Does AI Have Speech Rights? - Columbia Journalism Review

-->

About

Memberships

Donate

Advertise

Contact

About

Memberships

Donate

Advertise

Contact

Ask Anika

Image courtesy of Anika Collier Navaroli / Illustration by CJR

Sign up for the daily CJR newsletter .

Welcome back to Ask Anika, a column for addressing questions about the ethics, legal considerations, and best practices of adopting new technologies in journalism. Send me your questions at askanika@cjr.org.

Q: I’m considering using AI to write a draft of a story. But I’ve heard that there are some legal fights brewing over whether AI output could be protected under the First Amendment. Is that true, and if so, does that mean my work wouldn’t be protected?

A: As I wrote in my last column, each time a journalist uses generative AI in their work, they are opening themselves up to novel legal risks. Right now, without concrete legal precedent, the answer to your question mostly depends on who you ask. Since this is Ask Anika, here’s my answer.

More and more, lawyers for tech companies are arguing that the words spit out by generative AI’s math machines should be protected by the First Amendment. You may recall that a chatbot designed by Character.AI seemingly encouraged a teenager to commit suicide and that OpenAI’s ChatGPT allegedly aided and abetted a university mass shooting. In the defenses of these cases, the companies claimed speech rights for a “nonhuman speaker” (the chatbot). Tech lawyers have argued that chatbot output is analogous to video games or algorithmic content moderation on social media, which courts have repeatedly determined to be protected speech. Courts haven’t yet given us definitive guidance—the Character.AI suit was settled out of court, and the OpenAI case is still in the early stages.

So what else do we know about whether an LLM can actually produce “speech,” which would therefore be protected by the First Amendment? An LLM is a parrot with a penchant for probability, according to a seminal 2021 article from AI researchers Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell. Chatbots, they wrote, were created as a “system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning.” Although the overall effect can cause humans to mistake “output for meaningful text,” LLM output “is not grounded in communicative intent” and therefore not speech at all.

Sign up for CJR’s daily email

A cohort of legal practitioners and scholars use different reasoning but ultimately agree that the output of LLMs should not be considered speech protected by the First Amendment. Meetali Jain—the founder of the Tech Justice Law Project, which brought the Character.AI suit and others—argues that “there is a distinction to make between words and speech, and that it’s really only the latter that is deserving of First Amendment protection.” Stanford Law grads also addressed the question and proposed the notion of “speech certainty,” or “the basic idea that speech is only speech if the speaker knows what he said when he said it.” They argue that all outputs of machine-learning algorithms that rely on a mathematical technique called “gradient descent” lack such certainty and are thus unprotected non-speech. That criterion would apply to a range of technologies, from spam filters to chatbots.

Other countries are further along than the United States in sorting out the question of AI text as speech. In June, a court in Germany (where neither the First Amendment nor its absolutist ideas apply) ruled that Google was liable for false information provided by its AI Overview feature in search results. The court wrote that the text generated by AI was “not the expression of an acquired conviction of the persons expressing it, but the result of an algorithm.”

You do not have to stretch your imagination (or interrogate your gracious data-science colleagues about obscure machine-learning concepts, like I did) to fully understand the implications of these arguments for journalists. If machines that mimic language are not meaningfully speaking within the context of the First Amendment, then the output generated using artificial intelligence to, say, write a technology advice column does not enjoy the same constitutional protection as this draft that I am writing in Google Docs. That leaves journalists using generative AI in their work in uncharted legal territory.

The risks may include lack of protection against the forced disclosure of sources, confidential information, and even the prompts fed to a chatbot. And nothing prohibits a government from subpoenaing the text of an unpublished story that has been input into an LLM. If LLM output were considered protected speech, that could be beneficial for a journalist, since the previous scenarios would be treated under...