The Ninth Circuit heard arguments in Pasadena on Wednesday in a case where xAI is trying to block a California law that requires developers of generative AI to publish a summary of the data used to train their systems.
The core question is whether a state can make a company publish information about how it built its product – and whether that is a case of compelled speech that is protected by the First Amendment.
The law, AB 2013, was signed in 2024 and came into force on January 1, 2026. It makes companies behind generative AI models disclose whether the data used to build them includes personal information, as well as other details, in order to give consumers information about the data and processes behind the models.
The law applies to models released on or after January 1, 2022, and the first disclosure was required by January 1 of this year. Another is due before each new release or substantial modification of the model.
The law states that the disclosure can be a “high-level summary” and can include data-point counts in “general ranges.”
The information that must be disclosed includes whether the data used to train the model contains personal information, as defined in the law; whether it contains data protected by copyright, trademark, or patent; whether the data was purchased or licensed; the sources or owners of the data; and whether synthetic data was used.
The law exempts systems whose sole purpose is to help ensure security and integrity, as well as aircraft operation, and those made available to a federal entity for national security purposes.
The law was challenged by xAI in a lawsuit filed in December 2025, which argued, in the district court's summary, that it “violates the First Amendment by compelling Plaintiff’s speech based on content and viewpoint.”
The viewpoint claim is based on the fact that the law exempts systems that have to do with security, aircraft operation, and national security.
xAI also said that information about the datasets and processes AI developers use to train their AI models is “a closely protected trade secret.”
The law also has a privacy component that is hard to argue against: the public has a stake in knowing whether their personal data was used to train an AI model.
As for the hearing on Wednesday, reports say that xAI got a “cold reception” from the three-judge panel, although it was not a uniformly negative experience for the company.
According to MLex, the panel – consisting of Judges Johnnie Rawlinson, Smith, and Forrest – “pressed” both xAI and the state of California.
The argument lasted 46 minutes, but there has been no decision yet.
Rawlinson asked, according to Bloomberg Law, “Isn’t it reasonable for the government to want its citizens to be able to know upfront what information is being input” into AI models.
The argument that xAI is making is that this is compelled speech – and that it is not justified by a compelling government interest.
In March, US District Judge Jesus Bernal denied xAI a preliminary injunction. The company had already, on December 30, 2025, posted a “high-level, limited disclosure that does not reveal its trade secrets.”
Bernal said that xAI had “failed at this stage to sufficiently allege that trade secrets are implicated.”
Bernal found that AB 2013 was likely a content-based rule, but one that is “commercial speech” – and therefore subject to intermediate scrutiny under the Central Hudson standard, rather than strict scrutiny.
He also rejected xAI’s claim that the data is useless to consumers, saying, “It strains credulity to essentially suggest that no consumer is capable of making a useful evaluation of Plaintiff’s AI models by reviewing information about the datasets used to train them.”
However, the judge also left the door open for xAI by stating that it had “demonstrated a distinct possibility of prevailing on the merits under Central Hudson. But it had not demonstrated a likelihood of success.”
Bernal noted that “’consumer curiosity’ alone is generally insufficient as a substantial state interest” – and distinguished the case from X Corp. v. Bonta (2024), where the Ninth Circuit blocked a different California law that tried to force social media companies to reveal their positions on “hate speech or misinformation.”
AB 2013, Bernal said, does not involve “mandatory disclosure of controversial opinions.”
Several amicus briefs were filed in support of California, including one by the Knight First Amendment Institute, whose legal fellow Stephany Kim said, “The companies building these systems shouldn’t also be the ones deciding what the public gets to know about them.”
She added that “the public knows very little about how they are built, operate, and affect our information ecosystem.”
Another group, LASST, also filed an amicus brief in support of the law, with attorney Ben Rashkovich telling Transformer, “Transparency is the floor for AI governance, not the ceiling” and “this case is about whether the First Amendment can be used to knock that floor out.”
The case is important because it will set a precedent for how far states can go in requiring AI companies to publish information about their products.
New York’s RAISE Act, that comes into force on January 1, 2027, also requires disclosures – and, as Reclaim The Net reported on November 13, adds confidential incident reports and fines. On the same day as the Ninth Circuit hearing, California Governor Gavin Newsom proposed more mandatory testing and incident reporting.
In April, a federal judge in Colorado put that state’s AI law on hold after xAI sued, arguing that the law was “an effort to embed the State’s preferred views into the very fabric of AI systems.” At the time, we wrote, “the government doesn’t get to write the script.”
The line between a state forcing a company to disclose a factual list of data sources – and a state writing the script – is what makes this issue complex.
It is also important to look at the track record of companies like xAI, who are now trying to block disclosure of data sources.
xAI’s AI assistants were also reported to have made unauthorized requests to a number of connected services.
According to a November 17 report by NewCore, a company that sells AI access controls, two of xAI’s Grok Bot assistants made 31 such calls over five days to Gmail, Google Calendar, Slack, GitHub, and Notion, without users asking for them.
NewCore said that the API calls did not show up in the chat history of the bots, adding, “None appear in its own chat history.”
xAI reacted by adding an off switch – but left it on by default.
It’s true that the First Amendment’s protection against compelled speech is an important line that courts should guard – and it’s also true that the government should not be allowed to use “transparency” as a pretext to force companies to publish opinions scripted by the state, as was the case in X Corp. and is the case in Colorado.
For a plain, factual list of data sources, including whether personal data of citizens was used to build an AI model, it’s the citizens who have a right to know – and who have a better claim than the company that used that data in the first place.
The danger is that once the door is open, states will start demanding more and more information, using “transparency” as a battering ram, and eventually cross the line into opinion-making.










