- in United States
- within International Law, Finance and Banking and Real Estate and Construction topic(s)
- with readers working within the Consumer Industries industries
In this blog series, we explore the promises and the pitfalls of AI tools in the insurance coverage context, offering practical guidance for lawyers and business professionals to harness these technologies without getting burned.
If you practice insurance coverage law, you've been there: staring at an undefined term in a policy, toggling between three dictionaries that each say something slightly different, and wondering whether the word means what you think it means or what opposing counsel insists it means. The ordinary-meaning inquiry — foundational to coverage disputes across virtually every jurisdiction — is often more art than science. Dictionaries help, but they don't always resolve close calls.
That's what makes Judge Kevin Newsom's concurrence in Snell v. United Specialty Insurance Co., 102 F.4th 1208, 1221-234 (11th Cir. 2024), so worth reading. In what he acknowledges is an unusual concurring opinion, Judge Newsom pulls back the curtain on his own interpretive process and offers an early judicial framework for using large language models (LLMs) — tools like ChatGPT, Gemini, and Claude — to help determine the ordinary meaning of policy language. Whether you find that exciting or unsettling, Snell reflects a development for the legal community's use of LLMs that coverage attorneys should understand.
What Happened in Snell
The underlying dispute was straightforward insurance coverage fare. The insurer limited coverage to specified operations — here it was "landscaping." Snell, the insured, a landscaping company, had dug a pit, built a retaining wall, and installed a trampoline with a wooden cap. A child was later injured on the trampoline, and the child's family sued Snell. Snell tendered the claim to his insurer under a general commercial liability policy. The insurer denied coverage, reasoning that the injury didn't arise from Snell's "landscaping" — the only specified operation listed in the policy.
The majority affirmed summary judgment for the insurer, but it did so largely by sidestepping the definitional fight. Under Alabama law, Snell's application was incorporated into the insurance policy, and in that application, Snell had denied that his work included recreational equipment construction. That was enough to resolve the case without settling whether the trampoline installation counts as "landscaping."
Judge Newsom joined the majority opinion. But he wrote separately because, before that off-ramp became clear, he had spent considerable time wrestling with the ordinary meaning of "landscaping" — and in that process, he'd done something he initially thought was absurd: He asked ChatGPT for help.
The Framework: Why LLMs Deserve a Seat at the Table
What makes Judge Newsom's concurrence valuable for practitioners isn't just that an Eleventh Circuit judge consulted an AI model. It's that Judge Newsom methodically discussed the potential use of LLMs as a legitimate interpretive tool and identified the guardrails that should accompany their use. Here are the key points he raised:
1. LLMS Are Trained on How People Actually Talk
The ordinary-meaning rule rests on the premise that words in legal instruments should be understood the way that regular people use them. LLMs are trained on enormous quantities of real-world language — internet text spanning professional websites, news articles, blogs, forums, government records, and casual conversation. Chat GPT-3.5 Turbo alone trained using roughly 400 to 500 billion words. That dataset, Judge Newsom observed, may be better suited to capturing everyday usage than a dictionary compiled by a small team of editors reading selected published sources to determine definitions.
For coverage lawyers, think of it this way: When arguing whether "collapse" requires actual structural falling down or includes substantial impairment, interpretation typically considers the understanding of ordinary policyholders — and opposing counsel — to understand the word. An LLM's response is drawn from a massive sample based on actual usage of the term by millions of people.
2. LLMs Understand Context — Not just Definitions
Dictionaries provide a list of possible meanings, but LLMs can do something more: They can assess a word's function in a particular context. Judge Newsom describes modern LLMs as "high-octane language-prediction machines" capable of probabilistically mapping ordinary people's usage of words and phrases in context. He illustrated the point with the word "bat" — modern LLMs can easily distinguish between the flying mammal that may be living in the attic and the wooden implement used to hit home runs. That contextual sensitivity is directly relevant to ordinary-meaning analysis, where the goal is not just to catalog a word's possible definitions but to identify the meaning that fits the instrument at hand.
3. LLMs Are Accessible and Inexpensive
Judge Newsom noted a practical advantage that matters in the real world of litigation: LLM queries cost little or nothing, especially compared to legal subscriptions, expert linguists, or the survey-based methods some academics have proposed for gauging ordinary meaning. Any lawyer — or judge — can run a query in minutes. That accessibility also has a democratizing effect, potentially giving smaller firms and solo practitioners access to interpretive insights that might otherwise require expensive expert testimony.
4. LLMs May Actually Be More Transparent Than Dictionaries
This is a counterintuitive but compelling point. Courts and litigants tend to treat dictionaries as authoritative without scrutinizing their underlying premises — who chose the definitions, what sources the editors reviewed, and their ordering of multiple senses of a word. Judge Newsom cited Justice Antonin Scalia and legal scholar and lexicographer Bryan Garner's own warnings about the risks of uncritical reliance on dictionaries. By contrast, an LLM query can be fully disclosed: the exact prompt, the model used, and the complete response. Judge Newsom modeled this transparency by appending his ChatGPT and Bard queries and their full responses to his opinion. That kind of full disclosure approach with LLMs may reduce the opacity that critics have long associated with dictionary shopping.
5. LLMs Hold Advantages Over Other Empirical Methods
Judge Newsom also compared LLMs favorably to two other empirical approaches that legal scholars have proposed: public surveys and corpus linguistics. Surveys are interesting in theory but wildly impractical for everyday litigation — no judge has time to poll citizens about policy language. Corpus linguistics offer rigorous quantitative analysis but has been criticized for the discretion involved in selecting datasets. LLMs, Judge Newsom suggested, offer a middle path: empirically grounded, practically accessible, and less susceptible to selection bias.
The Caveats Worth Knowing
Judge Newsom identified several risks: hallucinations (where LLMs sometimes generate plausible sounding but fabricated information), the underrepresentation of offline communities in training data, the potential for strategic manipulation of queries, and the concern that reliance on AI could put us on a path toward "robo judges." On each point, he argued that the risk was real but manageable — and often no worse than the risks already inherent in dictionary-based interpretation.
He also offered concrete suggestions for best practices: Use general definitional queries rather than asking the model to resolve the ultimate legal question; try multiple prompts and manipulate models to test robustness; seek confidence levels or run repeated queries to surface the probabilistic nature of the results; and always disclose the methodology utilized.
What This Means for Coverage Attorneys
Judge Newsom's concurrence is not binding law, and he was explicit that he wasn't suggesting judges mechanistically defer to AI outputs. His proposal is modest: LLMs should be considered as one tool alongside dictionaries, canons of construction, and contextual analysis in the ordinary-meaning toolkit.
But for insurance coverage attorneys, the practical implications are real. If judges start consulting LLMs — and Judge Newsom's opinion signals that at least some already are — then practitioners should be prepared to engage with that evidence. That means understanding how to craft effective prompts, knowing the strength and limitations of different models, and being ready to challenge or corroborate an LLM's output with traditional sources.
More immediately, there's nothing stopping attorneys from using LLMs right now in their own coverage analyses. When drafting a brief about the ordinary meaning of "sudden," "occurrence," "professional services," or any other disputed policy term, an LLM query can serve as a quick gut-check against dictionary research — or it might surface a usage pattern not previously considered. Just as Judge Newsom acknowledges many judges are already using LLMs, coverage attorneys must consider that insurers and opposing counsel likewise incorporate LLMs into their own work. The more that coverage attorneys can anticipate LLM outputs on the ordinary meaning of key policy language, the better positioned they may be to argue for coverage.
Judge Newsom ended his concurrence by remarking that AI is here to stay and that the time is now to figure out how the legal community should use LLMs profitably and responsibly. While plenty of questions remain about the use of LLMs in the interpretive enterprise, Judge Newsome concluded, "[b]ut—and this is my bottom line—I think LLMs have promise."
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
