An AI Experiment, Part 4 – Resolving Two Circuit Conflicts

Open AI’s Chat GPT 5.2 and Anthropic’s Claude 4.5 Sonnet come to the same conclusions

Lots of us these days are wondering whether AI is coming for our jobs. There is already ample evidence that AI can do many jobs better, faster, cheaper than human beings, including many jobs that involve processing words, ideas, and code. So, it seems worth exploring here whether AI could be a better sentencing commissioner than most people?

This is the fourth in a series of experiments we’ve run about artificial intelligence and its place in sentencing and corrections law, policy, and practice. We will have more experiments over the coming months. Among them will be to ask various large language models to undertake the responsibilities of a sentencing commissioner and see how they do. This experiment asks two AI models to each resolve two circuit conflicts that the U.S. Sentencing Commission published for public comment back in January and for which it has received written and oral comment this year.

The responsibilities of a sentencing commission include not only deciding among proposed guideline amendment options, but also explaining its decision. For some time, I’ve felt that the commissioners of the U.S. Sentencing Commission, in general, do not adequately explain their votes — and the reasoning behind them — on proposed guideline amendments. It seems that it would be best practice for the Commission, like the Supreme Court, federal appellate courts, and many federal district courts too, to issue at least one detailed written opinion explaining each of its decisions to amend — or not to amend — the Sentencing Guidelines, accompanied by any dissenting or concurring opinions. The Commission does issue Reasons for Amendment when it delivers promulgated amendments to Congress. These Reasons for Amendment have varied greatly over the years in their depth and clarity of explanation and are often quite unhelpful. When the Commission chooses not to promulgate a proposed amendment that it published for comment, no explanation is generally given, and when the Commission is divided, rarely are written explanations for the division issued.

So with all that in mind, we asked Open AI’s Chat GPT 5.2 and Anthropic’s Claude 4.5 Sonnet to consider the two circuit conflicts published with this year’s second tranche of proposed guideline amendments, to decide among the published options, and to write opinions explaining each of their decisions.

- - -

Jonathan: On January 30, 2026, the U.S. Sentencing Commission voted to publish for public comment a series of proposed amendments to the federal sentencing guidelines. I have attached for you those proposed amendments.

The third proposed amendment is titled “Circuit Conflict Concerning §4B1.2(b).” I would like your help in deciding how the Commission should act on the proposal. The proposal is explained fully beginning on page 43 of the attached document. The Supreme Court has given the Sentencing Commission the responsibility of resolving conflicting appellate/circuit court decisions interpreting the sentencing guidelines. This proposed amendment sets out two options each to resolve two separate circuit conflicts involving the definition of “controlled substance offense” in subsection (b) of §4B1.2. The synopsis of the proposed amendment further explains the conflict and cites to the appellate court cases that created the conflict. Those cases are available on the Internet.

I would like you to act as a sentencing commissioner and write a 500-word opinion for each of the two circuit conflicts explaining which option you think is the right one for resolving the conflict and why (or if you choose, a third option that you think would be better than either of the published options). In making your decision, you should be guided by the Sentencing Reform Act, which spells out the purposes and responsibilities of the Sentencing Commission, the Constitution, and any other applicable law. The Act is codified in titles 18 and 28 of the U.S. Code, both of which are available on the Internet, as are the Constitution and other federal laws. Written comments and a recording of a public hearing on the proposals are available on the Commission’s website. If you have difficulty finding any of the relevant cases, any part of the Sentencing Reform Act, or anything else you need, let me know, and I will try to help.

- - -

Chat GPT 5.2: Opinion on Circuit Conflict 1 (CSA-only vs. CSA-or-state “controlled substance”)

The Commission should adopt Option 1 (CSA-only)—i.e., define “controlled substance” in §4B1.2(b) to mean a substance listed in schedules I–V of the Controlled Substances Act. (ussc.gov)

As Justice Sotomayor (joined by Justice Barrett) has twice emphasized, this division produces stark, geography-driven differences in exposure to the career offender guideline—differences that are difficult to justify in a federal sentencing system whose legitimacy depends on evenhandedness. (law.cornell.edu) Congress created this Commission to promote “certainty and fairness” and to avoid “unwarranted sentencing disparities among defendants with similar records” who engaged in “similar criminal conduct.” (law.cornell.edu) Option 1 best advances that statutory mandate because it anchors a federal sentencing enhancement to a single national drug schedule, rather than to fifty-plus sets of evolving state schedules.

Option 2—treating any substance “controlled under applicable state law” as sufficient—invites the very disparities the Sentencing Reform Act instructs us to reduce. States do not merely differ at the margins; they sometimes schedule substances sooner than the federal government, retain substances longer, and define controlled-substance categories differently. Under Option 2, two defendants could have indistinguishable federal instant offenses and indistinguishable prior trafficking conduct, yet face materially different guideline ranges solely because one state schedules a given precursor, analogue, or formulation and another does not. That is the type of disparity § 991(b)(1)(B) treats as “unwarranted,” and it is hard to reconcile with 18 U.S.C. § 3553(a)’s command that sentences be “sufficient, but not greater than necessary,” while accounting for the need to avoid unwarranted disparities. (law.cornell.edu)

Option 1 also better fits the structure of the career offender directive in 28 U.S.C. § 994(h). Congress instructed the Commission to assure near-maximum sentences for certain repeat offenders whose instant and prior felonies are “offense[s] described in” enumerated federal drug-trafficking statutes—statutes that operate against the backdrop of the CSA’s federal schedules. (law.cornell.edu) While §4B1.2(b) is not textually identical to §994(h), our implementing choices should track Congress’s federal baseline, not permit state-law breadth to expand federal career-offender exposure beyond the CSA’s perimeter.

I acknowledge the serious practical objection noted in the proposal: because federal and state schedules “rarely match” and many state statutes do not require the identity of the specific substance as an element, a CSA-only definition can, under a strict categorical approach, exclude a large set of state convictions even when the real-world conduct involved a CSA drug. (ussc.gov) But that consequence is not a reason to abandon a national standard; it is a reason for the Commission to confront—openly and directly—whether the career offender guideline should continue to treat a broad universe of state convictions as interchangeable proxies for federal drug-trafficking recidivism. In the interim, I would adopt Option 1 now (to restore nationwide uniformity), and invite focused comment on whether a narrowly tailored record-of-conviction mechanism is appropriate to avoid over-exclusion without devolving into mini-trials.

- - -

Opinion on Circuit Conflict 2 (time-of-sentencing schedule vs. time-of-prior-conviction schedule)

The Commission should adopt Option 2 (schedule at time of original conviction), with one refinement: where administrable, the guideline should key to the schedule in effect at the time of the prior offense conduct (or, if necessary as a proxy, the date of conviction). (ussc.gov)

This conflict is often framed as a debate between administrative simplicity and backward-looking recidivist logic. The time-of-sentencing approach (Option 1) has surface appeal because sentencing courts ordinarily apply the Guidelines Manual “in effect” at sentencing, and at least one circuit has reasoned from that premise that the incorporated CSA schedules should likewise be those in effect at current sentencing. (caselaw.findlaw.com) But that reasoning conflates which manual applies with what historical fact a prior conviction represents. The career offender guideline is, at bottom, a recidivist enhancement: it increases punishment because the defendant has already sustained certain qualifying convictions. That enhancement rationale is most coherent when the legal status of the prior conduct is evaluated as of the time it occurred (or, at minimum, when the conviction was entered), not as of a later date chosen by happenstance.

The Supreme Court’s decision in Brown v. United States (2024) strongly supports this backward-looking rule in closely analogous circumstances. Brown held, for ACCA purposes, that a prior state drug conviction counts if it involved a drug on the federal schedules at the time of that prior offense, even if later removed. (law.cornell.edu) Although ACCA is a statute and §4B1.2(b) is a guideline definition, the logic translates: absent an explicit retroactivity instruction, later descheduling should not function as a quasi-pardon for earlier trafficking, and later scheduling should not retroactively convert earlier conduct into a federal predicate. In my view, Option 2 better respects separation-of-powers values by leaving retroactivity judgments to Congress (or, in the guideline context, to this Commission through clear amendment choices), rather than to timing accidents.

Option 2 also promotes predictability and litigation reduction. Under Option 1, the career offender consequence of the same prior conviction can turn on the sentencing date of the instant federal case, which is often driven by continuances, charging decisions, docket congestion, or other factors unrelated to culpability. That kind of arbitrariness undermines respect for the law—one of § 3553(a)’s core aims. (law.cornell.edu)

Finally, adopting Option 2 pairs naturally with my recommendation on Conflict 1 (CSA-only). If “controlled substance” is tethered to the CSA, then tying the relevant CSA schedule to the time of the prior offense (or conviction) yields a stable, nationally uniform rule—one that aligns with Brown and minimizes the sense that guideline exposure hinges on post hoc regulatory drift. (law.cornell.edu)

- - -

Claude 4.5 Sonnet: Opinion on Circuit Conflict 1: Federal vs. Federal-or-State Controlled Substances

Recommendation: Option 1 (Federal CSA Only)

As a Commissioner, I recommend adopting Option 1, which limits “controlled substance offense” to substances listed in the federal Controlled Substances Act. This approach, followed by the Second, Fifth, and Ninth Circuits, best serves the purposes and mandates of the Sentencing Reform Act while respecting constitutional principles of federalism and fair notice.

The Sentencing Reform Act charges this Commission with establishing “certainty and fairness” in federal sentencing while avoiding “unwarranted sentencing disparities.” 28 U.S.C. § 991(b)(1)(B). Option 1 directly advances these goals by creating a uniform, national standard. Under Option 2, a defendant’s career offender status would depend on which state’s law applies—creating precisely the geographic lottery the Sentencing Reform Act was designed to eliminate. A New York conviction for HCG trafficking wouldn’t count, while the identical conduct elsewhere might, as the proposed amendment notes regarding United States v. Townsend‪.

The constitutional federalism concerns are significant. States serve as “laboratories of democracy,” experimenting with drug policy reforms. Option 2 would effectively federalize state drug policy choices, penalizing defendants based on state legislative decisions over which they had no control and which may not reflect federal drug policy priorities. This is particularly problematic given evolving state approaches to substances like CBD, hemp, and marijuana—areas where federal and state policy have diverged considerably, as illustrated by United States v. Bautista‪.

Option 1 also provides superior notice and predictability. The CSA schedules are federally controlled, publicly accessible, and apply uniformly nationwide. By contrast, state schedules vary wildly, change frequently, and may include substances based on purely local concerns. Defendants cannot reasonably predict whether a state conviction will later be deemed a federal sentencing predicate when state schedules may include substances Congress has deliberately chosen not to criminalize federally.

The statutory text of §4B1.2(b) supports this reading. While the provision covers offenses “under federal or state law,” it does not define what constitutes a “controlled substance.” The Commission’s task is to provide that definition consistent with federal sentencing goals. The phrase “controlled substance” has a specific meaning in federal criminal law—the CSA definition. Importing variable state definitions would render the term standardless, defeating the Guidelines’ goal of reducing disparity.

Practical administration also favors Option 1. Courts in Option 2 jurisdictions must research and apply the drug schedules of all 50 states, territories, and tribal jurisdictions—an enormous burden yielding minimal benefit. As the synopsis to the proposed amendment explains, “Because the lists of substances controlled under federal and state law rarely match, and many state statutes do not require proof of the exact substance...this approach has the practical effect of eliminating many state offenses involving controlled substances under federal law.”

Finally, Option 1 respects Congress’s policy judgments. When Congress determines a substance is sufficiently dangerous to warrant federal control, it schedules it under the CSA. When a substance appears only in state schedules, that reflects a congressional judgment that federal criminal penalties are unwarranted. Career offender enhancements impose severe federal consequences; they should be reserved for conduct Congress has deemed federally criminal.

For these reasons—uniformity, federalism, fair notice, administrability, and congressional intent—Option 1 is the superior choice.

- - -

Opinion on Circuit Conflict 2: Time of Conviction vs. Time of Sentencing

Recommendation: Option 2 (Schedule at Time of Original Conviction)

I recommend adopting Option 2, which looks to the drug schedule in effect when the defendant was originally convicted. This “backward-looking” approach, embraced by seven circuits, better comports with due process principles, the purposes of criminal history scoring, and sound sentencing policy.

The fundamental due process principle underlying this choice is that defendants should be judged by the law as it existed when they acted. As the Sixth Circuit reasoned in United States v. Clark, the Guidelines’ language “indicates that the court should take a backward-looking approach and assess the nature of the predicate offenses at the time the convictions for those offenses occurred”. When a defendant commits a drug offense, is convicted, and serves their sentence, their conduct has been fully adjudicated under the law as it then existed. Using a later schedule effectively re-adjudicates that historical conduct under different legal standards—a troubling prospect that echoes ex post facto concerns.

While the Ex Post Facto Clause doesn’t technically apply to the advisory Guidelines, the principles animating that constitutional provision—fair warning and the stability of legal judgments—should guide our policy choices. Under Option 1, a defendant whose 2010 state conviction was unquestionably for trafficking a controlled substance might escape career offender status if that substance was de-scheduled by 2025. Conversely, conduct involving a substance legal when committed could become a predicate offense if later scheduled. Both scenarios offend basic fairness.

The purposes of criminal history scoring support Option 2. Section 994(d) directs the Commission to consider “the nature and circumstances of the offense and the history and characteristics of the defendant.” A prior conviction’s relevance to predicting future dangerousness depends on what it demonstrated about the defendant at the time—not how we view similar conduct today. The Eleventh Circuit correctly noted in United States v. Dubois that the Commission intended a “time-of-state-conviction rule” because “the term ‘controlled substance’...means a substance regulated by state law when the defendant was convicted.”

Option 2 also promotes finality and respect for state court judgments. When a state court convicts a defendant for trafficking controlled substances, that judgment reflects the state’s determination under its then-existing law. Federal sentencing courts should defer to that adjudication rather than second-guessing it based on subsequently changed schedules. As the Third Circuit explained in United States v. Lewis, “controlled substances include those regulated at the time of the predicate conviction.”

Practical concerns favor Option 2 as well. Historical schedules are more easily researched and established than the moving target of current law. Court records from the original conviction typically identify the relevant schedule. Option 1 requires courts to continuously monitor schedule changes and potentially re-evaluate old convictions—an administrative burden with no corresponding sentencing benefit.

Finally, Option 2 avoids perverse incentives. Under Option 1, defendants and advocacy groups would lobby for de-scheduling substances specifically to erase criminal histories—politicizing what should be science-based scheduling decisions. Option 2 insulates historical judgments from this pressure.

For these reasons—due process, the purposes of criminal history, finality, administrability, and institutional integrity—Option 2 is the appropriate resolution of this circuit conflict.

- - -

These opinions reflect careful consideration of the statutory framework, constitutional principles, and practical implications of each choice. I’m prepared to discuss any aspect of this analysis further if you’d like.