To Err Is Human, But Artificial Intelligence Will Make Human Errors Worse

A recent blog post advocating using Artificial Intelligence (“AI”) to “automate” criminal appeals instead raises the concern that common flaws in contemporary judicial decision-making will only get worse if we cede legal writing to AI Large Language Models. The post analyzes an appellate decision interpreting the federal sentencing guidelines in which all three (human) judges reach the startling conclusion that their decision is incorrect under the guidelines themselves. That conclusion is deeply flawed—the court was not bound to reach the wrong result—but at least the judges understood the result was wrong, which leaves an ember of hope that another court will get it right. AI, by contrast, likely would magnify the errors that led a succession of human judges astray.

To illustrate this point, take the same case discussed in the blog, the Eleventh Circuit’s recent panel decision in United States v. Deleon.¹ Deleon robbed a convenience store at gunpoint, holding the weapon on the cashier until he was satisfied that he had everything from the register. Deleon never touched the cashier, and the incident lasted less than a minute. A jury convicted Deleon of armed robbery and of brandishing a firearm during a crime of violence. The district court enhanced Deleon’s robbery sentence under section 2B3.1(b)(4)(B) of the federal sentencing guidelines for “physically restrain[ing]” the cashier. That sentence enhancement raised two questions: (1) whether the conduct was inherent in the base robbery offense and therefore not a proper basis for enhancement; and (2) whether the conduct was covered by the separate firearm brandishing offense and therefore imposing the enhancement would be double punishment for the same thing.  

Section 2B3.1(b)(4)(B) is part of the specific guideline for robbery offenses. Starting with a “base offense level,” the guidelines provide for increases in the offense level for harmful conduct beyond what is inherent in any federally prosecutable robbery. That includes a two-level increase “[i]f any person was physically restrained to facilitate commission of the offense or to facilitate escape.” An application note explains, “[t]he guideline provides an enhancement for robberies where a victim … was physically restrained by being tied, bound, or locked up.” The reason for the enhancement is not stated, but it’s not difficult to imagine why physical restraint adds culpability to an act that inherently involves virtual restraint. (What robber ever said or implied: “you are free to leave at any time, including before I get your money.”) Physical restraint exposes victims to a different type of harm, and it frees the robber to move about the crime scene and create additional harms, including by pointing the weapon at other victims. The natural reading of the guideline is that enhancement is permitted only if the conduct is not part and parcel of an ordinary robbery and is at least similar in effect to tying, binding, or locking the victim up, even if the conduct listed in the note is only illustrative. There really isn’t much ambiguity to the guideline as written, and certainly not much need for the kind of AI disambiguation the Unikowsky blog post suggests.  

So how did the Eleventh Circuit get to the point where it would take the full court to apply the guideline correctly? As is often the case, it happened step by step when a court followed an imprecisely stated standard in a prior decision without appreciating what facts the standard applied to or checking that judicially articulated standard for conformity to what everyone should agree is the governing positive law—the guideline itself. Like the kid’s game of telephone, the message passes from one appellate panel to the next until the meaning of the standard (how it applies to fact) bears little resemblance to what the guideline actually means or even what the original author of the judicial standard meant.  

The Deleon panel thought it was required to uphold a sentence based on the pretty obviously incorrect application of the “restraint” enhancement to conduct that is inherent in robbery and brandishing a weapon. And that can’t be because the enhancement is not strictly limited to the specific examples in the application note. Normally, even if examples are illustrative rather than exhaustive, they illustrate a category of similar things. Holding a gun on someone while they empty the cash drawer is not similar to tying, binding, or locking up. The two-judge concurring opinion convincingly shows that the majority opinion reaches the wrong result. And the other concurrence agrees. Moreover, none of the other circuits that the concurrence says construe the “restraint” enhancement broadly go as far as Deleon; they all include additional conduct restricting movement._² If “physical restraint” means “creat[ing] circumstances allowing the persons no alternative but compliance,” then the enhancement applies to every armed robbery (which uses the threat of force to obtain property against the will of the victim). But that can’t be right.³

So, we are left with the duty of an appellate panel to follow the holdings of prior panels of the court;⁴ in this case, three prior Eleventh Circuit decisions were deemed binding. In the most recent case, United States v. Ware, 69 F.4th 830 (11th Cir. 2023), the court responded to the argument that applying the “restraint” enhancement would make it applicable to every robbery with a firearm by distinguishing the facts of that case from one in which the robber “entered these establishments, pointed a gun at the victim behind the welcome counter, demanded, ‘Your money or your life,’ obtained money from the victim, and left without further incident.”⁵ The additional facts in Ware included instructions not to move, grabbing an employee by the neck, and forcing another to the ground.⁶ But, as the court in Deleon noted, “that hypothetical situation from Ware is nearly identical to the fact pattern here.”⁷  

In United States v. Victor, 719 F.3d 1288 (11th Cir. 2013), the Eleventh Circuit upheld the application of the “restraint” enhancement on these facts: “Victor found a lobby employee and, while holding in his pocket what the employee believed to be a gun pointed at her, directed her to the teller line. Victor yelled that he had a gun and would kill any bank employee who did not comply with his demands. With the lobby employee frozen in fear beside him, Victor demanded money from the tellers, who gathered cash from their drawers.”⁸ While that wanders pretty far from tying, binding, or locking up, it differs from the facts of Deleon because the defendant restrained the employee’s movement by forcing her into the line.

In the first case, United States v. Jones, 32 F.3d 1512 (11th Cir. 1994), “[t]he robbers also ordered employees and customers into the safe room, where they were told to lie face down on the floor. The robbers closed the door to the safe room and left.”⁹ While that sounds like locking up, along the way to upholding the enhancement, the court held that the restraints in the application note are illustrative and that physical restraint can be satisfied by gunpoint compulsion without any touching. The court cited three cases from two other circuits, all of which involved confinement in particular spaces at gunpoint.¹⁰  

None of the Eleventh Circuit’s prior holdings involved the application of the “restraint” enhancement to conduct limited to the facts of Deleon. Yet the Deleon panel holds that the defendant’s arguments—which a majority of the panel explicitly agree with in a concurrence—fail because “our precedent binds us to conclude that the enhancement applies to conduct like Deleon’s,” even though none of the prior decisions actually involves only such conduct. What the court appears to really be saying is that it is bound by a verbal formula from Jones: “we adopted the Seventh and Eighth Circuits’ understanding of the definition of ‘physically restrained’: ‘a defendant physically restrains his victims if he creates circumstances allowing the persons no alternative but compliance.’”¹¹ But as we’ve seen, neither Jones nor the out-of-circuit cases it “adopted” involved conduct like the defendant’s in Deleon. In short, the panel considered itself bound to rule contrary to the actual sentencing guideline based on language in a prior opinion purporting to interpret the guideline on very different facts. And it did so even after recognizing the conflict between the cautionary hypothetical invoked in Ware and what the court thought was the standard applied in that case: “under the test we again approved in Ware—the enhancement applies where a defendant creates circumstances allowing [his victims] no alternative but compliance—we don’t see how the enhancement wouldn’t apply to the then-hypothetical scenario we raised in Ware. In that case, the robber would have created circumstances allowing his victim no realistic alternative but compliance. After all, who would think compliance was optional in that situation?”¹² The “test” takes precedence over cautionary language about how far to take it on different facts.

The decision in Deleon seems like a perfect example of the trend in modern caselaw to follow the implications of language in prior decisions without regard to what the earlier decision actually held—how it applied the law to the facts. And the case illustrates what happens when courts do that. Inch by inch, case by case, a standard extends until its application has passed the breaking point. Instead of distinguishing prior cases and tacking back to port, the law veers further out to sea.  

One worry about AI is that it seems likely to accelerate the trend towards treating disembodied judicial phrases as inexorable commands. AI gobbles up language, but it is unclear where we will find in Large Language Models the mechanism that will tell AI when to stop, that it has gone too far and has too directly contradicted what has to be the governing positive law (here the guideline itself). AI-drafted briefs (worse, AI-drafted opinions) have the potential to just reinforce that approach.

Some evidence of that appears in the AI-drafted petition for rehearing en banc that Unikowsky generated and attached to his blog post. He characterizes the petition as “fine,” if “wooden” and bearing “some mistakes.” The petition’s lead argument is that the panel decision conflicts with the plain text of § 2B3.1(b)(4)(B). But for whatever reason—and perhaps it’s because of limited inputs—the petition shows no more appreciation of how Eleventh Circuit law went wrong than the panel decision itself. Unikowsky speculates that the 1994 Jones decision was badly litigated and “led the court down the wrong path,” and “[a]s a result, for 30 years, federal courts in the Eleventh Circuit have been stuck deciding cases incorrectly.” While the speculation about the advocacy in Jones seems plausible, in my view the post-Jones panels in the Eleventh Circuit were not stuck deciding cases incorrectly. They did that to themselves by applying a verbal “rule” from Jones without considering the operative facts in that case.

The problem of judicial standards becoming obsolete or otherwise discredited is not new. One of the glories of the common law methodology is the ability to self-correct. Time and again courts of the past recognized that they had become trapped in doctrinal blind alleys or overly rigid and obsolete frameworks and doubled back. Law professors earned their keep by tracking these jurisprudential retreats and creating new vocabularies to describe what courts were actually doing when backing away from judicial disaster.  

Perhaps it is a result of the abject pledge of obedience to precedent that every judicial nominee must nowadays offer before confirmation, but today’s federal judiciary seems to have lost the knack for course correction, and automating criminal appeals with AI bots that cannot question the validity of their inputs will only make it worse. Does a bot know when to stop?

¹ ___ F.4th ___, 2024 WL 4048332 (11th Cir. Sep. 5, 2024).
² Of four out-of-circuit cases cited by the concurrence as consistent with the Eleventh Circuit rule, United States v. Miera, 539 F.3d 1232 (10th Cir. 2008), is explicit about requiring “something more,” which it says was met by pointing the gun, telling people not to move, and standing in front of the doors. Two others both involved ordering tellers to the floor. See United States v. Howell, 17 F.4th 673 (6th Cir. 2021); United States v. Dimache, 665 F.3d 603 (4th Cir. 2011). And in the fourth, United States v. Wallace, 461 F.3d 15 (1st Cir. 2006), the district court imposed sentence under post-Booker advisory guidelines, but only after the jury had made a unanimous factual finding of “restraint” under special jury instructions to implement mandatory guidelines. The appellate court upheld the application of the enhancement because “Notably, the defendant’s co-conspirator jumped in front of Gallinelli when she tried to escape, blocking her path and ordering her at gunpoint to stop. At the same time, the defendant kept his gun pointed directly at DiBiasio’s face and chest, at close range, commanding him to look straight ahead into the gun and not to move.” Id. at 34. The court also overturned the sentence because of other errors.
³ The supposed standard seems more useful in deciding when “physical restraint” might be found without directly touching the victim—a partial rather than complete explication of the guideline.  
⁴ For a recent critique of a circuit practice requiring adherence to prior panel dicta, see Stein v. Kaiser Foundation Health Plan, No. 22-15862 (9th Cir. Sept. 24, 2024) (Forrest, J., concurring).    
⁵ 69 F.4th at 855.
⁶ Id. 
⁷ Deleon, 2024 WL 4048332 at *3 (11th Cir. Sep. 5, 2024). 
⁸ 719 F.3d at 1289.  
⁹ 32 F.3d at 1515; accord id. at 1519. 
¹⁰ See Jones, 32 F.3d at 1519.  
¹¹ 2024 WL 4048332 at *2. 
¹² Id. at *3 (citations and internal quotation marks omitted).