AI and the Future of Free Speech

Artificial intelligence is reshaping the free speech landscape faster than any technology since the internet itself. The transformation is happening across multiple dimensions simultaneously — and the legal frameworks designed for human speakers are proving inadequate to address any of them.

AI as content creator: Large language models can generate text, images, audio, and video that is difficult or impossible to distinguish from human-created content. This creates new categories of speech — AI-generated political commentary, synthetic news articles, deepfake video and audio — for which existing legal frameworks were not designed. Questions left open include: Who is legally responsible for AI-generated defamation — the model developer, the deployer, the user who prompted it? Can AI-generated content infringe copyright? Andersen v. Stability AI and Kadrey v. Meta, both filed in 2023, challenged whether training AI systems on copyrighted works without licensing constitutes infringement — with implications for the entire generative AI industry. State legislatures have begun passing deepfake laws targeting election-related synthetic media; by 2025, over 20 states had enacted some form of deepfake regulation, with varying scope and First Amendment vulnerability.

AI as content moderator: Platforms rely on AI systems to moderate billions of posts daily — a volume that makes human review impossible at scale. These systems operate with imprecision, systematic biases, and opacity that makes accountability difficult. AI moderation errors are not random: they systematically over-flag certain languages, accents, and topics, creating disparate impacts on minority communities whose speech patterns differ from the dominant training data. When an AI system removes a post, the speaker typically receives a generic notice with no meaningful explanation and limited ability to appeal. The First Amendment governs government censorship, not private platform decisions, but the scale of AI moderation means that private systems now determine the practical speech environment for billions of people in ways that raise serious democratic concerns even absent constitutional application.

AI as speech suppressor: AI content filters embedded in search engines, recommendation algorithms, and platform moderation can effectively silence speakers without any human decision. When an algorithm reduces the distribution of content classified as harmful, the speaker may not know their reach has been limited, may have no recourse, and may never understand why. "Shadow banning" — limiting a user's visibility without notifying them — illustrates the problem. Moody v. NetChoice (2024) addressed whether platforms' editorial discretion in algorithmic curation receives First Amendment protection; the Court's fractured decision left the fundamental question open while signaling that content-neutral algorithmic systems are treated differently from viewpoint-based editorial decisions.

AI as speech actor: Questions once theoretical are becoming practical. Do AI systems themselves have anything analogous to First Amendment rights? Can AI be defamed? Can the government compel AI systems to express — or suppress — particular views? The compelled speech doctrine, rooted in West Virginia v. Barnette (1943) and extended in cases like Wooley v. Maynard (1977) and Hurley v. Irish-American Gay Group (1995), protects speakers from being compelled to convey government-approved messages. As AI systems become primary information intermediaries — answering billions of queries daily — government attempts to require particular outputs raise serious compelled speech questions. Conversely, when a state requires AI systems to include health warnings on certain content, or to label AI-generated material, does that compel speech in a constitutionally problematic way?

The EU AI Act (2024), the most comprehensive AI regulation enacted anywhere, takes a risk-based approach that has significant speech implications. "High-risk" AI systems, including those used in critical infrastructure and some content moderation contexts, face extensive transparency and human oversight requirements. The Act's provisions on "general purpose AI models" — the large foundation models like GPT-4 and Claude — require disclosure of training data and compliance with EU copyright law, with implications for how AI developers can build systems. The United States has taken a different path: the Biden administration's Executive Order on AI (2023) established risk evaluation frameworks through NIST but avoided binding regulation, and the Trump administration reversed much of this framework in 2025.

The questions being decided now — in AI labs, in early litigation, in regulatory proceedings — will set the frame for the next generation of free speech law. The decisions about what AI systems can refuse to generate, what they are required to disclose, how they moderate content, and whether they have any expressive rights of their own will shape the speech environment for billions of people in ways that make today's platform content moderation debates look narrow. The First Amendment doctrine that protected individual speakers from government suppression is being stress-tested by technology that disaggregates speaker, platform, moderator, and censor in ways that existing doctrine cannot easily address.