Week 6 - What is Non-Human Speech?

Welcome back, Free Speakers!

In my last post, we gave some evidence for Week 4’s dissent, and I talked a little bit about creating a ‘middle ground’ for AI outputs. But what would that really look like?

Money Talks

There’s actually a pretty well-documented history for First Amendment rights given to non-traditionally human entities. In these cases, the First Amendment may apply to its fullest extent, or to a more limited capacity. You may be familiar with the saying, “corporations are people” (and if you’re not, you’re in for a doozy with this post!). The saying was popularized by the landmark 5-4 decision in the famed case Citizens United v. FEC, which solidified First Amendment rights for corporate speakers. Citizens United was a non-profit advocacy organization that created a political movie on Hilary Clinton, and wanted to spend some of the corporation’s money to do so.^[1] There were existing restrictions on the amount of money a corporation could spend on communications, but ultimately, the Court sided with Citizens United. In doing so, the Court bestowed core First Amendment protections upon corporate speech, giving corporations full speech rights, as if they were a regular ol’ human. 

But, this actually wasn’t the first case where a corporation was afforded rights.^[2] Bank of the United States v. Deveaux was the first case that gave corporations the right to sue—under the pretense that the action was one taken by the employees of the corporation (who were, of course, citizens).^[3] Following that logic, corporations were given First Amendment protections in First Nat’l Bank of Boston v. Bellotti, where a bank sought to influence the results of a ballot initiative using its profits.^[4] The Court afforded a bank First Amendment protections “based not only on the role of the First Amendment in fostering individual self-expression but also on its role in affording the public access to discussion, debate, and the dissemination of information and ideas.”^[5] Essentially, the Court sought to protect the ideas we talked about in Week 2. Both these cases preceded Citizens United and laid the jurisprudential groundwork for Justice Kennedy’s opinion in the divisive case.

The lineage of corporate First Amendment rights outlines what speech rights may look like when it comes from a ‘speaker’ that’s a bit detached from a human. In both Citizens United and Bellotti, the corporations used a non-pure method of communication, and still got away with having their speech protected. The Court reasoned that money used by the corporation can enable speech, and limiting the expenditures “necessarily reduces the quantity of expression by restricting the number of issues discussed, the depth of their exploration, and the size of the audience reached.”^[6]

Who’s Expressing What?

So, money is speech because of the ways in which it facilitates expression and the dissemination of ideas in our beloved marketplace. The natural question that follows is, “can this be applied to LLMs?” 

LLMs, too, facilitate expression. The corporations that create LLMs arguably use them as means to get their ideas out there. If a writer has a word on the tip of their tongue, and they can think of a few words that are similar to—but not really—the word they’re looking for, an LLM (that they could create) can help them search for that word. The model allows the writer to better express themselves by increasing the precision of their diction. We noted in Week 5 that every other First Amendment case which hadn’t questioned expressive intent had a human doing the expressing. In Moody v. Netchoice, Justice Barrett recognized the nascency and complication in the use of AI when expressing oneself. In the case, where a corporation’s content-moderation was considered expression, the Justice wrote, “technology may attenuate the connection between content-moderation actions (e.g., removing posts) and human beings’ constitutionally protected right to ‘decide for [themselves] the ideas and beliefs deserving of expression, consideration, and adherence.’”^[7] In other words, the Justice questions how we should know if a model is truly expressing our desires as users. If the writer in our previous hypothetical doesn’t get the word they were looking for after a few inquiries from the model, is it still self-expression? 

Using current legal frameworks, expression should be tied to the corporation.^[8] Esteemed First Amendment scholar, Eugene Volokh, reasoned with colleagues that, “the choices companies and their employees make about what sources to train on and what results to modify using human feedback directly or indirectly influence the output of their AIs.”^[9] When creating any LLM model, there is a deliberate decision made by the engineers in deciding what raw data is given to the model. If a software engineer wanted to create a model that was really good at roasting their friend using inside jokes, they might give the model messages between them and the friend (deliberately excluding any messages between, say, the engineer and their father, so as not to confuse the model). This is the same reasoning that the Court uses in Moody to attach expressive intent to the content moderation (companies deliberately making the choice to exclude certain content). But, there can’t be intention in an output that the engineer didn’t want. Models can be jailbroken—which is just the process of removing a limitation or guardrail from a piece of technology—and when they are, the model outputs information the company wouldn’t willingly disseminate.^[10] When this happens, it’s not the company trying to secretly give erudite users access to illicit information. It’s the user who’s looking for this response. The user makes deliberate attempts to try and force specific outputs that they desire, against the wishes of the corporation that created the model. So, that kinda looks like the user expressing their own beliefs… right?

Eh… maybe not. Volokh et al. argue that “[r]egardless of whether any speaker interests are involved in an AI program’s output, readers can gain at least as much from what the program communicates as they do from commercial advertising, corporate speech, and speech by foreign propagandists—three kinds of speech that have been held to be protected in large part because of listener interests.”^[11] Because of the listeners’ rights we talked about in Week 3, Volokh claims the First Amendment needn’t worry about any expressive intent that the user has. The use of LLMs to find information is analogous to that of Google, Bing, or other search engines—it’s just that LLMs have a little more variance in what they’re able to say.^[12] To further support this point, the Court reasoned in Citizen’s United that “[i]t is irrelevant for purposes of the First Amendment that corporate funds may ‘have little or no correlation to the public’s support for the corporation’s political ideas’… the First Amendment protects [] speech, even if it was enabled by economic transactions with persons or entities who disagree with the speaker’s ideas.”^[13] The Court is basically saying that a corporation can do what it pleases with my funds if I support it economically—even though money is speech. Meaning, that corporation can use that money to support a cause that I don’t agree with. This seems analogous to the case at hand, affirming Volokh’s claim. If I input one thing into a model, the model can really output whatever it wants (under generally applicable laws like fraud and defamation, of course), because it’s up to the company to decide what it wants the model to express. 

But, doesn’t the variability in AI outputs make LLMs distinct in some way from other non-human speakers? Well, yes, it would. And we’ll get into that next week!

References

[1] Citizens United v. FEC, 558 U.S. 310 (2010).

[2] See Adam Winkler, We the Corporations: How American Businesses Won Their Civil Rights, Liveright Publishing Corporation, a Division of W.W. Norton & Company (February 27, 2018). 

[3] Bank of the United States v. Deveaux, 9 U.S. 61 (1809).

[4] First Nat’l Bank of Boston v. Bellotti, 435 U.S. 765 (1978).

[5] Id.

[6] Citizens United, supra note 1 (quoting Buckley v. Valeo, 424 U.S. 1 (1976)).

[7] Moody v. NetChoice, LLC, 603 U.S. ___ (2024) (Barrett, J., concurring, quoting Turner Broadcasting System, Inc. v. FCC, 512 U. S. 622, 641 (1994))

[8] Eugene Volokh, Mark A. Lemley, and Peter Henderson, Freedom of Speech and AI Output, Journal of Free Speech Law (August 3, 2023). [https://ssrn.com/abstract=4531003]

[9] Id.

[10] See Vladyslav Podoliako, Jailbreaking Claude, Vladyslav Podoliako and Belkins Inc. (April 17, 2026). https://www.vladsnewsletter.com/p/jailbreaking-claude; Alexey Guzey, A Two sentence Jailbreak for GPT-4 and Claude & Why Nobody Knows How to Fix It, Alexey Guzey, https://guzey.com/ai/two-sentence-universal-jailbreak/ (April 17, 2026); see also Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks, Anthropic (April 17, 2026) https://www.anthropic.com/research/next-generation-constitutional-classifiers

[11] Volokh et. al., supra note 7. 

[12] Id.

[13] Citizens United, supra note 1.