Reps. Jay Obernolte and Lori Trahan recently released a discussion draft of an AI bill known as the Great American Artificial Intelligence Act (GAAIA). At 269 pages, the draft covers topics including energy, labor, innovation, education, cybersecurity, preemption, and more. My colleagues have already covered many of these aspects, but I want to provide a closer analysis of the bill’s free speech implications.
Such a focus on speech is especially important because, late last week, the Trump administration effectively forced major AI models to shut down without any public explanation. The administration imposed export controls on Anthropic’s newest AI models, Fable 5 and Mythos 5, resulting in Anthropic blocking access to those models for all users.
While GAAIA does not create major new challenges for AI-powered speech, it misses an opportunity to meaningfully protect expression. In light of the Trump administration’s recent action against Anthropic’s newest AI models, GAAIA should do more to safeguard speech.
Lack of Direct Protections for Speech
GAAIA dedicates Section 141 to free expression. Unfortunately, this section does not provide substantial protection for free expression; instead, it directs the Secretary of Commerce to study various free-expression issues.
Primarily, the study would focus on jawboning: how the government interacts with private AI platforms, how such interactions may affect protected expression, and which policy options could curtail such jawboning. It is useful for Congress to consider how government power may be used to censor or compel speech, but a study falls short of addressing potential jawboning, especially given the significant research and congressional interest already devoted to the problem.
A transparency regime that required government officials to publicly disclose relevant communications, with necessary protections for classified and private information, would be an obvious next step to which the bill could have committed. Within a week of this draft being released, Senators Cruz and Wyden introduced the JAWBONE Act to tackle just this problem.
Perhaps more importantly, this section does not address the other ways government rules and regulations around AI can harm free expression. As we saw in the dust-up over Fable 5 and Mythos 5, the government’s creative use of export controls can have a far greater impact on free speech and access to information than government pressure or coercion against particular kinds of speech.
Standards and Preemption
Beyond these issues, GAAIA also has several other policies with free expression implications. Two worth closer consideration are the impact of the independent standards system and preemption.
As my colleague Juan Londoño has covered in significant detail, GAAIA creates a pseudo-independent testing regime. Large frontier model developers are required to create risk management frameworks and transparency reports to head off catastrophic risks. These companies must be regularly reviewed and audited by independent verification organizations (IVOs) licensed by the Center for AI Standards and Innovation (CAISI). Companies face significant civil penalties of up to $1 million per violation per day if they knowingly make inaccurate statements or convey false impressions regarding risk management disclosures.
Similar penalties may be imposed if a developer fails to comply with IVO-related requirements, and IVOs must refer matters to the Attorney General when they determine that noncompliance poses an imminent catastrophic risk. CAISI is also tasked with crafting various “voluntary” standards and best practices around detecting, labeling, or otherwise tracking the provenance of AI content. While these standards are voluntary, there is always the concern that government-determined best practices may effectively become mandatory when combined with procurement policies, liability norms, or expectations from other private actors.
The potential free speech concern is that the combination of “voluntary” standards, required disclosures, high compliance costs, and incentives for IVOs to audit aggressively all push developers toward compliance choices that narrow the range of lawful speech deemed safe. The bill could mitigate this trend toward collateral censorship by requiring standards and compliance obligations to be content-neutral and avoid suppressing content beyond limited, enumerated catastrophic harms.
On the other hand, one potential speech benefit of the bill is its preemption of state AI laws. Creating a single set of light-touch federal rules can prevent a patchwork of state laws that would inhibit competition and reduce the diversity of AI tools. The bill also includes significant open-source protections, again promoting a wider range of viewpoints within AI models.
However, the major question is whether the bill does enough to prevent the patchwork. The preemption lasts only three years and is fairly narrow, limited to AI development and not addressing the many applications of AI. Whether this preemption will yield significant benefits in competition and expression is therefore unclear.
Protecting Speech
Given this mixed or unclear impact on speech, what could this or another bill do to better address free expression concerns around AI? As noted above, there are specific anti-jawboning policies, such as transparency, that could be included rather than just a study.
GAAIA should also scrutinize and restrict the whole range of executive authorities that could be weaponized against AI-generated speech and the speech rights of AI developers. When a decision by one government official can silence entire models, more guardrails are necessary to prevent abuse. Export controls, tariffs, and other declarations of various emergencies by the executive branch not only present major separation-of-powers issues, but they also allow speech rights to be threatened by the whims of the executive.
GAAIA also needs a more meaningful preemption to prevent a patchwork of state laws that will hamstring competition and innovation, and thus speech. Failure will likely result in a Sacramento or Brussels effect — i.e., the ability of large states to adopt burdensome regulations that companies end up adopting as national or global standards rather than deal with many different standards, coming to dominate and restrict AI.
Policymakers should pursue an innovation-first approach that embraces soft law developments. GAAIA does this in part by creating independent verification organizations and best practices, but these depend in large part on government involvement rather than truly organic development. Policymakers should also rely as much as possible on the common law application of existing law to new technologies. Embracing such a path will allow the AI industry to flourish, and speech right along with it.
Conclusion
AI is often the focus of debate because of its powerful and life-altering potential, from cybersecurity to transportation. But with the effective shutdown of Fable 5 and Mythos 5, we must also recognize that AI tools are speech. They represent the views and choices of their developers and empower users with countless expressive choices and opportunities. While the Great American Artificial Intelligence Act focuses on innovation, development, and risk regulation, it should place greater emphasis on protecting the expressive elements of AI, including from arbitrary government decisions such as the Trump administration’s exercise of export controls.
