Lawyers, Linguists, Computer Scientists, and the Meaning of “Physically Restrained”
The third in a series of essays on the 2024-25 U.S. Sentencing Commission guideline amendment year
In the first essay in this series about the recently completed 2024-25 U.S. Sentencing Commission guideline amendment year, I found fault with many of the Commission’s decisions and some of its processes too. But there was one small guideline amendment, approved by the Commission unanimously in April and submitted to Congress earlier this month, that I found quite sensible. It addressed conflicting circuit court decisions on whether the “physically restrained” enhancement at §2B3.1(b)(4)(B) in the robbery guideline should be applied to cases in which a victim is restricted from moving at gunpoint but is not otherwise immobilized through physical measures.
Late last year, I was introduced to an extraordinary group of lawyers, linguists, and computer scientists, who happened to be studying this very issue. They were writing an article about the use of artificial intelligence (AI) and large language models (LLMs) to help judges with statutory and other legal interpretation. Judge Kevin Newsom of the United States Court of Appeals for the Eleventh Circuit had issued two opinions on the use of AI and LLMs in interpreting first, terms in a contract, and second, terms in a guideline. In Snell v. United Specialty Insurance Co., 102 F.4th 1208 (11th Cir. 2024), Judge Newsom floated this suggestion: “Those, like me, who believe that ‘ordinary meaning’ is the foundational rule for evaluation of legal text should consider – consider – whether and how AI-powered large language models like OpenAI’s ChapGPT, Goggle’s Gemini, and Anthropic’s Claude might – might – inform the interpretative analysis.” That opinion was later followed by one in United States v. Deleon, which just so happened to address the meaning of “physically restrained” in §2B3.1(b)(4)(B) and in which Judge Newsom again played with the idea of using AI and LLMs in deciding the case. The article written by the group of lawyers, linguists, and computer scientists dove into Judge Newsom’s thinking.
When I first read the article, I was reminded of how interdisciplinary study can lead to new insights and how we lawyers mistakenly tend to think our training is sufficient to answer most questions faced by courts. We want to believe that being a generalist with good analytical skills is sufficient.
I reached out to the authors and encouraged them to join me in sharing their work with the Commission. It wasn’t so much the analysis of how using AI might be used in guideline interpretation that I thought was relevant, but rather the application of linguistic analysis to the term “physically restrained” that seemed to me to resolve definitively what was dividing the circuit courts and what the Commission was taking up.
Last summer, I wrote an essay for this Substack about how diversity of thought can bring new insights to the work of the Sentencing Commission and federal sentencing law and policy. It included this –
When Judge Marvin Frankel first advocated for a “Commission on Sentencing” in his classic work, Criminal Sentences: Law Without Order, he envisioned a commission that would include “lawyers, judges, penologists, and criminologists ... sociologists, psychologists, business people, artists, and, lastly for emphasis, former or present prison inmates.” We need to remind ourselves of this wisdom from Judge Frankel now and then; that people who look at the world differently from us have much to contribute. Lawyers are great. A diverse set of lawyers are even better. But even more diversity at the Commission – by tapping into new people, new skills, new backgrounds, and machine learning too – might be better still and might bring with it some new thinking and insights and perhaps better sentencing policy.
I think we should add linguists and computer scientists to the list of diverse views that might be part of the Sentencing Commission someday.
Our letter to the Commission addressing the circuit conflict is below. I encourage you to read the full paper referenced in it. It’s really interesting. The article explains a lot about how AI is constructed and how it works, some of the traps to avoid in using it, and some of the ways to use it effectively. The letter lays out the position we encouraged the Commission to take on the meaning of “physically restrained” and how linguistics led us to that position. We were gratified when the Commission adopted our position in April.
- - -
February 15, 2025
The Honorable Carlton W. Reeves, Chair
United States Sentencing Commission
One Columbus Circle, NE
Suite 2-500, South Lobby
Washington, DC 20002-8002
Dear Judge Reeves:
This is in response to the Commission’s Proposed Amendment to the Sentencing Guidelines, Circuit Conflicts, Part A, published in the Federal Register on January 2, 2025, and to the comments submitted on this proposed amendment on or before February 3, 2025. The proposed amendment and the comments address a circuit conflict concerning whether the physically restrained enhancement at §2B3.1(b)(4)(B) of the robbery guideline should be applied to cases in which a victim is restricted from moving at gunpoint but is not otherwise immobilized through physical measures. We appreciate the opportunity to share a few thoughts with you on this proposal.
We are lawyers, linguists, and computer scientists, who authored an article, titled Large Language Models and Legal Interpretation? Don’t Take Their Word for It (available at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5123124), and also a former Sentencing Commissioner who read and commented on the article and then suggested to the authors to submit it to the Commission. This article as a whole examines the use of Artificial Intelligence in interpreting legal text. It was inspired in part by two concurring opinions of Judge Kevin Newsom of the United States Court of Appeals for the Eleventh Circuit, including his opinion in United States v. Deleon, which addresses the meaning of physically restrained in §2B3.1(b)(4)(B). We especially want to draw your attention to the article’s pages 60-67, which present an analysis of the phrase physically restrained from a theoretical linguistic perspective. We believe that this linguistic analysis can inform the resolution of the circuit conflict over how to interpret physically restrained as it is currently used in the robbery guideline.
It is not at all surprising to a linguist that the term physically restrained could give rise to an interpretative dispute. This is because as a matter of ordinary meaning, without additional context, physically restrained is inherently ambiguous between two possible interpretations. According to one interpretation, the adverb physically is used to describe the means by which the robbery victim is restrained. The restraint is accomplished through physical means, such as the use of duct tape or rope or other physical means that limit the victim’s movement. We’ll call this the Means Interpretation. In contrast, consider a robber who does not use physical means but instead simply threatens the use of physical force by pointing a gun at the victim and telling the victim to hand over the money. Or consider a robber who says, with a raised fist, “hand me the money or I’ll break your nose.” Under the Means Interpretation, in these latter two scenarios, the victim is not physically restrained because no physical means are used to restrain the victim.
On a second interpretation of physically restrained, though, the adverb physically is used to describe the nature of the resulting restraint. We call this the Domain Interpretation. Under this linguistic interpretation, if the robbery victim exhibited restraint in their physical activity, they would be considered restrained from moving – physically restrained – regardless of the means used by the robber. Under this interpretation, both physical and non-physical actions, such as verbal threats, can result in physical restraint.
We bring all this to your attention, because we think it is clear that physically restrained has just one plausible interpretation – the Means Interpretation – in the context of §2B3.1(b)(4). That section specifies a sentencing enhancement. It is meant to identify an aggravating factor that warrants a sentence higher than the base case of robbery, which is accounted for in the Base Offense Level. That base case involves the use or threatened use of force to take the property of another. This is what robbery is. On a broad understanding of the Domain Interpretation, the victim of a robbery is always physically restrained, in the sense that the victim is physically alienated from their property and restricted in their physical movement (or at least there is an attempt to do so). Therefore, applying a broad Domain Interpretation to physically restrained in §2B3.1(b)(4) would result in an enhancement that applies in every – or nearly every – robbery. That is not the purpose of a specific offense characteristic or any identified aggravating factor in the Guidelines. We think the context in which physically restrained is used in §2B3.1(b)(4) definitively supports the Means Interpretation.
Moreover, the existing Background Commentary to the robbery guideline further supports the Means Interpretation. It states, “‘Physically restrained’ means the forcible restraint of the victim such as by being tied, bound, or locked up.” Crucially, and consistent with the language of §2B3.1(b)(4), this statutory definition contemplates the means by which a victim is forcibly restrained (tying, binding, locking), rather than the domain of the restraint.
As Judge Rosenbaum writes in his concurring opinion in Deleon, “I understand [§2B3.1(b)(4)(B)] to reflect Congress’s view that physical bounding adds another dimension to the intimidation a victim of an armed robbery endures. In other words, it’s bad enough to be robbed at gunpoint, but being tied up or otherwise physically bound is invasive, dangerous, and victimizing in yet another way.” Section 2B3.1(b)(4)(B), using the Means Interpretation of physically restrained, straightforwardly delivers an enhancement in line with this view: it contemplates a particularly “invasive, dangerous, and victimizing” means of restraining a person in the course of a robbery.
We hope you find this helpful. Please let us know if there’s anything more we might do to assist you in your work. We are grateful for this opportunity to weigh in.
Sincerely,
/s/ Brandon Waldon
Postdoctoral Fellow in Linguistics, Computer Science, and Public Policy
Georgetown University
/s/ Jonathan J. Wroblewski, Director, Semester in Washington Program
and Lecturer on Law
Harvard Law School
/s/ Nathan Schneider
Associate Professor of Linguistics and Computer Science
Georgetown University
/s/ Ethan Wilcox
Assistant Professor of Computational Linguistics
Georgetown University
/s/ Amir Zaldes
Associate Professor of Computational Linguistics
Georgetown University
/s/ Kevin Tobia
Professor of Law and Associate Professor of Philosophy
Georgetown University