The Plain and Ordinary Meaning of Insurance Contracts in a Large Language Model World

August 2026 By Ronald M. Sandgrund

Definitions of words and phrases generated by artificial intelligence large language models may influence how courts interpret insurance policies and other contracts.

article citation

This article examines whether courts can employ large language models (LLMs) like those produced by Anthropic (Claude), Google (Genesis), Open AI (ChatGPT), xAI (Grok), and others as tools for determining the meaning of words and phrases used in insurance policies. A provocative concurring opinion from the Eleventh Circuit case Snell v. United Specialty Insurance Co.1 raises and examines this issue, and its implications are further explored here. A detour into the LLM world raises serious questions of “accuracy and reliability, efficiency, accountability and transparency, fairness and bias, and security and privacy.”2

The Plain and Ordinary Meaning of Insurance Policies

Colorado’s long-standing black letter rule is that words used in insurance policies, as with other contracts, should be given their plain and ordinary meaning unless the policy expressly shows that the parties intended an alternative interpretation.3 An insurance contract’s meaning is not determined by “experts” in the construction of insurance contracts.4 Rather, an insurance contract must be construed by reference to the meaning a person of ordinary intelligence would attach to it.5 An “insurance company has an obligation to write the critical portions of its policies in plain English and with such precision that a reasonable lay person can, by reading the policy, understand the coverage provided.”6 Colorado courts rarely examine the intent of the parties to an insurance contract. Instead, they usually resolve any uncertainty or ambiguity concerning coverage in favor of the policyholder. While a huge body of case law and commentary exists regarding how to decide if a term or phrase is ambiguous, and how to resolve that ambiguity, such discussion is beyond the scope of this article.

The Meaning of Technical or Specialized Terms

As with any general rule, there are exceptions. The policy itself may provide definitions for specific terms and phrases,7 and extrinsic evidence is sometimes considered when construing technical or specialized terms.8

Reliance on Dictionary Meanings

Courts may rely on dictionary definitions to determine the plain and ordinary meaning of words.9 Sometimes different dictionaries provide differing definitions, and courts must determine which meaning best fits the context where a term is susceptible to more than one meaning. When a term is ambiguous, the term will be construed in favor of the insured and coverage.

The Emergence of Large Language Models

LLMs like ChatGPT 3.5 seemed like cool toys only a few years ago. Since then, they have evolved substantially (ChatGPT 3.5 has been supplanted by ChatGPT 5.5 as of June 2026). They are more powerful, their reach and limitations are better understood, and they have begun to reshape our world, our work, and our relationships. Although ostensibly constituting nothing more than sophisticated pattern recognition algorithms, programmed using and updated regularly on massive databases, they can seem like a modern miracle, mimicking human thought, speech, intuition, and emotion. Some believe—and not just futurists, science fiction writers, and cranks—that LLMs could become autonomous and sentient. But, in the meantime, here on planet Earth, they remain a useful tool, albeit subject to error and bias. Still, do they have a place in legal thought and analysis?

Judge Newsom’s Provocative Concurrence in Snell v. United Specialty Insurance Co.

In Snell, the Eleventh Circuit Court of Appeals held that where a landscaping company had dug a pit, built a retaining wall, and installed a trampoline with a decorative wooden cap, the insured’s alleged liability for a later resulting trampoline injury was not encompassed by the “landscaping” hazard coverage provided in the insured’s policy because, in its insurance application, the insured had denied engaging in “any recreational or playground equipment construction.”10 This part of the opinion is unremarkable.

However, in what this author found to be a remarkable, lucid, somewhat self-deprecating, and smile-inducing (although maybe only for insurance geeks) concurrence, Judge Newsom posited that AI LLM definitions of words and phrases may offer an alternative and more accurate perspective than dictionaries on the common and ordinary meaning of those terms.11 He acknowledged that the court’s holding mooted his analysis, but onward he marched into the world of LLMs, asking the reader to humor and accompany him on his journey.12 He admits being “unabashedly a plain-language guy,” and that “[t]he ordinary meaning rule is the most fundamental semantic rule of interpretation,” quoting former Justice Scalia.13 After surveying a “buffet of [dictionary and encyclopedia] definitions,” he felt the exercise “left a little something to be desired,” since, “[f]rom their definitions alone, it was tough to discern a single controlling criterion.”14 He had a clerk query the meaning of “landscaping” on ChatGPT and had an “aha” moment, observing: “ChatGPT’s explanation seemed more sensible than I had thought it might—and definitely less nutty than I had feared.”15

Judge Newsom expanded on his thinking, again in a concurrence, in United States v. Deleon.16 There, the Eleventh Circuit needed to determine whether the defendant’s victim was “physically restrained” under the US Sentencing Guidelines Manual, § 2B3.1(b)(4)(B), thereby justifying a sentencing enhancement.17 While the defendant pointed a gun and threatened the cashier-victim, he never touched the cashier, who remained behind the counter.18 Still, the court affirmed the enhanced sentence, concluding that it was bound by precedent.19

Concurring with the result, Judge Newsom again advocated for courts to consider the possible usefulness of LLMs, this time to help find the ordinary meaning of a composite phrase (“physically restrained”), rather than a single word (“landscaping”) as discussed in his prior concurrence.20 Judge Newsom pointed out that phrases are not typically defined by dictionaries, but are often broken down into their constituent word elements by courts, and then pieced back together “into a coherent whole.”21 Judge Newsom queried various LLMs for the meaning of “physically restrained” and found that the responses not only varied from LLM to LLM, but also varied within each LLM when identical queries were provided multiple times.22 Judge Newsom acknowledges being both surprised and concerned by these results (they “rocked [him] back on [his] heels”23), but, after further research and consideration, he realized that (1) LLM responses are “probabilistic,” meaning that slightly differing responses are both expected and intended from a technical standpoint, and that these are a “feature” not a “bug” of LLMs; and (2) although the LLM responses he got “exhibited some minor variations in structure and phrasing,” these “marginal divergences” are expected, and the responses “coalesce[d], substantively, around a common core—there was an objectively verifiable throughline.”24 In the end, Judge Newsom concluded that the varying but similar LLM query responses “closely mimic what we would expect to see, and in fact do see, in everyday speech patterns.”25 In other words, LLM variability in assessing “ordinary meaning” is a “virtue rather than a vice.”26 Judge Newsom concludes that “LLMs may well serve a valuable auxiliary role as we aim to triangulate ordinary meaning.”27

LLM Advantages

Judge Newsom identifies various bases supporting the use of LLMs when interpreting words in insurance contracts, among other contracts and writings. LLMs (1) are trained on how people actually talk; (2) understand context, not just definitions; (3) are accessible and inexpensive; (4) may be more “transparent” than dictionaries; and (5) hold advantages over other empirical methods.28 A discussion of these benefits follows.

Actual Usage

LLMs are programmed based on the actual words found in internet text across disparate websites, including news articles, blogs, government records, professional websites, and regular conversation. These datasets may be superior to dictionaries in identifying ordinary usage. Dictionaries rely on small editorial teams reviewing a subset of published sources to establish definitions. In theory, an LLM definition of a term may be drawn from hundreds of billions of text samples based on actual usage by millions and millions of people—text generated by the highbrow and the lowbrow and everyone in between.29 Thus, the LLM may be better at identifying the “ordinary meaning” of words as used in “common speech.” Still, these massive databases do not tap into “pure offline uses,” such as private correspondence and conversations, which are absent from LLMs training pools of text.30

Context

Another advantage LLMs may have over dictionaries is their ability to assess how words are used in context—Judge Newsom describes them as “high-octane language-prediction machines.”31 “GPT” means “generative pre-trained transformer”—a computer program that creates new content after being trained on massive datasets to understand language patterns and uses a specialized “neural network architecture” to understand context and relationships within text.32 Contextual “sensitivity” allows an LLM to find a meaning that fits how a term is used in the actual contract a court is construing rather than all the possible meanings of that word. LLMs are designed “to detect language patterns at a granular level,” and these patterns illuminate what words mean in various contexts, so that they understand the difference between an echo-locating flying mammal and the lever Aaron Judge uses to smash dingers.33

Access and Expense

Ignoring the monthly cost of subscribing to one of the latest and greatest LLM models, individual queries cost users pennies or nothing to employ, and running queries and obtaining responses happen at the speed of electricity. The cost of accessing specialized legal databases or employing linguistic experts or conducting public surveys can be avoided. LLMs “leverage” inputs from ordinary people while also being available to ordinary people, thereby “‘democratizing’ the interpretative enterprise.”34 While online dictionaries are similarly cheap, access to the most comprehensive dictionaries’ resources often lies behind a paywall.35

A Better Tool

Although debatable, Judge Newsom forcefully argues that LLMs are more “transparent” and “reliable” than dictionaries.36 While many treat dictionaries as authoritative, most people do not appreciate who decides a given word’s definition, their qualifications and biases, what sources they draw on, and how they order the various meanings a given word may have within the definition itself. Although most reputable dictionaries explain how they are compiled and how human judgment is employed to arrive at a word’s definition or definitions, the process is less transparent than it may appear. Then, afterward, judges may engage in “comparative weighing” of definitions that vary across the dictionary corpus (an exercise a cynic might refer to as “dictionary shopping”).37

While Judge Newsom recognizes our similar lack of “perfect knowledge” about the training data and algorithms that LLMs employ, he argues that one should consider the real possibility that LLMs may be a powerful tool to add to a judge’s interpretive arsenal.38 Moreover, the LLM query that elicited a term’s meaning can be revealed: the precise prompt, the LLM model employed, and the LLM’s verbatim response. Also, various LLM model responses can be compared and contrasted, and the same query can be run to test its consistency. The LLM can be asked to rate its “confidence” level in its response and even identify weaknesses in its programming and methodology.

LLM Risks

Judge Newsom also identifies various reasons why the use of LLMs should be employed with caution: (1) their propensity for “hallucinations,” (2) the underrepresentation of offline communities, (3) the risk of strategic manipulation of queries, and (4) the risk of letting machines assume human decision-making responsibilities.39 A discussion of these concerns follows.

Hallucinations

Judge Newsom agrees that LLMs’ well-publicized “hallucinations,” including the fabricated legal citations that have showed up in some briefs, are a real problem.40 However, he suggests that as technology improves and users become more aware, this problem will dissipate over time, and that it is less of an issue when seeking the ordinary meaning of a word or phrase than when searching for a specific answer to a specific question.41

Underrepresentation

Datasets used to train LLMs may not adequately reflect the ordinary meaning of words used by people with limited internet access—a community that Judge Newsom postulates is disproportionately comprised of “minorities” (his word) and rural dwellers.42 But, he maintains that this same problem taints dictionary definitions.43 Still, with an increase in the number and quality of speech-to-text transcription and translation programs, those who might never have put pen to paper now have a conduit to add their speech to the internet’s expanding dataset.

Manipulation

It is possible that judges, lawyers, sophisticated litigants, and others may seek to poison the well by using “LLMs [to] strategically [] reverse-engineer a preferred answer,” by “shopping around among the available models or manipulating queries.”44 Of course, a similar problem exists in a non-LLM world, as lawyers and judges have been known to “cast about for advantageous dictionary definitions and exploit the interpretive canons.”45 On balance, Judge Newsom concludes that LLM responses are less subject to manipulation if their use is accompanied by adequate and detailed disclosure.46

Also, parties may try to distort the data upon which LLMs train and base their query responses, especially if those parties are the companies that generate the LLMs, as they are often the target of lawsuits and may have a stake in a query’s response.47 Still, an LLM’s architecture, such as using dataset training “cutoff” dates, may mitigate against retroactive manipulation.48 Also, there are practical problems in successfully manipulating a dataset consisting of billions of data points—and, if a whistleblower were ever to publicize such an effort, that LLM’s credibility, and its owner’s stock value, would surely take a hit.

Ex Machina

Judge Newsom’s final caution was whether relying on LLM output might set us down the royal road to robot judges.49 This result seems highly unlikely since human judgment and discretion are central to our ideal of justice and to fair dispute resolution. Moreover, Judge Newsom is not suggesting that LLMs should be employed differently from any other judicial or legal tool—but simply as an aid to and not a substitute for wisdom and the human element.

Best Practices

If LLMs are used in the legal arena to aid in interpretation, Judge Newsom suggests some ways to maximize their utility: (1) clarify one’s objective and fashion one’s query appropriately; (2) because LLMs are highly sensitive to prompts and often respond with varying answers, try different prompts and disclose the prompts and results; (3) seek not just answers, but also have the LLM rate the confidence level with which it calculates the accuracy of its response, such as high, low, or ambiguous; and (4) consider the fact that the meaning of words may change over time, so that the dataset upon which an LLM is trained may “bound” its response.50 This “temporality” issue is discussed more fully below.

Temporality

A thorny issue is the fact that LLMs are constantly being updated by regularly scraping the internet and other sources of information in an ongoing effort to stay current and expand their knowledge base. Moreover, their underlying algorithms may also be simultaneously tweaked. In today’s world of “originalists,” who seek the original intended meaning of words as used in the time and place of their utterance or publication, there may be a need for a “time machine” function that limits query responses to data generated before a particular date or during a particular time frame.

Convergent “Facts”

Judge Newsom’s discussion and analysis implicate a related concept emerging from the widespread use of LLMs. This is the possibility LLMs are creating a less biased source of information distinct from the internet silos and informational echo chambers that pervade our world.51 As one commentator argues, “LLMs are a kind of anti-social media,”52 opening the door for people to access less purposefully biased information.53 Access to alternate conclusions unmoored from the so-called left and right wings and mainstream media information sources that pervade our world and shape our perspectives may prove to be useful.

Anecdotal LLM Test

In the fall of 2023, this author wrote a two-part article on the emerging use of LLMs in legal practice.54 Given how much LLMs have advanced, the article seems like it was written so very, very long ago: much about LLMs has changed and improved, and the universe of their promise and peril has expanded.

Since that article was published during the summer of 2023, the author has conducted a regular and quite amateurish test of LLM capabilities. He has asked various LLMs to summarize the history and rules of “Wilderness Frisbee,” a not very well-known game for which the author believes there is only one published source of information: the August 2025 issue of Colorado Lawyer.55 In a sense, the query has asked for the meaning of the term “Wilderness Frisbee.” To the author’s amusement (and bemusement), for nearly three years no two queries out of the dozens he inputted produced the same response, even from the same LLM. Moreover, with each succeeding month, the LLM responses incorporated more and more hallucinations—often to the point where the game and its creators were almost unrecognizable.56 Maybe this was a lousy or unfair test—the relevant dataset is tiny, so all I got in response to my queries was a “scrape” of the article, with the LLM “filling in” the blanks it gleaned with its best prediction of what the missing text said. However, to my surprise, while editing this article during June 2026, the hallucinations disappeared and the three leading LLMs generated accurate and consistent responses to my queries. Still, at a minimum, this experience highlights a potential vulnerability and weakness in turning to LLMs for help.

Conclusion

Technology is progressing, if not accelerating, much more quickly than people can imagine.57 The future is coming at us faster than ever. Judge Newsom warns that the future of AI is already here, and that lawyers and judges need to understand it more fully, and to appreciate its power and influence within the legal system so it can be used responsibly, even if a lot of questions about its use and abilities remain.58

Epilogue

Just before this article was heading to the typesetter, Judge Newsom and his chief law clerk, Alana Frederick, co-authored “Meaning, Understanding, and Contextual Textualism.”59 In that law review article, they argue that there is a significant difference between the “ordinary meaning” and the “ordinary understanding” of words and phrases, and that context distinguishes the two.60 They posit that “contextual textualism” recognizes “that extratextual presuppositions are often subliminally and inextricably baked into” one’s understanding of the written word—whether that word appears in contracts, statutes, or our Constitution.61

The authors remain convinced that LLMs offer a unique potential and may provide useful insight into the ordinary person’s understanding of words and phrases.62 In exploring their thesis, the authors maintain their sense of humor by analyzing US Supreme Court Justice Breyer’s “Great Snail Debate” (Does a teacher carrying a basket of live snails on a train need to buy the snails a ticket since the train’s policy requires tickets for animals?)63 and Professor H.L.A. Hart’s “No vehicles in the park!” legendary hypothetical (Are bikes and toy cars “vehicles” subject to being fined?).64 The authors conclude that “meaning is clean, abstract, and academic,” while “understanding is organic and gritty”; that “on-the-ground conversational and cultural context” matters; and that LLMs may provide a path to a greater understanding of the meaning of words and phrases.65

Ronald M. Sandgrund is of counsel with the construction defect group of Burg Simpson Eldredge Hersh Jardine PC. The group represents commercial and residential property owners, homeowner associations and unit owners, and construction professionals and insurers in construction defect, product liability, and insurance coverage disputes. He is a frequent author and lecturer on these topics, as well as on the practical aspects of being a lawyer. He has taught courses on entrepreneurial innovation and public policy and trial advocacy, and has lectured on legal ethics, construction law, mass tort litigation, consumer rights, and other subjects at Colorado Law—rms.sandgrund@gmail.com. Coordinating Editor: Jennifer Seidman, jseidman@hearnfleener.com.

Notes

citation Sandgrund, “The Plain and Ordinary Meaning of Insurance Contracts in a Large Language Model World,” 55 Colo. Law. 52 (Aug. 2026), https://cl.cobar.org/features/the-plain-and-ordinary-meaning-of-insurance-contracts-in-a-large-language-model-world.

1. Snell v. United Specialty Ins. Co., 102 F.4th 1208, 1221–34 (11th Cir. 2024).

2. Avery, “Introducing the ‘Hard’ Problem of Artificial Legal Intelligence,” 101 NYU L. Rev. Online 116, 117–18 (Apr. 2026), https://nyulawreview.org/online-features/introducing-the-hard-problem-of-artificial-legal-intelligence. See also Grimmelman et al., “Generative Misinterpretation,” 63 Harv. J. on Legis. 229 (Winter 2026) (arguing that “LLMs are not yet fit for use in judicial chambers” due to reliability (consistency) and epistemic (validity) gaps, rendering them socially illegitimate, and diving deeply into how LLMs actually work from a technical and programming standpoint in founding their criticism). Grimmelman et al. criticize Arbel and Hoffman, “Generative Interpretation,” 99 N.Y.U. L. Rev. 451 (2024), upon which Judge Newsom’s concurrence in Snell, 102 F.4th 1208, partly relied.

3. See Compass Ins. Co. v. Littleton, 984 P.2d 606, 613 (Colo. 1999) (quoting Chacon v. Am. Fam. Mut. Ins. Co., 788 P.2d 748, 750 (Colo. 1990)).

4. See Simon v. Shelter Gen. Ins. Co., 842 P.2d 236, 240 (Colo. 1992).

5. Id. See also Compass Ins. Co., 984 P.2d at 617 (policy to be “interpreted according to the understanding of the average purchaser of insurance”).

6. Travelers Indem. Co. v. Howard Elec. Co., 879 P.2d 431, 434 (Colo.App. 1994).

7. Turner, Insurance Coverage of Construction Disputes § 1:11 (2d ed. Dec. 2025 update) (“Most policies provide the definitionsof certain terms of art used in the policy in an effort by insurers to avoid policyholders’ claims of ambiguity”).

8. See Travelers Indem. Co., 879 P.2d at 434–35 (insurance industry term “allocated loss adjustment expense” is ambiguous, warranting admission of extrinsic evidence to ascertain the parties’ understanding of the agreement). See also Sorensen v. USAA Cas. Ins. Co., No. 24CA0035 ¶ 8, 2024 WL 4850221, at *2 (Colo.App. Nov. 21, 2024) (unpublished) (Colorado gives insurance policy words and phrases “their plain, everyday meaning and may not force strained constructions”; “[b]ut when a contract is within a trade or technical field, like insurance, and unless a different intent is manifested, technical terms and words of art are given their technical meaning when used in a transaction within their technical field.”) ((citation modified))).

9. Miller v. Hartford Cas. Ins. Co., 160 P.3d 408, 410–11 (Colo.App. 2007) (court looked to three different dictionary definitions of “long-term”). See also Hecla Mining Co. v. N.H. Ins. Co., 811 P.2d 1083, 1091 (Colo. 1991) (court turned to three different dictionary definitions to clarify undefined word “sudden” as used in insurance policy); Cyprus Amax Mins. Co. v. Lexington Ins. Co., 74 P.3d 294, 304 (Colo. 2003) (looking to dictionary definition to explain the ambiguous term “property damage”).

10. Snell, 102 F.4th at 1222.

11. Id. at 1221–34.

12. Id. at 1221.

13. Id. at 1222 (quoting Scalia and Garner, Reading Law: The Interpretation of Legal Texts 69 (Thomson/West 2012)).

14. Snell, 102 F.4th at 1223.

15. Id. at 1225.

16. United States v. Deleon, 116 F.4th 1260 (11th Cir. 2024).

17. Id. at 1261.

18. Id. at 1261–64.

19. Id.

20. Id. at 1270.

21. Id. 1270–71.

22. Id. at 1272–75.

23. Id. at 1273.

24. Id. at 1274–75.

25. Id. at 1275.

26. Id. at 1277.

27. Id.

28. See Henry and Adams, “The Eleventh Circuit Just Gave Us a Roadmap for Using AI to Interpret Insurance Policies—And Coverage Lawyers Should Pay Attention,” Bradley, It Pays to Be Covered (blog) (Mar. 16, 2026) (summarizing and analyzing Judge Newsom’s concurrence in Snell, and discussing “best practices” when using LLMs), https://www.itpaystobecovered.com/2026/03/the-eleventh-circuit-just-gave-us-a-roadmap-for-using-ai-to-interpret-insurance-policies-and-coverage-lawyers-should-pay-attention/?utm_source=mondaq&utm_medium=syndication&utm_content=sourceoriginal&utm_campaign=.

29. As Judge Newsom puts it, “LLMs can provide useful statistical predictions about how, in the main, ordinary people ordinarily use words and phrases in ordinary life.” Snell, 102 F.4th at 1226. See also id. at n.7 (discussing how these “statistical predictions” are arrived at). Judge Newsom, “unabashedly a plain-language guy,” adds that “[t]he 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.” Id. at 1222, 1226, 1228 n.11.

30. Id. at 1227.

31. Id. at 1228.

32. See Wikipedia, “Generative Pre-Trained Transformer,” https://en.wikipedia.org/wiki/Generative_pre-trained_transformer. It seems both ironic and quaint to cite Wikipedia for this general definition.

33. Snell, 102 F.4th at 1227–28.

34. Id. at 1228.

35. Id. While the typical legal researcher might not think to look in the Urban Dictionary, that online dictionary’s dataset may be a part of what an LLM “scrapes” in formulating a query response. Where else might one find the definition of “gorehound” and other neologisms (i.e., newly coined words and phrases, or giving existing words brand-new meaning)? See Gorehound, Urban Dictionary, https://www.urbandictionary.com/define.php?term=Gorehound. Three Tenth Circuit district court judges have cited the Urban Dictionary. See Keenan-Coniglio v. Cumbres & Toltec Scenic Operating Comm’n, 769 F.Supp. 3d 1216, 1239 n.11 (D.N.M. 2025) (collecting cases, and noting that in the matter at hand, “‘dictionaries le[ave] a little something to be desired’” (quoting Snell, 102 F.4th at 1223 (Newsom, J., concurring))). The court added that “instead of relying solely on the dictionary definition—the Court has looked to Wikipedia and Urban Dictionary” for the “contemporary meanings and usage” of “gringo” and “guero” in a race discrimination case, and observed that, “These resources have value, too.” Id.

36. Snell, 102 F.4th at 1228–29.

37. Id. at 1229 (quoting Scalia and Garner, “A Note on the Use of Dictionaries,” 16 Green Bag 2d 419, 422 (2013)).

38. Id.

39. See Henry and Adams, supra note 28.

40. Cf. Al-Hamim v. Star Hearthstone, LLC, 2024 COA 128, ¶¶ 2–4, 25–41 (discussing appropriate sanctions against pro se litigant who relied on an LLM to help write a brief that contained eight hallucinated case cites, citing in its discussion to Snell, 102 F.4th at 1226 n.7 (Newsom, J., concurring)). See also Avery, supra note 2 at 118–19 (LLMs’ “[b]lack-box systems frustrate review” because one cannot “see how inputs match to outputs.”).

41. Snell, 102 F.4th at 1230.

42. Id. at 1231.

43. Id.

44. Id.

45. Id.

46. Id. at 1229, 1232.

47. Id. at 1231–32. LLM data manipulation by interested parties may be more than a theoretical concern. One intrepid scientist made up a “fake” disease, “bixonimania,” and published online two fake papers involving fake bixonimania studies. Various chatbots picked up on the papers and treated bixonimania as a real disease. Equally worrying, the fake papers were later cited in peer-reviewed literature. Stokel-Walker, “Scientists Invented a Fake Disease. AI Told People It Was Real,” Nature (Apr. 7, 2026), https://www.nature.com/articles/d41586-026-01100-y. Evidently, some scientists, like some lawyers, are prone to cite “published” matter that is phony or that doesn’t exist.

48. Snell, 102 F.4th at 1232. Interestingly, LLM programs often incorporate algorithms that “shape” their responses to the same user over time to provide personalized, affirming answers in an effort to better “connect” with the human on the other end of the “conversation.” Such algorithms, by integrating demographics and past queries in formulating responses, may prioritize emotional connection over accuracy, thereby inadvertently changing their responses based on context over time rather than being strategically manipulated. See generally Zhang et al., “Personalize Before Retrieve: LLM-Based Personalized Query Expansion for User-Centric Retrieval,” arXiv.org (Oct. 10, 2025), https://arxiv.org/html/2510.08935v1#:~:text=Specifically%2C%20we%20identify%20two%20core,%2DYingyi/PBR%2Dcode.

49. Snell, 102 F.4th at 1232. Cf. Avery, supra note 2 at 120, 128–31 (asking whether “law produced or substantively shaped by artificial systems” can maintain the kind of legitimacy associated with law generated by human deliberative and collective participation, and if there is a danger of reducing personal responsibility on the part of those who decide or enforce the law).

50. Snell, 102 F.4th at 1232–34.

51. See Williams, “How AI Will Reshape Public Opinion,” Conspicuous Cognition (blog) (Mar. 3, 2026), https://www.conspicuouscognition.com/p/how-ai-will-reshape-public-opinion (“in liberal democracies where governments don’t exert significant censorship and control over LLMs, their most consequential impact on public opinion will involve technocratisation: shifting people’s beliefs towards expert opinion”).

52. Id. The same commentator adds: “When it comes to the effects of LLMs on public epistemics and our information environment, the most likely impact is simply that they greatly improve people’s access to expert-level information. This doesn’t mean there is nothing to worry about. Even when it comes to this technocratising tendency of LLMs, there are important grounds for concern and vigilance. For example, expert opinion is often biased and wrong, and there is a significant risk that the technocratising, epistemically converging features of LLMs might reduce epistemic diversity in broader society. . . . [T]he flaws of expert opinion, and the benefits of democratic diversity and debate—remain. However, we can only face up to these problems if we recognise LLMs for what they are: not a continuation of social media, but a powerful corrective to it.” Id. (emphasis added).

53. Id.

54. Sandgrund, “Who Can Write a Better Brief: Chat AI or a Recent Law School Graduate? Part 1,” 52 Colo. Law. 24 (July/Aug. 2023), https://cl.cobar.org/departments/who-can-write-a-better-brief-chat-ai-or-a-recent-law-school-graduate-part-1; Sandgrund, “Who Can Write a Better Brief: Chat AI or a Recent Law School Graduate? Part 2,” 52 Colo. Law. 16 (Sept. 2023), https://cl.cobar.org/departments/who-can-write-a-better-brief-chat-ai-or-a-recent-law-school-graduate-part-2.

55. Sandgrund, “Wilderness Frisbee,” 54 Colo. Law. 16 (Aug. 2025), https://cl.cobar.org/departments/wilderness-frisbee. Once this article is published, there will be two published sources of information about Wilderness Frisbee.

56. Interestingly, I ran a second experiment on my Wilderness Frisbee article to test the accuracy of an online tool that claims it can detect AI-generated writing. The program reported that there was less than a 5% chance that AI wrote this article—the tool’s lowest possible score. But, alas, it found that another article I wrote bore a 100% likelihood of having been produced with AI as my “creative partner [to] generate the heavy-lifting historical dialogue, compile the quote-stuffed arguments, and map out the debate. [The author] then edited it, added his signature local legal framing, and approved the wonderfully absurd illustrations.” But, wait, the article was 100% me—I swear! See Sandgrund, “What Can Benjamin Franklin and Genghis Kahn Teach Us About the Rule of Law?,” 55 Colo. Law. 24 (July 2026), https://cl.cobar.org/departments/what-can-benjamin-franklin-and-genghis-kahn-teach-us-about-the-rule-of-law.

57. See Diamandis and Kotler, The Future Is Faster Than You Think: How Converging Technologies Are Transforming Business, Industries, and Our Lives (Simon & Schuster 2020).

58. Judge Newsom’s thought experiment has gained some traction. See, e.g., Chaudhry v. Thorsen, No. 3:20-CV-50381, 2026 U.S. Dist. LEXIS 57070, at *54 n.26 (N.D.Ill. Mar. 18, 2026) (in defamation case involving statements suggesting teacher was “grooming” a student, the court rejected defendants’ reliance on “a generative artificial intelligence question and answer” intended to persuade the court that the term “groomer” can have an “innocent construction,” such as meaning the defendant was a “man on his wedding day” or was a “servant/official in a royal household,” or was referencing “animal care.” Rather, the court concluded: “The obvious and only connotation is to establish trust with someone, often a child or vulnerable person, in order to exploit or abuse them.” (citation modified)). The court concluded that defendant’s definitional efforts were not a “proper use of generative artificial intelligence to define a term as envisioned by Judge Newsom.” Id. See also Ross v. United States, 331 A.3d 220, 229 (D.D.C. 2025) (Deahl, J., dissenting) (in appeal reversing conviction for animal cruelty arising from leaving a dog in a car for over an hour on a hot day because, as a matter of law, such act, by itself, could not establish guilt beyond a reasonable doubt, the dissent argued that it was “common knowledge” such conduct “created ‘a plain and strong likelihood’ that [the] dog would be harmed,” citing to responses to various queries run on ChatGPT, and citing to Judge Newsom’s Snell concurrence). A separate concurrence in Ross acknowledged the dissent’s arguments and AI’s potential utility, but raised concerns about AI bias and reliability and other pitfalls of using AI in crafting judicial opinions, advocating a “cautious and proactive” approach to AI’s use). Id. at 236–37 (Howard, J., concurring). See also Avery, supra note 2 at 117, 125 (describing Judge Newsom’s concurrence as an experiment in querying LLMs to define key or contested terms). See also Simon, “Statistically Significant Judging: Mechanizing Originalism Through Corpus Linguistics and AI,” 23 Geo. J.L. & Pub. Pol’y 577, 590–91, 596–99 (Summer 2025), https://www.law.georgetown.edu/public-policy-journal/wp-content/uploads/sites/23/2026/06/23.2-Simon.pdf (discussing Judge Newsom’s thesis).

59. Frederick and Newsom, “Meaning, Understanding, and Contextual Textualism,” 135 Yale L.J. 2614 (May 2026), https://yalelawjournal.org/feature/meaning-understanding-and-contextual-textualism.

60. Id. at 2614.

61. Id. at 2614, 2626–29.

62. Id. at 2668–71. But because LLMs are starting to replace internet search functions, “prior definitions of commonly understood words are being systematically replaced with stripped down (operationalized) definitions by AI.” Urie, “Why AI Doesn’t Think, Cannot Reason, Isn’t Intelligent and Will Never Achieve Consciousness: A Bundle of Algorithms is Like a Rock in Certain Respects,” J. of Belligerent Pontification (blog) (July 2, 2026), https://roburie.substack.com/p/why-ai-doesnt-think-cannot-reason And “[t]his stripping down creates the sense of a consensus view on every topic that is incorrect. Linguistic diversity is being eliminated from the discourse.” Id.

63. Frederick and Newsom, supra note 59 at 2631–37.

64. Id. at 2650–54.

65. Id. at 2673.