Week 5 - Are LLM Outputs Pure Speech?

Welcome back, Free Speakers!

In my post last week, I disagreed with the Court’s decision to analyze expressive intent in LLM outputs. We left off by saying that words are presumed to be expressive by First Amendment jurisprudence. Let’s look at some evidence of our conclusion.

To Analyze or Not to Analyze?

In Character.AI’s motion for interlocutory appeal, the company argued that a number of circuit courts have held in differing ways that words need not undergo expressive analysis.^[1] In Otto v. City of Boca Raton, “the Eleventh Circuit rejected the argument that speech-based therapy was ‘non-expressive conduct’ that did ‘not implicate the First Amendment’ because it ‘consists—entirely—of words.’” In Anderson v. City of Hermosa Beach, “the Ninth Circuit [held] that ‘words’ are a ‘form[] of pure expression,’” and that they receive “‘full First Amendment protection.”^[2] The Court also found that it was unnecessary to conduct any analysis to figure out if the words have enough “elements of communication.”^[3] Other Courts have held similar principles.^[4] Most notably, the case of Walters v. OpenAI, L.L.C. came before the Georgia Supreme Court, when Walters filed for defamation against OpenAI. The words which triggered the defamation suit were generated by OpenAI’s LLM. Defamation cases require First Amendment scrutiny—similar to the scrutiny undergone in Garcia—but when the Court analyzed the outputs under the First Amendment, it did not use the expressive conduct test. The Court in Walters made the decision not to analyze the expressive intent behind the LLM’s output. That decision is relevant to our case, because the facts of the cases that ultimately lead both of the courts to make their respective decisions are largely the same. Both cases implicate the First Amendment. Both cases analyze LLM outputs under the First Amendment. In Garcia, the Court asks the defendants if their purported speech is really speech, but in Walters, the Court doesn’t concern itself with that question.

“But wait,” you may begin to ask, “if the cases are functionally the same, how do they arrive at two different conclusions?” Well, the two courts just disagreed with one another. The expressive conduct test is triggered in Garcia, because the Court believed it was ambiguous whether the outputs were speech or not.^[5] The expressive conduct test is not triggered in Walters for the same reasoning I gave last week—it implicitly presumes the words to be expressive.^[6] Additionally, it’s important to note that Walters is not binding (i.e. courts outside its jurisdiction don’t have to follow their decision). Walters was a case in the Georgia Supreme Court, while Garcia is in the Florida District Court. That said, Walters and the other cited cases show that when faced with an analogous issue, Courts have considered and declined using the expressive conduct test. And I tend to agree with those Courts, opposed to the one at bar.

A Middle Ground for Rights

We now have some citable law to back up last week’s wild dissent. Now, I know I disagreed with the Garcia Court’s reasoning pretty hard last week, but maybe the Court was onto something. We’re exploring unmarked lands with this ‘AI outputs is free speech’ talk. In other cases where pure speech isn’t questioned for its expressive intent, humans were the ones expressing themselves. By not doing the expressive conduct test, Courts inadvertently allow LLM outputs to have the same protections as regular human speech. It makes sense why courts don’t want to do that—especially given the sweeping implications.^[7] If Courts find that LLM outputs have no First Amendment rights (i.e. if Courts find LLM outputs fall under an unprotected speech category), then it can be wholly regulated and limited by the government.^[8] But what if we didn’t have to choose between ‘AI has the same rights as humans’ and ‘LLM outputs can be regulated as much as the government pleases’? Well, lucky for us, speech categories exist that allow for this First Amendment middle ground. This category of speech can be referred to as ‘less protected speech’ or ‘partially protected speech.’ And—you guessed it—we’ll talk more about it next week. 

I’ll see you Free Speakers then!

References

[1] Garcia v. Character Technologies, Inc., 6:24-cv-01903, (M.D. Fla. Jun 18, 2025) ECF No. 130 at 12-13

[2] As cited in Id. at 12

[3] Id. at 12, citing Spence v. Washington, 418 U.S. 405, 409 (1974), quotations omitted. 

[4] See Bery v. City of New York, 97 F.3d 689, 695 (2d Cir. 1996) (where the Second Circuit refused to evaluate if inherently expressive communications were “lacking in communicative concepts or ideas”);  Longoria ex rel. M.L. v. San Benito Consol. Indep. Sch. Dist., No. 1:17-cv-160, 2018 WL 6288142, at *7 n.2 (S.D. Tex. July 31, 2018), aff’d, 942 F.3d 258 (5th Cir. 2019) (cited Doc. 98 at 5) (“Because [student] engaged in pure speech, the Court need not apply the symbolic speech test … to determine whether [the] expression showed intent to convey a particularized message.”) (as cited in Garcia at 13, supra note 1); see also Bernstein v. U.S. Dep’t of State, 922 F. Supp. 1426, 1434–35 (where the Northern District Court of California found it needn’t evaluate the expressiveness of code, because “‘[l]anguage is by definition speech’” (quoting Yniguez v. Arizonans for Off. Eng., 69 F.3d 920, 935 (9th Cir. 1995), vacated on other grounds, 520 U.S. 43 (1997))) (as cited in Garcia at 12-13, supra note 1).

[5] Garcia v. Character Technologies, Inc., 6:24-cv-01903, (M.D. Fla. May 21, 2025) ECF No. 115 (holding “the purpose of the expressive conduct test is to determine whether conduct is sufficiently similar to speech so as to warrant First Amendment protections.”) 

[6] See Walters v. OpenAI, L.L.C., No. 23-A-04860-2 (Ga. Sup. Ct. May. 19, 2025) (the Court uses language such as: “a reasonable reader in Riehl’s position could not have concluded that the challenged ChatGPT output communicated actual facts,” illustrating the LLM in fact communicated using language via its outputs (quotes and citation omitted)).

[7] See Jodi Duckett, AI, Free Speech, and the Future of Democracy: Navigating Truth in the Age of Synthetic Media, Leigh University College of Arts and Sciences (November 19, 2024) https://cas.lehigh.edu/articles/ai-free-speech-future-democracy

[8] Chaplinsky v. New Hampshire, 315 U.S. 568 (1942)