Branch, Circuit Judge

Appeal from the United States District Court for the Southern District of Alabama, D.C. Docket No. 1:21-cv-00229-CG-M Thomas M. O'Hara, O'Hara Law Firm, LLC, Daphne, AL, for Plaintiff-Appellant. Cooper Campbell Thurber, Allen Eugene Graham, Phelps Dunbar, LLP, Mobile, AL, Douglas Kleeman, Phelps Dunbar, LLP, New Orleans, LA, for Defendant-Appellee. Before Newsom, Branch, and Luck, Circuit Judges. Branch, Circuit Judge:

United Casualty Insurance Company ("United") refused to defend James Snell, a landscaper, in a civil lawsuit alleging that Snell had negligently installed a ground-level trampoline in a client's backyard. Snell sued, contending that United had breached its insurance contract with him in bad faith and seeking a declaratory judgment that United had a duty to defend and indemnify Snell. The district court granted summary judgment for United, holding that the accident did not "arise from" Snell's "landscaping" work within the meaning of his commercial general liability policy. On appeal, Snell argues that the district court misconstrued the policy and his complaint, prematurely granted summary judgment on United's duty to indemnify him while the underlying suit was still pending, and erred in granting summary judgment on his bad faith claim.

After review, and with the benefit of oral argument, we affirm. In the first place, the district court correctly concluded that the allegations of the complaint did not trigger United's duty to defend and, though the facts United discovered in investigating the claim present a closer question, Snell's insurance application makes clear that the policy does not cover his work here. Next, Alabama law does not, as Snell suggests, preclude a decision on the duty to indemnify before judgment in the underlying case. Finally, Snell's bad faith claim necessarily fails because United had a lawful basis to deny the claim. But even if that were not so, Snell failed to meet the high burden of showing bad faith under Alabama law because he points to no evidence that United wholly failed to investigate any part of his claim. Instead, he simply insists that if United had done a more thorough investigation, it would agree with him, which is insufficient to show bad faith under Alabama law.

I. Background

A. Factual Background

Snell's landscaping company is named "Outdoor Expressions," and it is located in Fairhope, Alabama.

Snell was hired by a family, the Westons, to turn an above ground trampoline into a ground level trampoline. According to Snell, the Westons wanted an in-ground trampoline because it would be more aesthetically pleasing and sturdy than an above-ground trampoline.

Snell's project for the Westons involved various tasks like tree pruning and removal; installation of shrubs, trees, and sod; and setting up a sprinkler irrigation system. The trampoline aspect of the project involved "site work" to make a place for the trampoline and "assembly and installation" of the trampoline. Snell's site work included excavation of a pit, installation of a drain and drainage sand, excavation of a trench to install a drainage pipe, installation of the drainage pipe and of a drain pump, and, most relevant here, construction of concrete block retainer walls and installation of a wood cap on the retainer walls. The retaining walls aimed to prevent erosion and collapse of the structure; the wood cap was for aesthetics. Then, after all that, Snell unboxed the trampoline, assembled it, and lowered it into the pit.

A few years later, Matthew Burton sued the Westons for injuries his daughter suffered on the Westons' trampoline. Burton's original complaint alleged that his daughter was injured when she "fell off of the trampoline and struck her face on the wooden board" surrounding the trampoline. The original complaint further alleged that "[t]he trampoline had no padding and no netting," and that "[t]he conditions around the trampoline were dangerous and inherently dangerous for those who used it."

Burton later amended his complaint to add Snell, Outdoor Expressions, and the trampoline manufacturer as defendants. The amended complaint purported to incorporate all the allegations of the original complaint.

Burton alleged two claims against the trampoline manufacturer: the first under Alabama's Extended Manufacturer's Liability Doctrine, and the second for Failure to Warn. These claims alleged that the trampoline itself was defective and not in a reasonably safe condition, and that the trampoline manufacturer "failed to adequately warn the minor Plaintiff[ ] of the dangers associated with the use of the trampoline in question." "As installed," Burton alleged, "the trampoline had no netting, inadequate padding and was surrounded by wooden decking."

Burton alleged just one claim against Snell and Outdoor Expressions: "Negligence and/or Wantonness." Burton alleged that Snell "wantonly assembled, constructed and installed the trampoline in the backyard of the Weston[s]'s home," creating "an unreasonably dangerous condition and structure on the property." This "negligence and/or wantonness," Burton alleged, "was the proximate cause of the serious injuries suffered by the minor Plaintiff[.]"

Snell notified United (his insurance carrier) of the lawsuit in late December 2020. On January 5, 2021, United e-mailed Snell stating that it was reviewing the policy to determine if there was coverage. The following week, United informed Snell that it would not defend him in the lawsuit. The denial letter noted that the allegations dealt with wanton assembly of the trampoline and explained that United had no duty to defend Snell against such an action.

To explain its denial of coverage, United turned first to the general coverage provisions of the policy, which provided as follows:

[United] will pay those sums that the insured becomes legally obligated to pay as damages because of "bodily injury"[] or "property damage"[] to which this insurance applies. [United] will have the right and duty to defend the insured against any "suit" seeking those damages. However, [United] will have no duty to defend the insured against any "suit" seeking damages for "bodily injury"

or "property damage" to which this insurance does not apply.

Further, and though United did not specifically rely on it in the letter, Snell's application for insurance is relevant. The application, which predated the policy, specifically asked "[d]o you do any recreational or playground equipment construction or erection?" In response, Snell checked the "No" box.

After receiving the denial letter, Snell's attorney charged that the denial of coverage was in "bad faith" and demanded a retraction and reimbursement for all legal expenses. United stood by the decision.

B. Procedural History

After United refused to defend or indemnify him, Snell filed this lawsuit in Alabama state court. He alleged that United was obligated to defend and indemnify him in the underlying lawsuit. His complaint asserted claims for breach of contract and bad faith denial of coverage, and he sought a declaratory judgment that United had a duty to defend and indemnify him.

United removed the case to federal court based on diversity jurisdiction. After discovery, the district court granted summary judgment in favor of United.

As to the breach of contract and declaratory judgment claims, the district court reasoned that "whether the claims are covered depends upon whether the performance of 'landscaping' would include Snell's installation of the trampoline" under the specified operations provision. The court noted that the policy did not define the word "landscaping," and it concluded that the "common, everyday meaning of the word" did not include trampoline installation "even [under] the definitions submitted by Snell" in his briefing. The court further concluded that "the context of the Policy makes it clear that" the trampoline-related work "is not covered." Under Alabama law, the district court explained, an insurance contract must be "construed according to the entirety of its terms and conditions" including "as . . . modified by any . . . application which is a part of the policy." Ala. Code. § 27-14-17(a). And "Snell was asked in the application whether his work included 'any recreational or playground equipment construction or erection' and Snell answered 'No.' " "If Snell had answered 'Yes' to that question," the district court said, "or if he had informed United [ ] at some time later that his operations were going to include structural work for recreational equipment and the installation of recreational equipment, then United [ ] could have added that coverage and made any appropriate adjustments to Snell's rate." Thus, the district court concluded that the policy was not ambiguous and that United had not breached the contract because it had no duty to defend the lawsuit.

As to Snell's bad faith claim, the district court concluded that, because United "had a lawful basis for denying Snell's claim and [ ] Snell's breach of contract claim fail[ed]," his "bad faith denial [claim] also fail[ed]."

Snell appealed.

II. Standard of Review

"We review de novo a district court's grant of summary judgment, applying the same standard as the district court." Bailey v. Allgas, Inc., 284 F.3d 1237, 1243 (11th Cir. 2002). Namely, summary judgment is appropriate "if the movant shows that there is no genuine dispute as to any material fact and the movant is entitled to judgment as a matter of law." Fed. R. Civ. P. 56(a). The court's function is not "to weigh the evidence and determine the truth of the matter but to determine whether there is a genuine issue for trial." Anderson v. Liberty Lobby, Inc., 477 U.S. 242, 249, 106 S.Ct. 2505, 91 L.Ed.2d 202 (1986). "The mere existence of some evidence to support the nonmoving party is not sufficient for denial of summary judgment; there must be 'sufficient evidence favoring the nonmoving party for a jury to return a verdict for that party.' " Bailey, 284 F.3d at 1243 (quoting Anderson, 477 U.S. at 249, 106 S.Ct. 2505). In applying this standard, the court must view all evidence in the light most favorable to the non-moving party and draw all reasonable inferences in the non-movant's favor. Burton v. City of Belle Glade, 178 F.3d 1175, 1187 (11th Cir. 1999).

III. Discussion

Snell makes three arguments on appeal. First, he contends that the district court erred in granting summary judgment on the duty to defend claim. Second, he argues that even if we affirm the duty to defend claim, the district court's ruling on the duty to indemnify was premature. Third, he submits that his bad faith claim should have survived along with his duty to defend claim. We reject all three arguments.

A. Duty to Defend

Whether United had a duty to defend Snell depends on the scope of coverage in Snell's commercial general liability policy. As we explain below, the district court correctly held that Snell failed to meet his burden to show coverage applies.

i. Snell has the burden to show coverage.

The parties first debate who has the burden of proof in this dispute. Snell contends that United had the burden to show there was no duty to defend the underlying lawsuit because the Specified Operations provision was "essentially an exclusion." It appears that the district court believed Snell had the burden to show coverage and, for the following reasons, we agree.

Under Alabama law, "the party seeking coverage under the Policy[ ] bears the burden of proving that coverage exists," Pa. Nat'l Mut. Cas. Ins. Co. v. St. Catherine of Siena Parish, 790 F.3d 1173, 1180 (11th Cir. 2015) (citing Ala. Hosp. Ass'n Trust v. Mut. Assurance Soc'y of Ala., 538 So. 2d 1209, 1216 (Ala. 1989)), but the insurer bears the burden of proving an exclusion applies, Acceptance Ins. Co. v. Brown, 832 So. 2d 1, 12 (Ala. 2001). See also Belt Auto. Indem. Ass'n v. Ensley Transfer & Supply Co., 211 Ala. 84, 99 So. 787, 790 (Ala. 1924) (explaining that the burden is on the insurer to show an exclusion applies because, "[w]hen an insurance policy contains a general liability clause, followed by a clause or clauses which except specified cases from the operation of the policy," the exceptions are in the nature "of [an] affirmative defense," as to which "the onus of allegation and proof is . . . upon the defendant" (emphasis added)).

While the distinction between limits to coverage and exclusions from coverage may be murky in some cases, the policy here makes clear that the Specified Operations provision is a limit—not an exclusion.

To begin with, the Policy's "Schedule of Forms and Endorsements" describes 27 different "exclusions"—and "Specified Operations" is not one of them. "Specified Operations" is instead described as a "Limitation of Coverage."

But we need not rest on the policy's description of the Specified Operations provision, because the operation of the policy confirms its status. Commercial general liability policies generally "give[ ] coverage through the general coverage provision, and 'take[ ] away' coverage through the various exclusions." Douglas L. Patin, 4 Law and Prac. of Ins. Coverage Litig. § 45:11 (July 2023 Update) (emphasis added) (quotation omitted). Recall, the general coverage provision provides that

[United] will pay those sums that the insured becomes legally obligated to pay as damages because of "bodily injury" or "property damage" to which this insurance applies. [United] will have the right and duty to defend the insured against any "suit" seeking those damages. However, [United] will have no duty to defend the insured against any "suit" seeking damages for "bodily injury" or "property damage" to which this insurance does not apply.

b. This insurance applies to "bodily injury" and "property damage" only if:

(4) The "bodily injury" or "property damage" arises from one or more of the operations shown above; and [i]f also scheduled above[.]"

Thus, the Specified Operations provision is a limitation of coverage—not an exclusion—and Snell, "as the party seeking coverage under the Policy, bears the burden [under Alabama law] of proving that coverage exists." St. Catherine of Siena Parish, 790 F.3d at 1180.

ii. Snell has not shown that United had a duty to defend him.

Snell argues United had a duty to defend him. He asserts that the district court arrived at the opposite conclusion because it "ignored" or mistook "the plain meaning of undefined terms" like "landscaping," failed to give proper breadth to the phrase "arises from," and "applied an unreasonably narrow scope to the claims and allegations in the Underlying Complaint[.]" Because Snell's insurance application—which Alabama law requires us to consider part of the policy—expressly disclaims the work he did here, we disagree.

Under Alabama law, insurance contracts are subject to the same rules of interpretation as any other contract. Safeway Ins. Co. of Ala. v. Herrera, 912 So. 2d 1140, 1143 (Ala. 2005); Twin City Fire Ins. Co. v. Alfa Mut. Ins. Co., 817 So. 2d 687, 691 (Ala. 2001). If the terms of the insurance contract are unambiguous, the court must enforce them as written. Safeway, 912 So. 2d at 1143; Liggans R.V. Ctr. v. John Deere Ins. Co., 575 So. 2d 567, 569 (Ala. 1991).

"Whether an insurance company owes its insured a duty to provide a defense in proceedings instituted against the insured is determined" under Alabama law "primarily by the allegations contained in the complaint." Tanner v. State Farm Fire & Cas. Co., 874 So. 2d 1058, 1063 (Ala. 2003). And an "insurer's duty to defend is more extensive than its duty to [indemnify]," Porterfield v. Audubon Indem. Co., 856 So. 2d 789, 791 (Ala. 2002) (alteration in original) (quoting United States Fid. & Guar. Co. v. Armstrong, 479 So. 2d 1164, 1168 (Ala. 1985)), meaning that, "[i]f the allegations of the injured party's complaint show an accident or occurrence which comes within the coverage of the policy, the insurer is obligated to defend regardless of the ultimate liability of the insured," Chandler v. Ala. Mun. Ins. Co., 585 So. 2d 1365, 1367 (Ala. 1991) (citation omitted). But the duty to defend does not exclusively depend on the complaint: "[i]f the complaint against the insured does not, on its face, allege a covered accident or occurrence, but the evidence proves one, then the insurer likewise owes the duty to defend." Tanner, 874 So. 2d at 1065. Thus, "[t]he insurer owes no duty to defend only if neither . . . the complaint against the insured allege[s] . . . nor . . . the evidence in the litigation . . . prove[s] a covered accident or occurrence." Tanner, 874 So. 2d at 1065.

The parties expend significant energy parsing the words of the policy, including whether the site work necessary to install the trampoline was "landscaping" and whether the trampoline injury "arises from" that work. We conclude we need not resolve those issues here. Even taking the term "landscaping" as ambiguous, construing it in Snell's favor, and applying Alabama law's broad understanding of the causal term "arises out of," Snell's claim still fails. As the district court explained, under Alabama law, "[e]very insurance contract shall be construed . . . as . . . modified by any . . . application which is a part of the policy." Ala. Code. § 27-14-17(a). And the district court's analysis of Snell's application under that statute was correct:

Snell was asked in the application whether his work included "any recreational or playground equipment construction or erection" and Snell answered "No." It is undisputed that the trampoline is "recreational equipment." If Snell had answered "Yes" to that question or if he had informed United Specialty at some time later that his operations were going to include structural work for recreational equipment and the installation of recreational equipment, then United Specialty could have added that coverage and made any appropriate adjustments to Snell's rate.

Snell, for his part, contends that the district court overread § 27-14-17(a) to mean that an application is necessarily part of the insurance contract, rather than part of the contract if it is expressly incorporated. But, however we might read that statute in the first instance, the Alabama Supreme Court disagrees with Snell. In Atlanta Casualty Company v. Russell, the Alabama Supreme Court held that that § 27-14-17(a) means that "the insurance application is to be construed as a part of the policy itself." 798 So. 2d 664, 667 (Ala. 2001) (emphasis added). And we are bound by Alabama's interpretation of its own laws. See Turner v. Wells, 879 F.3d 1254, 1262 (11th Cir. 2018) ("Where the highest court" of a state "has spoken on the topic, we follow its rule.").

Taking the application as part of the policy itself, we agree with the district court that Snell expressly disclaimed doing any of the sort of work he did here—including the site work necessary to install the trampoline that he now claims is "landscaping" out of which the underlying injury "arises."

In sum, the district court correctly held that Snell's insurance application—which Alabama law requires us to consider part of the policy—expressly disclaims the work he did here. Accordingly, we affirm the grant of summary judgment on Snell's duty-to-defend claim against United.

B. Duty to Indemnify

Snell next argues that, even if he loses on the duty to defend claim, "[g]ranting summary judgment on the duty to indemnify was premature." But Snell's reading of Alabama law—i.e., as prohibiting a decision on the duty to indemnify until the underlying case goes to judgment—is incorrect.

True, under Alabama law, it is "often" the case that "a court can only determine whether there is . . . a duty to indemnify—based on facts adduced at the trial of the underlying action." James River Ins. Co. v. Ultratec Special Effects Inc., 22 F.4th 1246, 1252 n.5 (11th Cir. 2022); Hartford Cas. Ins. Co. v. Merchants & Farmers Bank, 928 So. 2d 1006, 1013 (Ala. 2005) (similar). And of course, "there will be nothing to indemnify unless and until damages are assessed against the insured in [the underlying] action." James River Ins. Co., 22 F.4th at 1252 n.5. But that does not mean, as Snell suggests, that "the duty to indemnify is not ripe" for decision under Alabama law unless and until "the insured is held liable in the underlying litigation."

The principal case on which Snell relies, Guaranty National Insurance Company v. Beeline Stores, Inc., held that, under Alabama law, "a determination of the duty to indemnify cannot be made at a preliminary stage in the proceedings, when it is still possible for the plaintiff in the underlying lawsuit to change the theory of liability and assert a claim that is covered by the policy at issue." 945 F. Supp. 1510, 1514 (M.D. Ala. 1996). We are not persuaded. As an initial matter, Beeline is a district court case and therefore does not bind us. Further, the three Alabama cases Beeline cited for that proposition of Alabama law do not support it. See id. at 1514-15 (citing Ladner & Co. v. S. Guar. Ins. Co., 347 So. 2d 100, 104 (Ala. 1977), Home Ins. Co. v. Rice, 585 So. 2d 859, 861 (Ala. 1991), and Tapscott v. Allstate Ins. Co., 526 So. 2d 570, 573-75 (Ala. 1988)).

To start, Ladner did not hold that a court can never grant summary judgment on the duty to indemnify before the underlying case goes to judgment. The Alabama Supreme Court certainly declined to decide the duty to indemnify despite finding no duty to defend, but that was because, in that case, "[i]f the plaintiffs change[d] their theory of liability and assert[ed] a claim against Ladner which [was] covered by the policies," the insurance companies "may well be obligated to defend Ladner and pay any judgment . . . rendered against him[.]" Ladner & Co., 347 So. 2d at 104. We do not take the refusal to decide the question in that case to represent a categorical rule.

The same goes for Rice. As with Ladner, Rice did not definitively hold that the duty to indemnify can never be summarily adjudicated along with the duty to defend—it merely held that a ruling on the duty to defend did not necessarily foreclose any question of the duty to indemnify. See Rice, 585 So. 2d at 861 (explaining that, as a result, the insurer's contention that the court had effectively created coverage by estoppel was misplaced).

Finally, Tapscott affirmatively cuts against Snell's argument because it shows that the duty to indemnify can, under some circumstances, be decided before the underlying case reaches judgment. Tapscott was an appeal from an order granting judgment to the insurer on both the duty to defend and the duty to indemnify—which the Alabama Supreme Court affirmed. 526 So. 2d at 571. The Alabama Supreme Court "conclude[d] that Allstate [was] not required to defend or indemnify [ ] Tapscott" because the complaint alleged only intentional torts, which were not covered by the policy. Id. at 572, 575. However, the court said, "if the complaint [was] later amended to include claims for unintentional torts, then Allstate may be required to defend and indemnify for those claims." Id. at 575. Thus, far from holding that a duty to indemnify claim is not ripe until there is an underlying merits judgment, Tapscott shows that a court can decide at summary judgment that an insurer "is not required to defend or indemnify" the insured—although that determination is subject to later developments. Id. (emphasis added).

Simply put: Snell points us to no authority, and we are aware of none, holding that the duty to indemnify under Alabama law may not be decided before judgment in the underlying litigation.

Turning back to Snell's claim—we conclude that the duty to indemnify issue is ripe. Unlike in Ladner and Rice, nothing about this case suggests that the indemnity question is premature. To the contrary, Snell's insurance application forecloses any duty to indemnify for the same reason it forecloses any duty to defend. Put differently, we see no need to defer ruling on the duty to indemnify because our conclusions about United's obligations depend—not on the facts and legal theories alleged in the underlying complaint—but on Snell's insurance application expressly disclaiming the work he performed here.

In sum, we hold that the duty to indemnify under Alabama insurance law can, at least under some circumstances, be decided before the underlying action reaches judgment. So where, as here, it can be shown that there is no dispute of material fact on the duty to indemnify, summary judgment is appropriate. We therefore affirm the judgment of the district court on the breach of contract claim as to United's duty to indemnify.

C. Bad Faith

Finally, we reject Snell's argument that the district court erred in granting summary judgment on his bad faith denial claim.

Generally, to prove a claim for bad faith refusal to pay an insurance claim, the plaintiff has the burden to prove "(1) the existence of an insurance contract; (2) an intentional refusal to pay the claim; and (3) the absence of any lawful basis for the refusal and the insurer's knowledge of that fact or the insurer's intentional failure to determine whether there is any lawful basis for its refusal." Brown, 832 So. 2d at 16; see also Nat'l Sav. Life Ins. Co. v. Dutton, 419 So. 2d 1357, 1361 (Ala. 1982) (similar). The insured "has a heavy burden of proof," and "the remedy [of bad faith] will be afforded in only extreme cases." Blue Cross & Blue Shield of Ala. v. Granger, 461 So. 2d 1320, 1325, 1329 (Ala. 1984) (alteration in original) (quotation omitted). "Ordinarily, if the evidence produced by either side creates a fact issue with regard to the validity of the [insurance] claim and, thus, the legitimacy of the denial thereof, the [bad faith] tort claim must fail and should not be submitted to the jury." Dutton, 419 So. 2d at 1362.

Snell advances what the Alabama Supreme Court calls an "abnormal" bad faith claim—which is focused on bad faith failure to investigate. See State Farm Fire & Cas. Co. v. Slade, 747 So. 2d 293 (Ala. 1999). A plaintiff advancing an "abnormal" bad faith claim must prove that the insurer (1) intentionally or recklessly failed to investigate a claim; (2) intentionally or recklessly failed to properly subject a claim to a cognitive evaluation or review; (3) manufactured a debatable reason to deny a claim; or (4) relied on an ambiguous portion of a policy as a lawful basis for denying a claim. See id. at 306-07. The plaintiff must also show "that the insurer breached the contract for insurance coverage . . . when it refused to pay the insured's claim." Id. at 318; Ex parte Simmons, 791 So. 2d 371, 379 (Ala. 2000) (same).

The district court correctly explained, therefore, that having "found [ ] that United [ ] had a lawful basis for denying Snell's claim and that Snell's breach of contract claim fails," his "claim for bad faith denial also fails." See Slade, 747 So. 2d at 318.

Regardless, even if Snell's failure to show a breach of contract did not necessarily defeat his bad-faith claim, he has still failed to meet the high burden of showing abnormal bad faith under Alabama law. In Slade, for example, the insureds demonstrated "abnormal" bad faith by the insurance company's total failure to send a qualified lightning investigator to their home before denying their claim, or to interview any witnesses. Id. at 315. Here, by contrast, Snell complains mainly that the coverage denial letter does not (1) explain its definition of landscaping operations, (2) mention the allegation that the minor plaintiff was injured on the wood cap, or (3) make any showing that the claims attorney or anyone else was qualified to investigate "the sequence of events regarding the preparatory site work versus the assembly and placement of the trampoline into the pit." He also protests that he sent a letter to United pointing out these alleged deficiencies, and that even when United finally responded, it "failed to discuss the extent of its investigation, its cognitive evaluation of the facts, and other items raised by the undersigned's letter."

Unlike the situation in Slade, Snell does not point to any aspect of the case that United simply and intentionally ignored. See 747 So. 2d at 315 (holding that the insureds had demonstrated "abnormal" bad faith because the insurance company had totally failed to send a qualified lightning investigator to their home before denying their claim or even interviewed any witnesses). He simply insists that, if United had worked a little harder or thought a little more, it would have agreed with him. That insistence does not show an "intentional failure to determine whether there is a legitimate or arguable reason to refuse to pay the claim." State Farm Fire & Cas. Co. v. Brechbill, 144 So. 3d 248, 258-59 (Ala. 2013) ("State Farm may or may not have perfectly investigated . . . Brechbill's claim to his satisfaction, but perfection is not the standard here. . . . Bad faith . . . is not simply bad judgment or negligence. It imports a dishonest purpose and means a breach of known duty[.]" (quotations omitted)).

Thus, one way or the other, Snell has not shown it was error to grant summary judgment on his bad faith claim.

IV. Conclusion

In sum, we reject Snell's arguments that the district court erred in granting summary judgment for United. We therefore affirm the judgment of the district court.

AFFIRMED.

Newsom, Circuit Judge, concurring:

I concur in the Court's judgment and join its opinion in full. I write separately (and I'll confess this is a little unusual) simply to pull back the curtain on the process by which I thought through one of the issues in this case—and using my own experience here as backdrop, to make a modest proposal regarding courts' interpretations of the words and phrases used in legal instruments.

Here's the proposal, which I suspect many will reflexively condemn as heresy, but which I promise to unpack if given the chance: Those, like me, who believe that "ordinary meaning" is the foundational rule for the evaluation of legal texts should consider—consider—whether and how AI-powered large language models like OpenAI's ChatGPT, Google's Gemini, and Anthropic's Claude might—might—inform the interpretive analysis. There, having thought the unthinkable, I've said the unsayable.

Now let me explain myself.

I

First, a bit of background. As today's majority opinion recounts, both in the district court and before us, the parties litigated this as an "ordinary meaning" case. In particular, they waged war over whether James Snell's installation of an in-ground trampoline, an accompanying retaining wall, and a decorative wooden "cap" fit within the common understanding of the term "landscaping" as used in the insurance policy that Snell had purchased from United Specialty Insurance Company.

So, for instance, the district court observed that "whether the claims are covered depends upon whether the performance of 'landscaping' would include Snell's installation of the trampoline." Doc. 23 at 10. Because the policy didn't define the term "landscaping," the court said, the coverage determination turned on whether Snell's trampoline-related work fit the "common, everyday meaning of the word." Id. at 10-11. Having reviewed multiple dictionary definitions provided by the parties, the court concluded that Snell's work didn't constitute "landscaping." Id. at 13. As the majority opinion explains, the plain-meaning battle continued on appeal, with the parties "expend[ing] significant energy parsing the words of the policy, including whether the site work necessary to install the trampoline was 'landscaping.' " Maj. Op. at 1216. Snell insisted, for example, that the district court had erred by "ignor[ing] the plain meaning of undefined terms" in the policy—most notably, "landscaping." Br. of Appellant at 20, 21.

Now, as it turned out, we managed to resolve this case without having to delve too deeply into the definitional issue that the parties featured—due in large part to (1) a quirk of Alabama law that, according to the state supreme court, makes every insurance application ipso facto part of the policy that it precedes and (2) the fact that in his application Snell had expressly denied that his work included "any recreational or playground equipment construction or erection." Maj. Op. at 1217 (quotation marks omitted). Combined, those two premises yield the majority opinion's controlling conclusion: "Snell's insurance application—which Alabama law requires us to consider part of the policy—expressly disclaims the work he did here" and thus defeats his claim. Id. at 1217.

Importantly, though, that off-ramp wasn't always obviously available to us—or at least as I saw things, to me. Accordingly, I spent hours and hours (and hours) laboring over the question whether Snell's trampoline-installation project qualified as "landscaping" as that term is ordinarily understood. And it was midway along that journey that I had the disconcerting thought that underlies this separate writing: Is it absurd to think that ChatGPT might be able to shed some light on what the term "landscaping" means? Initially, I answered my own question in the affirmative: Yes, Kevin, that is positively absurd. But the longer and more deeply I considered it, the less absurd it seemed.

But I'm getting ahead of myself. I should tell the full story, from beginning to end. In what follows, I'll first explain how my initial efforts to pinpoint the ordinary meaning of the term "landscaping" left me feeling frustrated and stuck, and ultimately led me—initially half-jokingly, later more seriously—to wonder whether ChatGPT and other AI-powered large language models ("LLMs") might provide a helping hand. Next, I'll explore what I take to be some of the strengths and weaknesses of using LLMs to aid in ordinary-meaning interpretation. Finally, given the pros and cons as I see them, I'll offer a few ideas about how we—judges, lawyers, academics, and the broader AI community—might make LLMs more valuable to the interpretive enterprise.

II

First things first. I'm unabashedly a plain-language guy—firmly of the view that "[t]he ordinary meaning rule is the most fundamental semantic rule of interpretation" and that it should govern our reading not only of "constitutions, statutes, [and] rules," but also, as relevant here, of "private instruments." Antonin Scalia & Bryan A. Garner, Reading Law: The Interpretation of Legal Texts 69 (2012). Accordingly, I take it as gospel truth that absent a clear indication that some idiosyncratic, specialized meaning was intended, "[w]ords are to be understood in their ordinary, everyday meanings." Id.; accord, e.g., Shiloh Christian Ctr. v. Aspen Specialty Ins. Co., 65 F.4th 623, 629-30 (11th Cir. 2023) (Newsom, J.) (evaluating an insurance policy's "plain language"); Heyman v. Cooper, 31 F.4th 1315, 1319-20 (11th Cir. 2022) (Newsom, J.) (evaluating a municipal ordinance's "ordinary meaning"); Barton v. U.S. Att'y Gen., 904 F.3d 1294, 1298-99 (11th Cir. 2018) (Newsom, J.) (evaluating a federal statute's "ordinary meaning").

So, following the district court's lead, I did here what any self-respecting textualist would do when trying to assess the ordinary meaning of a particular word, here "landscaping": I went to the dictionaries. In his brief, Snell had served up a buffet of definitions, ranging from Dictionary.com's—"to improve the appearance of (an area of land, a highway, etc.) as by planting trees, shrubs, or grass, or altering the contours of the ground"—to Wikipedia's—"any activity that modifies the visible features of an area of land." See Br. of Appellant at 22-23. My own research revealed, in addition, that Webster's defined "landscaping" as "to modify or ornament (a natural landscape) by altering the plant cover," Merriam-Webster's Collegiate Dictionary 699 (11th ed. 2014), and that Oxford defined it to mean "improv[ing] the aesthetic appearance of (an area) by changing its contours, adding ornamental features, or by planting trees and shrubs," Oxford Dictionary of English 991 (3d ed. 2010).

As occasionally happens, the dictionaries left a little something to be desired. From their definitions alone, it was tough to discern a single controlling criterion. Must an improvement be natural to count as "landscaping"? Maybe, but that would presumably exclude walkways and accent lights, both of which intuitively seemed (to me, anyway) to qualify. Perhaps "landscaping" work has to be done for aesthetic reasons? That, though, would rule out, for instance, a project to regrade a yard, say away from a house's foundation to prevent basement flooding. I once regraded my own yard, and while my efforts did nothing to improve my house's curb appeal, they served an important functional purpose—and for what it's worth, I definitely thought I was engaged in "landscaping."

After languishing in definitional purgatory for a while, I decided to look at the case from a different perspective—and I do mean look. The record contains a series of photographs of Snell's trampoline-related project. Here's one, which shows his prep work—in particular, the empty sand pit and the below-ground retaining wall that reinforced its borders:

Image materials not available for display. And another, which depicts the finished product, including both the polypropylene mat (the fun part) and the decorative wooden "cap":

Image materials not available for display.

I'm not particularly proud of it, but I'll confess that the photos affected the way I thought about the case. Nothing in them really struck me as particularly "landscaping"-y. The problem, of course, was that I couldn't articulate why. And visceral, gut-instinct decisionmaking has always given me the willies—I definitely didn't want to be that guy. So in a way, I felt like I was back to square one.

And that's when things got weird. Perhaps in a fit of frustration, and most definitely on what can only be described as a lark, I said to one of my clerks, "I wonder what ChatGPT thinks about all this." So he ran a query: "What is the ordinary meaning of 'landscaping'?" Here's what ChatGPT said in response:

"Landscaping" refers to the process of altering the visible features of an area of land, typically a yard, garden or outdoor space, for aesthetic or practical purposes. This can include activities such as planting trees, shrubs, flowers, or grass, as well as installing paths, fences, water features, and other elements to enhance the appearance and functionality of the outdoor space.

Suffice it to say, my interest was piqued. But I definitely didn't want to fall into the trap of embracing ChatGPT's definition just because it aligned with my priors. (Bad.) So, in what might have been a mistake—more on that later—we went ahead and asked it the ultimate question: "Is installing an in-ground trampoline 'landscaping' "? ChatGPT responded as follows:

Yes, installing an in-ground trampoline can be considered a part of landscaping. Landscaping involves altering the visible features of an outdoor area for aesthetic or practical purposes, and adding an in-ground trampoline would modify the appearance and function of the space. It's a deliberate change to the outdoor environment, often aimed at enhancing the overall landscape and usability of the area.

* * *

As I've said, intervening developments—in particular, the focus on a statement in Snell's application, which Alabama law makes part of his policy and in which Snell disclaimed doing any recreation-related work—alleviated the need to settle on the ordinary meaning of the term "landscaping." But my own process of working through the plain-language issue was a valuable one, if only because it got me thinking about what was to me a previously unimaginable possibility: Might LLMs be useful in the interpretation of legal texts? Having initially thought the idea positively ludicrous, I think I'm now a pretty firm "maybe." At the very least, it seems to me, it's an issue worth exploring.

So let's explore.

III

In what follows, I'll try to sketch out what I take to be some of the primary benefits and risks of using LLMs—to be clear, as one implement among several in the textualist toolkit—to inform ordinary-meaning analyses of legal instruments.

A

I'll start with the pros as I see them, and then turn to the cons.

1. LLMs train on ordinary-language inputs. Let me begin with what I take to be the best reason to think that LLMs might provide useful information to those engaged in the interpretive enterprise. Recall what is (for many of us, anyway) the "most fundamental semantic rule of interpretation": Absent a clear indication that they bear some technical or specialized sense, the words and phrases used in written legal instruments "are to be understood in the ordinary, everyday meanings." Scalia & Garner, Reading Law, supra, at 69. The premise underlying the ordinary-meaning rule is that "[i]n everyday life, the people to whom rules are addressed continually understand and apply them." Id. at 71. Accordingly, the ordinary-meaning rule, as its name suggests, has always emphasized "common language," Nix v. Hedden, 149 U.S. 304, 307, 13 S.Ct. 881, 37 L.Ed. 745 (1893), "common speech," Sonn v. Magone, 159 U.S. 417, 421, 16 S.Ct. 67, 40 L.Ed. 203 (1895), and "common parlance," Helix Energy Sols. Grp. v. Hewitt, 598 U.S. 39, 52, 143 S.Ct. 677, 214 L.Ed.2d 409 (2023)—in short, as I've explained it elsewhere, "how people talk," United States v. Caniff, 916 F.3d 929, 941 (11th Cir. 2019) (Newsom, J., concurring in part and dissenting in part), vacated and superseded, 955 F.3d 1183 (11th Cir. 2020).

The ordinary-meaning rule's foundation in the common speech of common people matters here because LLMs are quite literally "taught" using data that aim to reflect and capture how individuals use language in their everyday lives. Specifically, the models train on a mind-bogglingly enormous amount of raw data taken from the internet—GPT-3.5 Turbo, for example, trained on between 400 and 500 billion words—and at least as I understand LLM design, those data run the gamut from the highest-minded to the lowest, from Hemmingway novels and Ph.D. dissertations to gossip rags and comment threads. Because they cast their nets so widely, LLMs can provide useful statistical predictions about how, in the main, ordinary people ordinarily use words and phrases in ordinary life. So, for instance, and as relevant here, LLMs can be expected to offer meaningful insight into the ordinary meaning of the term "landscaping" because the internet data on which they train contain so many uses of that term, from so many different sources—e.g., professional webpages, DIY sites, news stories, advertisements, government records, blog posts, and general online chatter about the topic.

To be sure, LLMs' training data aren't a perfect universe from which to draw hard-and-fast conclusions about ordinary meaning, principally because they don't capture what I'll call "pure offline" usages—i.e., those that neither (1) occur online in the first instance nor (2) originate offline, in hard copy, but are eventually digitized and uploaded to some online site. And indeed, the absence of offline usages from the training pool—and in particular, the implications for underrepresented populations—strikes me as a sufficiently serious concern that I've broken it out for separate discussion below. See infra at 1230-32. Even so, those omissions aside, it seems to me scarcely debatable that the LLMs' training data are at the very least relevant to the ordinary-meaning analysis. In fact, an LLMs' dataset may well be the most "perfectly imperfect" on offer because (1) scads of people either use the internet or create content that finds its way onto the internet (or more likely both), (2) the information available online reflects people's use of terminology in a wide array of contexts and settings, from the sublime to the ridiculous, and (3) there's little reason (that I can think of) to worry that writers and speakers whose communications end up online manipulate the inputs (i.e., their words) in a way that might artificially skew the data.

Put simply, ordinary-meaning interpretation aims to capture how normal people use language in their everyday lives—and the bulk of the LLMs' training data seem to reflect exactly that.

2. LLMs can "understand" context. So far as I can tell, researchers powering the AI revolution have created, and are continuing to develop, increasingly sophisticated ways to convert language (and I'm not making this up) into math that computers can "understand." See Yonathan A. Arbel & David A. Hoffman, Generative Interpretation, 99 N.Y.U. L. Rev. (forthcoming 2024) (manuscript at 26) (describing "attention mechanism," a feature of LLMs that facilitates the recognition of how words are used in context). The combination of the massive datasets used for training and this cutting-edge "mathematization" of language enables LLMs to absorb and assess the use of terminology in context and empowers them to detect language patterns at a granular level. So, for instance, modern LLMs can easily discern the difference—and distinguish—between the flying-mammal "bat" that uses echolocation and may or may not be living in your attic, on the one hand, and the wooden "bat" that Shohei Otani uses to hit dingers, on the other. See id. And that, as I understand it, is just the tip of the iceberg. LLM predictions about how we use words and phrases have gotten so sophisticated that they can (for better or worse) produce full-blown conversations, write essays and computer code, draft emails to co-workers, etc. And as anyone who has used them can attest, modern LLMs' results are often sensible—so sensible, in fact, that they can border on the creepy. Now let's be clear, LLMs aren't perfect—and again, we'll discuss their shortcomings in due course. But let's be equally clear about what they are: high-octane language-prediction machines capable of probabilistically mapping, among other things, how ordinary people use words and phrases in context.

3. LLMs are accessible. LLMs are readily accessible (and increasingly so) to judges, lawyers, and, perhaps most importantly, ordinary citizens. In recent years, the use of LLMs has proliferated, and as with all other internet-related tools, one can only assume that usage will continue to accelerate, likely at an exponential rate. The LLMs' easy accessibility is important in at least two respects. First, it offers the promise of "democratizing" the interpretive enterprise, both (as already explained) by leveraging inputs from ordinary people and by being available for use by ordinary people. Second, it provides judges, lawyers, and litigants an inexpensive research tool. My "landscaping"-related queries, for instance, while no doubt imperfect, cost me nothing. To be sure, querying a more advanced LLM may come with a pricetag, at least for now. But so does, for example, searching the Oxford English Dictionary, the online version of which exists behind a paywall. And I'd be willing to bet that the costs associated with even the more advanced LLMs pale in comparison to subscriptions for Westlaw and Lexis, which power most modern legal research, including some involving dictionaries. And of course there's always the promise that open-source LLMs might soon approximate the for-profit models' productivity.

4. LLM research is relatively transparent. Using LLMs to facilitate ordinary-meaning interpretation may actually enhance the transparency and reliability of the interpretive enterprise itself, at least vis-à-vis current practice. Two brief observations.

First, although we tend to take dictionaries for granted, as if delivered by a prophet, the precise details of their construction aren't always self-evident. Who exactly compiles them, and by what criteria do the compilers choose and order the definitions within any given entry? To be sure, we're not totally in the dark; the online version of Merriam-Webster's, for instance, provides a useful primer explaining "[h]ow . . . a word get[s] into" that dictionary. It describes a process by which human editors spend a couple of hours a day "reading a cross section of published material" and looking for new words, usages, and spellings, which they then mark for inclusion (along with surrounding context) in a "searchable text database" that totals "more than 70 million words drawn from a great variety of sources"—followed, as I understand things, by a step in which a "definer" consults the available evidence and exercises his or her judgment to "decide[ ] . . . the best course of action by reading through the citations and using the evidence in them to adjust entries or create new ones."

Such explainers aside, Justice Scalia and Bryan Garner famously warned against "an uncritical approach to dictionaries." Antonin Scalia & Bryan A. Garner, A Note on the Use of Dictionaries, 16 Green Bag 2d 419, 420 (2013). They highlighted as risks, for instance, that a volume could "have been hastily put together by two editors on short notice, and very much on the cheap," and that without "consult[ing] the prefatory material" one might not be able "to understand the principles on which the dictionary [was] assembled" or the "ordering of [the] senses" of a particular term. Id. at 420, 423.

To be clear, I'm neither a nihilist nor a conspiracy theorist, but I do think that we textualists need to acknowledge (and guard against the fact) that dictionary definitions present a few known unknowns. See id. at 419-28; cf. Thomas R. Lee & Stephen C. Mouritsen, The Corpus and the Critics, 88 U. Chi. L. Rev. 275, 286-88 (2021) (highlighting potential interpretive pitfalls associated with dictionaries). And while I certainly appreciate that we also lack perfect knowledge about the training data used by cutting-edge LLMs, many of which are proprietary in nature, see supra notes 6 & 8, I think it's fair to say that we do know both (1) what LLMs are learning from—namely, tons and tons of internet data—and (2) one of the things that makes LLMs so useful—namely, their ability to accurately predict how normal people use language in their everyday lives.

A second transparency-related thought: When a judge confronts a case that requires a careful assessment of a word's meaning, he'll typically consult a range of dictionary definitions, engage in a "comparative weighing," Scalia & Garner, A Note, supra, at 422, and, in his written opinion, deploy one, two, or a few of them. The cynic, of course, will insist that the judge just dictionary-shopped for the definitions that would enable him to reverse-engineer his preferred outcome. See James J. Brudney & Lawrence Baum, Oasis or Mirage: The Supreme Court's Thirst for Dictionaries in the Rehnquist and Roberts Eras, 55 Wm. & Mary L. Rev. 483, 539 (2013). I'm not so jaded; I trust that ordinary-meaning-focused judges genuinely seek out definitions that best fit the context of the instruments that they're charged with interpreting. See, e.g., Hoever v. Marks, 993 F.3d 1353, 1366-68 (11th Cir. 2021) (en banc) (Newsom, J., concurring in judgment in part and dissenting in part) (choosing, based on contextual clues, from among competing definitions of the word "for"). Even so, I have to admit (1) that the choice among dictionary definitions involves a measure of discretion and (2) that judges seldom "show their work"—that is, they rarely explain in any detail the process by which they selected one definition over others. Contrast my M.O. in this case, which I would recommend as a best practice: full disclosure of both the queries put to the LLMs (imperfect as mine might have been) and the models' answers.

Anyway, I don't mean to paint either too grim a picture of our current, dictionary-centric practice—my own opinions are chock full of dictionary definitions, I hope to good effect—or too rosy a picture of the LLMs' potentiality. My point is simply that I don't think using LLMs entails any more opacity or involves any more discretion than is already inherent in interpretive practices that we currently take for granted—and in fact, that on both scores it might actually involve less.

5. LLMs hold advantages over other empirical interpretive methods . One final point before moving on. Recently, some empiricists have begun to critique the traditional dictionary-focused approach to plain-meaning interpretation. Some, for instance, have conducted wide-ranging surveys of ordinary citizens, seeking to demonstrate that dictionaries don't always capture ordinary understandings of legal texts. See, e.g., Kevin P. Tobia, Testing Ordinary Meaning, 134 Harv. L. Rev. 726 (2020). Others have turned to corpus linguistics, which aims to gauge ordinary meaning by quantifying the patterns of words' usages and occurrences in large bodies of language. See, e.g., Thomas R. Lee & Stephen C. Mouritsen, Judging Ordinary Meaning, 127 Yale L.J. 788, 795 (2018).

On balance, reliance on LLMs seems to me preferable to both. The survey method is interesting, but it seems wildly impractical—judges and lawyers have neither the time nor the resources to poll ordinary citizens on a widespread basis. By contrast, as already explained, LLMs are widely available and easily accessible. And corpus methods have been challenged on the ground, among others, that those tasked with compiling the data exercise too much discretion in selecting among the inputs. See, e.g., Jonathan H. Choi, Measuring Clarity in Legal Text, 91 U. Chi. L. Rev. 1, 26 (2024). For reasons already explained, I don't think LLM-based methods necessarily carry the same risk.

* * *

For all these reasons, and perhaps others I haven't identified, it seems to me that it's at least worth considering whether and how we might leverage LLMs in the ordinary-meaning enterprise—again, not as the be all and end all, but rather as one aid to be used alongside dictionaries, the semantic canons, etc.

B

Now, let's examine a few potential drawbacks. I suppose it could turn out that one or more of them are deal-killers. I tend to doubt it, but let's put them on the table.

1. LLMs can "hallucinate." First, the elephant in the room: What about LLMs' now-infamous "hallucinations"? Put simply, an LLM "hallucinates" when, in response to a user's query, it generates facts that, well, just aren't true—or at least not quite true. See, e.g., Arbel & Hoffman, supra, at 48-50. Remember the lawyer who got caught using ChatGPT to draft a brief when it ad-libbed case citations—which is to say cited precedents that didn't exist? See, e.g., Benjamin Weiser, Here's What Happens When Your Lawyer Uses ChatGPT, N.Y. Times (May 29, 2023). To me, this is among the most serious objections to using LLMs in the search for ordinary meaning. Even so, I don't think it's a conversation-stopper. For one thing, LLM technology is improving at breakneck speed, and there's every reason to believe that hallucinations will become fewer and farther between. Moreover, hallucinations would seem to be most worrisome when asking a specific question that has a specific answer—less so, it seems to me, when more generally seeking the "ordinary meaning" of some word or phrase. Finally, let's shoot straight: Flesh-and-blood lawyers hallucinate too. Sometimes, their hallucinations are good-faith mistakes. But all too often, I'm afraid, they're quite intentional—in their zeal, attorneys sometimes shade facts, finesse (and even omit altogether) adverse authorities, etc. So at worst, the "hallucination" problem counsels against blind-faith reliance on LLM outputs—in exactly the same way that no conscientious judge would blind-faith rely on a lawyer's representations.

2. LLMs don't capture offline speech , and thus might not fully account for underrepresented populations' usages . I flagged this one earlier, but I think it's a serious enough concern to merit separate treatment. Here's the objection, as I see it: People living in poorer communities (perhaps disproportionately minorities and those in rural areas) are less likely to have ready internet access and thus may be less likely to contribute to the sources from which LLMs draw in crafting their responses to queries. Accordingly, the argument goes, their understandings—as manifested, for instance, in their written speech—won't get "counted" in the LLMs' ordinary-meaning assessment.

As I say, I think this is a serious issue. Even so, I don't believe it fatally undermines LLMs' utility, at least as one tool among many for evaluating ordinary meaning. Ideally, of course, the universe of information from which any source of meaning draws would capture every conceivable input. But we should guard against overreaction. Presumably, LLMs train not only on data that were born (so to speak) online but also on material that was created in the physical world and only thereafter digitized and uploaded to the internet. And there is (I think) less reason to fear that those in underserved communities are at a dramatic comparative disadvantage with respect to the latter category. Moreover, to the extent we're worried about a lack of real-world, documentary evidence representing underrepresented populations' usages, then we have bigger fish to fry, because there's reason to doubt the utility of dictionaries, as well—which, as Merriam-Webster's editors have explained, also rely on hard-copy sources to evaluate terms' ordinary meanings. See supra at 1228-29 & note 12.

Anyway, the risk that certain communities' word-usage outputs aren't adequately reflected in LLMs' training-data inputs is real, and I'd note it as a candidate for improvement, but I don't think it's either fatal or insurmountable.

3. Lawyers , judges , and would-be litigants might try to manipulate LLMs. I suppose there's a risk that lawyers and judges might try to use LLMs strategically to reverse-engineer a preferred answer—say, by shopping around among the available models or manipulating queries. Maybe, but that's an evergreen issue, isn't it? Although they shouldn't, lawyers and judges can cast about for advantageous dictionary definitions and exploit the interpretive canons, but no one thinks that's a sufficient reason to abandon those as interpretive tools. And if anything, I tend to think that the LLMs are probably less vulnerable to manipulation than dictionaries and canons, at least when coupled with (as I've tried to provide here) full disclosure of one's research process.

Relatedly, might prospective litigants seek to corrupt the inputs—the data on which the LLMs train and base their responses to user queries—in an effort to rig the system to spit out their preferred interpretations? It's a real concern—perhaps especially considering that the same AI companies that have developed and are training the LLMs might themselves be litigants. But given the nature of the technology as I understand it, hardly insurmountable. For one thing, most models embody some training "cutoff"—for instance, though things might have changed, it was once common knowledge that GPT-4 learned on data up to and including September 2021. See Open AI, GPT-4 Technical Report 10 (arXiv:2303.08774, 2024). Accordingly, it would likely be difficult, if not impossible, to pollute the inputs retroactively. More fundamentally, it seems almost inconceivable that a would-be malefactor could surreptitiously flood any given dataset with enough new inputs to move the needle—remember, just by way of example, that GPT-3.5 Turbo trained on more than 400 billion words. Finally, while I tend to doubt that any AI company would conclude that corrupting its own product in order to obtain an interpretive advantage in a single case was in its long-term business interest, that risk, it seems to me, could be mitigated, if not eliminated, by querying multiple models rather than just one.

4. Reliance on LLMs will lead us into dystopia. Would the consideration of LLM outputs in interpreting legal texts inevitably put us on some dystopian path toward "robo judges" algorithmically resolving human disputes? I don't think so. As Chief Justice Roberts recently observed, the law will always require "gray area[ ]" decisionmaking that entails the "application of human judgment." Chief Justice John G. Roberts, Jr., 2023 Year-End Report on the Federal Judiciary 6 (Dec. 31, 2023). And I hope it's clear by this point that I am not—not, not, not—suggesting that any judge should ever query an LLM concerning the ordinary meaning of some word (say, "landscaping") and then mechanistically apply it to her facts and render judgment. My only proposal—and, again, I think it's a pretty modest one—is that we consider whether LLMs might provide additional datapoints to be used alongside dictionaries, canons, and syntactical context in the assessment of terms' ordinary meaning. That's all; that's it.

IV

Which brings me to my final question: If I'm not all wet, and it's at least worth considering whether LLMs have a role to play in the interpretation of legal instruments, how might we maximize their utility? I've already flagged a few suggestions for improvement along the way—more data, from more sources, representing a more representative cross-section of Americans. But beyond the obvious, what else?

First , I think it'll be helpful to clarify the objective. Remember that in my clumsy first crack at this, I asked two different models two different questions: (1) "What is the ordinary meaning of 'landscaping'?"; and (2) "Is an in-ground trampoline 'landscaping'?" Which is the proper question? In retrospect, if my contention is—as it is—that LLMs might aid in the search for the ordinary, everyday meaning of common words and phrases, then it seems pretty clear to me that my first, more general query is the more appropriate one. The models' highest and best use is (like a dictionary) helping to discern how normal people use and understand language, not in applying a particular meaning to a particular set of facts to suggest an answer to a particular question.

Second , and relatedly, how can we best query LLMs? Those in the know refer to the question a user asks a model as a "prompt." I'll confess that I gave relatively little thought to my own prompts—they were just the questions that immediately sprang to mind. But research indicates that the models can be sensitive to prompts and that the results can vary accordingly. See, e.g., Arbel & Hoffman, supra, at 36. So it may be wise for users to try different prompts, and, importantly, to report the prompts they use and the range of results they obtain. Id. at 36-37. Better still to do all that and query multiple models to ensure that the results are consistent—or, in statistics-speak, "robust."

Third , we need to clarify the particular output we're after. The questions I asked sought a discrete, one-time answer. In particular, I asked for a single definition of "landscaping" and, separately, whether installation of an in-ground trampoline qualified. One potential challenge is that this approach obscures the fact, already explained, that LLMs make probabilistic, predictive judgments about language. With that in mind, some who have considered how LLMs might be used to interpret contracts have suggested that users seek not just answers but also "confidence" levels. See id. at 23. So, for instance, an LLM might reveal that its prediction about a provision's meaning is "high" or, by contrast, only "ambiguous." Alternatively, but to the same end, a researcher might ask an LLM the same question multiple times and note the percentage of instances in which it agrees that, say, installation of an in-ground trampoline is landscaping. See Christoph Engel & Richard H. McAdams, Asking GPT for the Ordinary Meaning of Statutory Terms 15 (Max Planck Inst. Discussion Paper 2024/5).

Fourth and finally, there are temporal considerations to mull. The ordinary-meaning rule has an important corollary—namely, that "[w]ords must be given the meaning they had when the text was adopted." Scalia & Garner, Reading Law, supra, at 78 (emphasis added). That principle—"originalism," if you will—most obviously applies to constitutional and statutory texts. See, e.g., United States v. Pate, 84 F.4th 1196, 1201 (11th Cir. 2023) (en banc) ("[W]hen called on to resolve a dispute over a statute's meaning, [a court] normally seeks to afford the law's terms their ordinary meaning at the time Congress adopted them." (quoting Niz-Chavez v. Garland, 593 U.S. 155, 160, 141 S.Ct. 1474, 209 L.Ed.2d 433 (2021)). To be clear, though, it holds for private instruments like contracts and insurance policies, as well. See Scalia & Garner, Reading Law, supra, at 78. This case didn't present a particularly knotty temporal difficulty because the policy in question was issued relatively recently, in 2017. But cases routinely arise in which the relevant question is what a particular term meant in the more distant past—say, in 1787, 1868, or 1964. If LLMs are to be deployed to aid more broadly in the search for ordinary meaning, it would be enormously helpful—if it's not already possible—for AI engineers to devise a way in which queries could be limited to particular timeframes. Cf. Engel & McAdams, supra, at 34-37.

V

In his most recent year-end report on the state of the federal judiciary, Chief Justice Roberts cautioned that the "use of AI requires caution and humility." Roberts, supra, at 5. I wholeheartedly agree. Importantly, though, I also agree with what I take to be the report's assumption that AI is here to stay. Now, it seems to me, is the time to figure out how to use it profitably and responsibly. It's in that spirit that I've offered these preliminary thoughts about whether and how LLMs might aid lawyers and judges in the interpretive enterprise. Plenty of questions remain, and I'm sure I haven't even identified all of them. But—and this is my bottom line—I think that LLMs have promise. At the very least, it no longer strikes me as ridiculous to think that an LLM like ChatGPT might have something useful to say about the common, everyday meaning of the words and phrases used in legal texts.

Just my two cents.

Appendix

Below are the questions that I asked OpenAI's ChatGPT and Google's Bard models, along with their responses.

ChatGPT

Question: What is the ordinary meaning of "landscaping"? Response: "Landscaping" refers to the process of altering the visible features of an area of land, typically a yard, garden or outdoor space, for aesthetic or practical purposes. This can include activities such as planting trees, shrubs, flowers, or grass, as well as installing paths, fences, water features, and other elements to enhance the appearance and functionality of the outdoor space. Question: Is installing an in-ground trampoline "landscaping"? Response: Yes, installing an in-ground trampoline can be considered a part of landscaping. Landscaping involves altering the visible features of an outdoor area for aesthetic or practical purposes, and adding an in-ground trampoline would modify the appearance and function of the space. It's a deliberate change to the outdoor environment, often aimed at enhancing the overall landscape and usability of the area.

Bard

Question: What is the ordinary meaning of "landscaping"? Response: The ordinary meaning of landscaping" is the process of modifying the visible features of an area of land, including: • Adding or removing vegetation, such as trees, shrubs, and flowers • Installing hardscape elements, such as walkways, patios, and retaining walls • Creating or enhancing natural features, such as ponds, waterfalls, and rock gardens Landscaping can be used to improve the aesthetics of an area, increase its functionality, and/or protect it from erosion. It can also be used to create a sense of privacy, define spaces, and add value to a property.1235Here are some examples of landscaping projects: • Planting a new lawn or garden • Adding a patio or walkway • Building a retaining wall • Installing a sprinkler system • Creating a pond or waterfall • Planting trees and shrubs • Adding mulch or gravel • Edging a flower bed • Trimming hedges Landscaping can be a DIY project, or it can be hired out to a professional landscaping company. The cost of landscaping will vary depending on the size and scope of the project. Question: Is installing an in-ground trampoline "landscaping"? Response: Whether or not installing an in-ground trampoline is considered landscaping depends on how you define the term. If you define landscaping broadly to include any modification to the visible features of an area of land, then installing an in-ground trampoline would be considered landscaping. However, if you define landscaping more narrowly to focus on the planting and maintenance of vegetation, then installing an in-ground trampoline would not be considered landscaping. Here is a table that summarizes the different ways to define landscaping and whether or not installing an in-ground trampoline would be considered landscaping under each definition: Definition Includes installing an in-ground trampoline? Broad definition: Any modification to the visible features of an area of land Yes Narrower definition: Planting and maintenance of vegetation No Ultimately, whether or not installing an in-ground trampoline is considered landscaping is a matter of opinion. There is no right or wrong answer.