First Amendment Questions for AI Transparency Laws

A bipartisan group of lawmakers in the U.S. House of Representatives recently introduced the AI Foundation Model Transparency Act, which would direct the Federal Trade Commission to set transparency requirements for the data used to train high-impact foundation models. Developers would need to provide information about where training data comes from, how models are trained, and whether user data is collected during use. The bill joins a growing roster of artificial intelligence (AI) transparency measures at the state and federal level, including California’s Assembly Bill (AB) No. 2013, which requires developers to publish high-level summaries of their training data; California’s Senate Bill (SB) 53, which requires frontier AI developers to publish safety frameworks and make public disclosures about risk assessment and mitigation measures; and New York’s RAISE Act, which imposes requirements similar to SB 53.

The basic idea behind AI transparency laws is straightforward: As AI plays a larger role in public life, the public should have access to basic information about how these systems are built and the risks they pose. But laws requiring companies to publish information about their AI systems can face First Amendment scrutiny. Legislators drafting disclosure requirements will need to do so with an eye toward how courts may evaluate those laws.

Under U.S. law, when the government compels a company to publish information about its products or services, it regulates the company’s speech. That means transparency laws trigger First Amendment scrutiny. If companies challenge these laws in court, the outcome can turn on what level of scrutiny a court applies—a question currently being litigated in a challenge to California’s AB 2013. How courts answer this question will matter far beyond AB 2013. While AI transparency laws remain viable, First Amendment doctrine is becoming less predictable, and drafting choices matter more than many policymakers assume. The questions that the U.S. Court of Appeals for the Ninth Circuit now confronts will not only affect AB 2013 but also may shape how courts evaluate other AI transparency laws going forward.

xAI’s Challenge to AB 2013

AB 2013 is an early test for how courts may evaluate training data transparency mandates. The statute requires AI developers who provide models in California to post high-level disclosures on their websites, including:

• General sources and characteristics of training data.
• How datasets relate to the system’s intended purpose.
• Approximate size of the data (expressed in ranges or estimates).
• Whether the data includes copyrighted or licensed material.
• Whether personal or aggregate consumer information is involved.
• Whether synthetic data are used.

As the law took effect in the new year, Elon Musk’s AI company, xAI, sued in the Northern District of California, claiming that AB 2013 was unconstitutional and seeking to block the law with a motion for a preliminary injunction. On March 4, U.S. District Judge Jesus Bernal of the Central District of California denied xAI’s motion. Even though California won at this stage, the case continues on appeal at the Ninth Circuit.

The appeal may address how courts evaluate disclosure requirements such as AB 2013 as compelled speech under the First Amendment. AB 2013 remains in effect for now, but the state’s victory was tempered by Bernal’s caution that xAI could have “a distinct possibility of prevailing on the merits” of its First Amendment challenge as the case develops, even if the Ninth Circuit affirms the denial of the preliminary injunction. Bernal also subjected AB 2013 to a more searching review than courts have traditionally applied to regulatory disclosure laws.

The Three Standards for Compelled Speech

To understand what happened in the AB 2013 case, it helps to begin with the standards courts use to evaluate laws that principally compel speech, including transparency laws that require companies to disclose information.

• Strict scrutiny is the default for content-based speech regulation. Laws that require a speaker to say something specific are usually evaluated as content-based regulations because they dictate the content of the speech. Viewpoint-based regulations—favoring one view over another—are an especially disfavored subset of content-based regulation. Under strict scrutiny, the government must show the law is “narrowly tailored to serve compelling state interests.” In practice, laws rarely survive it.
• Intermediate scrutiny applies to regulations on “commercial speech.” The Supreme Court described the test for intermediate scrutiny as applied to commercial speech in the Central Hudson case: The court asks whether the “government’s interest is substantial,” whether “the regulation directly advances” that interest, and “whether it is not more extensive than is necessary to serve that interest.” While the Central Hudson standard is easier than strict scrutiny, and laws can survive it, it’s still a form of “heightened” scrutiny.
• Zauderer review is the most permissive. It also applies to commercial speech, though some disagree about its scope. The Zauderer test permits compelled disclosures that are “purely factual and uncontroversial” and “reasonably related” to a substantial government interest, and not “unjustified or unduly burdensome.” In practice, when courts determine that Zauderer applies, most disclosure laws survive^.

How a court characterizes a disclosure requirement—such as whether it qualifies as commercial speech—can determine whether it survives a legal challenge. Two distinct questions therefore matter for transparency laws. First, when does a compelled disclosure count as commercial speech? Second, if it does, which lower standard applies—Central Hudson or Zauderer? For policymakers, the practical point is that the same disclosure requirement can face very different odds in court depending on how a judge answers those two questions.

Narrowing What Commercial Speech Qualifies for Deferential Review

Until relatively recently, courts usually treated factual disclosure requirements imposed on businesses and professionals as compelled commercial speech and reviewed them under Zauderer’s relatively permissive standard. However, several recent developments have unsettled that understanding. That matters because it can make disclosure laws harder to defend.

First, in 2018, the Supreme Court decided National Institute of Family and Life Advocates (NIFLA) v. Becerra, striking down a California law that required crisis pregnancy centers to make disclosures about their services. The Supreme Court rejected the argument that Zauderer applied to the disclosures, despite their regulation of speech by professionals, as they pertained to “abortion, anything but an ‘uncontroversial’ topic.” NIFLA breathed new life into Zauderer’s requirement that compelled disclosures be “uncontroversial.”

After NIFLA, companies challenging disclosure laws can argue that even required factual disclosures shouldn’t get Zauderer’s more deferential treatment when disclosures concern controversial topics. In 2024, the Ninth Circuit blocked a California law requiring social media companies to produce reports on their content moderation policies based on state-specified categories. The lower court had allowed the law under Zauderer, but the appeals panel decided that the law regulated noncommercial speech and applied strict scrutiny instead. The panel suggested that a law could mandate disclosing terms of service and “existing content moderating policies.” But the panel said the state’s reporting framework did more than require disclosure of existing policies. It required companies to address “intensely debated and politically fraught topics, including hate speech, racism, misinformation, and radicalization.”

While that decision invoked the political nature of the compelled speech, other cases in which courts found that speech was too controversial to be “commercial” don’t necessarily implicate hot-button political disputes. For example, in another case, the Ninth Circuit considered a different California law requiring online service providers to report on risks their products pose to children and the steps they take to mitigate them. The court also blocked the law under strict scrutiny, reasoning that the required reports didn’t qualify as “commercial speech” because they required “businesses to opine on and mitigate the risk that children are exposed to harmful content online.” Still, the reports implicated free speech values in a different sense: Identifying and characterizing “harmful” content can raise concerns about censorship and editorial judgment over other’s speech.

Even more remote from considerations of political disputes and free speech values are cases involving scientific information. For example, in 2019, shortly after NIFLA, the Ninth Circuit applied Zauderer and upheld a municipal ordinance requiring retailers to warn consumers that storing a cell phone in pants or a shirt pocket might exceed federal radiofrequency (RF) radiation guidelines. While it acknowledged disagreement over the dangers of RF radiation, the court pointed out that the ordinance did not “force cell phone retailers to take sides in a heated political controversy.” Even so, subsequent Ninth Circuit panels haven’t confined NIFLA to heated political controversies. For example, courts have struck down mandated warnings about cancer risks from acrylamide and glyphosate on grounds that scientific debate over the risk indicated a “controversy,” thereby failing Zauderer.

Recent cases not only have narrowed what counts as “uncontroversial” but also have renewed attention to the limits of what counts as “commercial speech.” The Supreme Court has described commercial speech as speech relating to “the proposal of a commercial transaction,” such as advertising. Courts often look to factors such as whether the speech promotes products or is economically motivated. But courts disagree over how far that category extends beyond direct commercial contexts. This creates uncertainty for AI transparency laws where the required disclosure does not tie closely to economic activity.

Courts also disagree about when Zauderer applies rather than Central Hudson. A broad view treats Zauderer as a more permissive rule for compelled commercial speech and Central Hudson as a more demanding standard for restrictions on commercial speech. Under this approach, Zauderer can apply across a wide range of state interests, so long as the required disclosures are “uncontroversial.” Courts have applied it to warnings about the dangers of cigarettes, sugar, and cell phone radiation, as well as nation-of-origin labeling on meat products and Securities and Exchange Commission-mandated share buyback rationale disclosures.

Other courts treat Zauderer more as an exception to Central Hudson than as an alternative. If a compelled disclosure doesn’t satisfy Zauderer, then the court may ask whether it can survive under Central Hudson instead. Some decisions avoid the Zauderer question altogether by starting with Central Hudson and upholding the law if it survives that more demanding test.

A more skeptical view, advanced by some judges and businesses challenging disclosure laws, is that Zauderer should be confined to its original anti-deception setting. They argue that Zauderer should, at most, apply only to laws that address some form of misleading or deceptive commercial speech, such as a disclosure meant to correct misleading advertising, because the Supreme Court has applied it only to that setting. This remains a minority position, but if it gains traction, it could unsettle the broader application of Zauderer in product labeling and other disclosure cases.

What this means is that First Amendment doctrine is becoming unsettled along two axes at once: what counts as commercial speech, and what standard applies once speech is treated as commercial.

One way to understand the recent cases is that courts may be drawing an implicit line between two sorts of compelled disclosure laws. In one category are disclosures that suggest a product is harmful or otherwise undercut the company’s own message. Judges may be more cautious about those. In the other category are disclosures that simply tell users how a product works without embedding the government’s own views or judgment. This isn’t a formal doctrinal rule, but it helps explain why some disclosure mandates attract more judicial skepticism than others. It can also guide how AI transparency laws can be designed to be more robust to First Amendment challenges.

The broader point is that AI transparency laws are not doomed, but these laws can no longer be drafted on the basis that factual disclosure requirements will automatically receive deferential review. Recent developments in commercial speech doctrine and Zauderer’s scope make that body of First Amendment law less settled than it once appeared.

How the District Court Evaluated AB 2013 and Why It Matters

Judge Bernal’s opinion understood that Ninth Circuit precedent makes strict scrutiny the default starting point for a direct public disclosure requirement unless the law qualifies as commercial speech. Bernal then asked whether AB 2013 regulated “commercial speech,” which would permit a lower standard of review. He concluded that it likely did, reasoning that the disclosures related to an “actual or potential” commercial transaction because they gave the public information relevant to comparing AI models on the market. He rejected xAI’s argument that AB 2013 was really aimed at rooting out certain kinds of bias in AI training data—an argument xAI supported with statements made in the Senate Floor Analysis. Bernal noted that the bias language came from a supporter’s testimony rather than any legislator’s statements or the statute itself. He emphasized that nothing in the law requires disclosures relating to bias or suggests an effort to steer model outputs.

Bernal then considered whether to apply the more lenient Zauderer standard. While he suggested he might be inclined to find that AB 2013’s disclosures were purely factual and uncontroversial, he reasoned that “this case is even further afield from the original context under which Zauderer arose,” given the Supreme Court’s limited use of Zauderer outside misleading advertising. He therefore applied the more stringent Central Hudson standard and asked whether the law was “more extensive than is necessary” to serve the government’s interest. Bernal concluded that it was too early in the litigation to determine whether AB 2013’s required “high level summaries” would satisfy Central Hudson’s tailoring requirement. That matters because Central Hudson gives courts more room to ask whether the legislature required more disclosure than necessary to achieve the law’s objectives, even if some disclosure would be permissible.

California won for now, but the opinion signals some of the First Amendment questions that similar transparency laws may face if challenged. If courts evaluating AI transparency laws follow the same path as the district court in AB 2013—or apply even higher scrutiny, as some Ninth Circuit opinions suggest—future challenges will be harder to predict. Government attorneys defending these laws may need to address not only whether required disclosures are factual and not “controversial” but also whether transparency laws are sufficiently connected to commercial activity and appropriately tailored to the interests they serve.

That’s why the Ninth Circuit’s decision could matter beyond AB 2013: Will courts continue to treat factual disclosure requirements for businesses as commercial speech? Will they limit Zauderer to cases where the government is trying to prevent deception in the marketplace?

Implications for Legislators

For legislatures, the practical lesson is that it increasingly matters how AI transparency laws are framed. The starting point is still what problems the law tries to solve, and what information is needed to address them. But in deciding how to structure public disclosure requirements, legislators may want to keep some basic principles in mind:

• Focus on defining public disclosure requirements such that they are aimed at eliciting facts rather than judgments.
• Connect disclosures to AI products or services and how they are developed or used, where possible, especially where that information may help users and consumers evaluate a model.
• Avoid requiring businesses to adopt the government’s framing of contested issues or make judgments based on it.

In that respect, training data transparency laws may be on firmer footing when they call for disclosure of facts that help consumers evaluate a model, such as its training data sources. These laws are harder to defend when they call for interpretive judgments instead, such as assessments of possible bias. The same principles can apply to other forms of AI transparency.

Most importantly, legislators should keep in mind that this doctrine is in flux. AI transparency laws like AB 2013 are arriving at a moment when at least some courts appear to be rethinking how to assess mandated corporate disclosure. First Amendment doctrine is shaped by whatever disputes reach the Supreme Court. Those disputes often arise in settings far removed from ordinary commercial disclosure and often involve highly charged political topics. But the rules those cases produce don’t usually stay in their original contexts; instead, they apply much more broadly. NIFLA, for example, arose in the context of abortion, yet it now influences how courts review compelled disclosures in very different settings, such as product risk warnings.

So even when a case seems to be unrelated to AI transparency, its outcome and reasoning may still matter. Policymakers interested in transparency laws therefore need to watch a broad range of First Amendment decisions. That’s also a reason to draft carefully. There’s an old legal adage that bad cases make bad law. An ill-designed statute may do more than just lose in court. It may create precedent that makes better statutes harder to sustain.