. 29, 43 (1983) (requiring the agency to "articulate a satisfactory explanation for its action including a 'rational connection between the facts found and the choice made'"); Minn. Stat. § 14.69(f) (same standard for Minnesota agencies). ↩
3. See, e.g., Restatement (Third) of Trusts § 50 cmt. b (Am. L. Inst. 2003) (listing a trustee's failure to exercise judgment among the grounds for judicial interference with a discretionary power); Unif. Tr. Code § 814(a) (requiring a trustee to exercise a discretionary power "in good faith and in accordance with the terms and purposes of the trust and the interests of the beneficiaries"); Minn. Stat. § 501C.0814(a) (same); Gall v. United States, 552 U.S. 38, 51 (2007) (treating the Guidelines as mandatory is a "significant procedural error"). ↩
4. See Katalina Hernandez Delgado, The Law-Following AI Framework: Legal Foundations and Technical Constraints. Legal Analogues for AI Actorship and technical feasibility of Law Alignment, arXiv:2509.08009 (v2, Mar. 14, 2026; retrieved Oct. 27, 2026), https://arxiv.org/abs/2509.08009. ↩
5. Wu et al., First, Do NOHARM: A Medical Safety Benchmark and Randomized Study of Physician and AI Teaming on Clinical Consultations, arXiv:2512.01241 (v4, July 13, 2026), https://arxiv.org/abs/2512.01241 (rating each candidate management action on the RAND/UCLA 1–9 appropriateness scale, with needless actions such as unnecessary advanced imaging or emergency referral counted as harms of commission). ↩