VIDEO | AUDIO | RECAP EN / ES / FR | ARCHIVE | PERMALINK
Location: Washington, DC / Online
Speakers: Megan Iorio - Senior Counsel, Electronic Privacy Information Center (EPIC); Corbin Barthold - Senior Litigation Fellow, NetChoice; Kate Ruane - Director, Free Expression, Center for Democracy & Technology
Moderator: Cristiano Lima-Strong - Senior AI & Government Reporter, Bloomberg Government
Opening
Tim Lordan - Executive Director, Congressional Internet Caucus Academy - opened the discussion by framing artificial intelligence as a new challenge for First Amendment law. He introduced the panel and noted that generative AI differs from earlier technologies because machines can now produce language that resembles human speech.
Lima-Strong said policymakers are increasingly confronting AI risks at the same time that courts and legal scholars are considering whether and how First Amendment protections apply to AI. He highlighted parallels with longstanding disputes over social media regulation and began with the fundamental question of why speech generated by machines should implicate the First Amendment at all.
Whose Speech Does AI Represent?
Iorio argued that the central First Amendment concern is not whether an AI system itself possesses rights, but whether government restrictions on AI interfere with people’s ability to receive information and viewpoints. She emphasized that some activities involved in developing AI systems may be expressive while others are not, making it necessary to examine precisely what a government intervention regulates.
Barthold similarly described an LLM as a tool rather than a person, comparing it to technologies such as cameras and other mechanisms used to communicate. He emphasized listeners’ rights to receive information and argued that AI developers also make expressive choices through training data, reinforcement learning and decisions about how systems present information.
Ruane agreed that AI systems themselves do not possess First Amendment rights but said users and developers can have significant speech interests in their creation and use. At the same time, she stressed that First Amendment protection is not absolute. Questions of harmful outputs, liability, civil rights law and increasingly autonomous AI agents therefore require determining which human actors bear responsibility.
Responsibility for AI-Generated Speech
The panel explored the difficulty of identifying a human speaker behind generative AI output. Iorio cited litigation involving Character.AI to illustrate the tension between listener rights and the absence of a clearly identifiable human speaker responsible for a chatbot’s individual responses. She argued that an important question is whether a human made decisions that resulted in the expression being regulated.
Barthold agreed that LLMs lack human intent, creating complications for doctrines such as true threats, incitement and conspiracy, which normally involve an intent requirement. He suggested that some form of principal-agent relationship may ultimately be needed to establish responsibility for AI-generated output while cautioning against imposing substantially stricter speech standards on AI than would apply to people.
Defamation, Harm and Liability
Turning to inaccurate or defamatory chatbot outputs, Ruane noted that generative systems can produce unpredictable results even when developers establish safeguards. She argued that liability questions should therefore examine the chain of development and deployment: what developers controlled, what safeguards they implemented and whether they took reasonable steps to prevent foreseeable harmful outputs.
Barthold emphasized that users can also acquire responsibility when they take information generated privately by an AI system and publish or repeat it. Ruane added that because a chatbot itself cannot be sued, courts will ultimately have to determine which person or organization in the development and deployment chain can be held responsible. Iorio characterized much of this discussion as a question of tort liability, centered on what humans actually controlled in training, fine-tuning and safety design.
Is AI Output Expressive?
The panel then considered whether generative AI output is itself expression. Iorio argued that this question does not necessarily determine whether the First Amendment applies. Even if chatbot output were considered non-expressive, a government rule requiring or prohibiting particular viewpoints could interfere with the expressive choices of developers. She connected this issue to compelled-speech and editorial-discretion precedents, including Moody v. NetChoice.
Ruane focused on the many editorial decisions made during AI development. Developers choose training data, remove certain material, establish safety policies and apply input and output filters. Users then influence the resulting expression through their prompts. She argued that First Amendment analysis should therefore examine this entire process rather than asking only whether an individual chatbot response represents the expression of a particular person.
Barthold took a broader position, arguing that the presence of ideas and even relatively minimal expressive choices can be sufficient to implicate First Amendment protections. He cautioned against creating a new requirement that expression must always be traceable to a specific human decision.
An Unsettled Area of Law
The speakers agreed that relatively little AI-specific First Amendment doctrine has been settled. Barthold noted that many relevant cases rely on analogies involving libraries, newspapers, parades and other traditional forms of expression rather than judicial decisions directly addressing generative AI. Settlements in lawsuits involving harmful AI interactions have likewise provided little precedent.
Autonomous Agents: Speech or Conduct?
The discussion shifted from chatbot speech to autonomous AI agents capable of taking actions. Iorio argued that activities such as unauthorized hacking are principally questions of conduct, cybersecurity and liability rather than novel First Amendment questions. Existing laws already regulate hacking even though code and language can be involved in carrying it out.
Barthold agreed that traditional distinctions between speech and conduct provide the most useful starting point. Breaking into a computer system remains conduct when performed through AI, he said, although increasingly autonomous agents will create difficult edge cases as systems become less directly connected to individual human actors.
Government Pressure and “Jawboning”
The panel also examined government pressure on AI companies. Ruane explained “jawboning” through existing First Amendment precedents involving officials using regulatory power or threatened enforcement to induce intermediaries to suppress speech. She said similar concerns could arise if access to government contracts or other benefits were conditioned on AI companies changing the viewpoints expressed by their models.
Barthold highlighted the dependence of AI companies on government decisions involving regulation, infrastructure and other resources. He argued that transparency is particularly important when government and AI companies communicate privately, so that legitimate cybersecurity oversight does not become a mechanism for influencing lawful speech.
Guidance for AI Policymakers
In closing, Iorio pointed to privacy, data protection and product liability as areas where policymakers can address chatbot harms while potentially avoiding some of the most difficult First Amendment questions. She highlighted concerns about AI systems collecting extensive personal information and using it to influence users.
Barthold called for clearer rules defining when developers or other principals are responsible for the actions of AI agents, while maintaining established First Amendment principles concerning government suppression of ideas. Ruane urged policymakers to concentrate on concrete and identifiable problems, including cybersecurity safeguards and incentives for safer system development, rather than assuming that every novel AI incident requires an entirely new legal framework.
RESOURCES
AI, Speech, and the First Amendment: Where Are the Constitutional Lines? — Congressional Internet Caucus Academy event page
Whose Speech Is It Anyway? The Constitutional Contours of Chatbot Regulation — the CDT paper cited by Megan Iorio
People-First Chatbot Bill — EPIC model bill with Consumer Federation of America and Fairplay
JAWBONE Act — Senate anti-jawboning bill discussed by Kate Ruane
Moody v. NetChoice — includes the Barrett concurrence
NRA v. Vullo — most recent Supreme Court jawboning case
Lamont v. Postmaster General — listener’s right to receive information
Bantam Books v. Sullivan — Rhode Island commission’s obscene-books letters
Hurley v. Irish-American GLIB — Boston St. Patrick’s Day parade case
On the Dangers of Stochastic Parrots — 2021 paper by Emily Bender, Timnit Gebru et al.