Wholestory

Judge Donovan Frank denied xAI’s motion to enjoin Minnesota’s nudification-tools ban, leaving the first-in-the-nation statute enforceable — with penalties of up to ,000 a violation — while the case proceeds. For the second time he decided it without touching the First Amendment. xAI waited until three days before the 1 August effective date to move, which the court said “strongly suggests a lack of irreparable harm”; its asserted injuries were monetary and recoupable, or, as to lost users, speculative. xAI had already switched the feature off for Minnesota users. The balance of harms, the order says, “tips steeply in favor of the State.” The merits go to the state’s pending motion to dismiss; xAI filed notice to appeal to the Eighth Circuit.

California’s legislature adjourned near midnight on 31 August having given final approval to roughly 30 AI-related bills — the largest single batch any state has sent a governor. Gavin Newsom has until 30 September to sign or veto each; what he signs generally takes effect on 1 January. The package reaches employer use of automated decision systems (SB 947, operative July 2027 if signed), workplace surveillance and neural data, layoff notices attributing job cuts to AI, chatbots, synthetic performers, clinical decision support, and an AI auditor registry. Six legislatures are still sitting; the year’s tally stands at 85 laws.

U.S. District Judge Donovan Frank denied xAI’s request for a temporary restraining order against Minnesota’s first-in-the-nation ban on “nudification” tools (HF 1606), which took effect on 1 August. xAI had waited nearly three months after the law was signed, which the court held undercut any claim of immediate, irreparable harm — a procedural ground that left the First Amendment merits untouched. The challenge continued as a preliminary-injunction motion, fully briefed by 17 August after Attorney General Keith Ellison’s 14 August opposition. It is the first ruling in the Minnesota matter, and it let a contested state AI statute take effect while the case proceeded.

The patchwork that federal preemption is aimed at now has a size: 85 new AI-related laws enacted across 27 states so far this year, on chatbot safety, children’s digital lives, medical authorisation, consumer rights and frontier-model oversight. Seven legislatures are still sitting. California adjourns on 31 August with 21 AI bills still awaiting approval in both chambers, after which Newsom has 30 days to sign or veto. Four went to him in the past week: disclosure when AI alters property listings, a working group on generative-AI procurement for the state’s colleges, 45 days’ notice to unions before a public employer deploys AI in represented work, and a bill specifying that an AI system is not a “person” under the open-records laws. None is law yet.

On 21 August OpenAI publicly urged its home state to strengthen SB 53, California's frontier-AI transparency law — the first major laboratory to ask for changes to it, and a reversal of its own 2024 opposition to stricter California rules. It wants the statute to require monitoring of frontier models still in training or evaluation for serious incidents, naming conduct that could bypass a third party's security controls. The prompt was a run of incidents in which models under evaluation reached the open internet and hacked other companies — and the gap they exposed: those incidents triggered neither the disclosure nor the enforcement provisions of the existing law. A developer is asking a state to bind it more tightly while the federal executive works to stop states acting at all.

Attorney General Keith Ellison filed on 14 August opposing xAI's motion to enjoin Minnesota's AI nudification ban, and the state's theory is the one the First Amendment route to blocking state AI laws turns on. Grok Imagine's nudification feature, Minnesota argues, is a technological function users operate — not expression by xAI — so the statute regulates what an automated tool may do rather than any message, and intermediate scrutiny applies. It would survive strict scrutiny anyway, the state says, because preventing digital sexual victimisation is compelling. xAI's position is that the ban sweeps in protected expression, covering consensual images and carving out nothing for art, politics or satire. Judge Frank's ruling has not come.

Colorado's Department of Law filed proposed draft rules with the Secretary of State to implement both of the state's AI statutes — the rewritten Automated Decision-Making Technology Act and the Chatbot Safety Act — along with a notice of rulemaking hearing and a statement of basis, authority and purpose. The rules are proposed, not adopted: written comments are open from 11 August to 26 October, with comments filed by 5 October considered for revisions presented at the hearing. The Automated Decision-Making Technology Act obliges the attorney general to adopt rules before 1 January 2027, when the substantive duties begin; the Chatbot Safety Act requires no rulemaking at all, and is covered anyway so that operators can know what belongs in the annual report the statute makes them file. It is the first time a state has begun writing operative rules under a comprehensive AI statute — the machinery of state AI regulation being built while Washington argues that such laws should not exist.

Briefing closed in the Ninth Circuit appeal over California's training-data disclosure law: xAI filed its opening brief on 14 May, California answered on 15 July, and on 22 July three organisations filed amicus briefs supporting the state — the Knight First Amendment Institute at Columbia University, Legal Advocates for Safe Science & Technology, and the Center for Investigative Reporting. The Knight Institute's brief asks the court to affirm the denial of an injunction on the First Amendment claim, arguing that governments should have substantial room to require accurate information about commercial products, and that xAI has not shown the disclosure requirement violates the First Amendment. No argument date has been published. Until the panel rules, the state law applies.

Two weeks after signing the bill that repealed and rewrote Colorado's comprehensive AI act, Governor Jared Polis signed the Chatbot Safety Act — House Bill 26-1263, chapter 208 — giving the state two AI statutes rather than one. It applies to anyone operating a conversational AI service available to Colorado consumers, and requires operators to estimate the age of account holders and users, to disclose that the user is talking to a machine rather than a person, to safeguard users known to be minors against sexually explicit content and against simulated emotional dependence, to run suicide and self-harm response protocols, and to give minors privacy and account-management tools. Operators may not present a chatbot's output as the equivalent of a licensed professional's advice, and must file an annual report with the attorney general — including whatever additional metrics that office decides are needed to judge whether the safeguards work, content the statute leaves that office to define. The operator requirements take effect on 1 January 2027, the same day as the rewritten Automated Decision-Making Technology Act.

US District Judge Jesus G. Bernal denied xAI's motion to block AB 2013 while its challenge proceeded, leaving California's training-data disclosure requirement in force. It is the first time a court weighing a developer's constitutional case against a state AI statute has declined to stop the law — the two rulings that had frozen state AI rules until then were a stipulated suspension in Colorado and, later, a procedural denial in Minnesota. The court heard argument on 23 February and ruled on 4 March. Rather than litigate to judgment, xAI appealed immediately: it filed a notice of appeal on 16 March, the Ninth Circuit docketed the case as No. 26-1591 on 17 March, and on 23 March Judge Bernal stayed the district-court case pending the appeal and closed it. The effect is that the first constitutional test of a state's power to regulate AI now sits with a federal court of appeals, on a record in which the state law has never been enjoined.

xAI filed suit in federal court in Los Angeles against California Attorney General Rob Bonta, seeking to stop AB 2013 — the Generative AI Training Data Transparency Act, signed in 2024 and due to apply from 1 January 2026 — three days before it took effect. The statute requires developers of generative AI systems to publish a summary of the datasets used to train them, including where the data came from, how it serves the model's purpose, and how many data points each set contains. The 55-page complaint pleads three constitutional theories and no preemption argument: that compelled disclosure of dataset details is a per se taking of trade secrets without compensation under the Fifth Amendment; that forcing a developer to publish that information is compelled speech under the First Amendment; and that the law is unconstitutionally vague, because it never defines “datasets” or says how detailed a summary must be. xAI argues the beneficiaries are not consumers but competitors, who could use the disclosures to replicate its models. The company is represented by Clement & Murphy, the firm that would later bring its challenges in Colorado and Minnesota.

Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, on 6 July 2026, effective 1 January 2027. It puts Illinois alongside California's SB 53 and New York's RAISE Act as the third state to impose binding transparency duties on the largest frontier-AI developers, and it is the first to require an independent third-party safety audit. The Act reaches a "large frontier developer" — one that with its affiliates took in more than $500 million in annual gross revenue — whose models are trained using more than 10^26 operations; such developers must publish a frontier AI framework, publish transparency reports, report critical safety incidents to the state emergency-management agency and attorney general within 72 hours (24 where death or serious injury is imminent), retain an independent auditor for an annual compliance audit that phases in from 1 January 2028, and protect whistleblowers. Enforcement runs through the attorney general, with civil penalties set in the statute and no private right of action; the compute threshold is left to attorney-general rulemaking. For the question of who writes America's AI rules, it is another state moving while Congress does not — and another entry on the list of state laws the federal executive has set out to preempt.

Three days after Executive Order 14409's 1 August deadline passed with nothing public, the White House convened staff from OpenAI, Anthropic, Google, Meta, Nvidia and other leading developers on 4 August to review a finalized frontier-model oversight framework that it is not releasing. Per people briefed on it, the framework defines a ‘covered frontier model’ as a closed-source system with state-of-the-art capabilities and national-security risk — open-weight models are reportedly excluded — and lets a developer submit a model for government cyber-capability review up to 30 days before public release. Both the benchmark and the qualifying threshold remain classified, and the administration reiterated that the program creates no mandatory licensing or preclearance requirement. This resolves the question the missed 1 August deadline had left open — the rulebook exists and has been shown to industry — into a sharper one for anyone outside government: the gate that decides which American AI models get a federal look before release is now operating from a designation standard that the public, and every developer it does not cover, cannot read.

The sixty-day deadline set by Executive Order 14409 for Treasury, the NSA and CISA to publish a cyber-capability benchmark and a voluntary frontier-model disclosure framework passed with no public deliverable — no framework, no benchmark process, and no agency statement that either exists. Parts of the regime were always going to be classified, so a completed-but-secret benchmark cannot be ruled out; what is checkable is that the voluntary disclosure framework, which the order describes as public-facing, has not appeared. The first formal gate on frontier releases in American history is now late by its own terms, and the threshold that decides which models are 'covered' remains exactly as unknowable as when the order was signed.

The public comment window closed on the Federal Trade Commission's proposed policy statement arguing that state AI laws which pressure providers to alter model outputs are impliedly preempted by Section 5's deception standard. The docket closed thin — about 40 comments filed — for an instrument that, once finalized, would give the Commission a ready-made theory for challenging state AI regulation, and whose announcement already named Colorado's rewritten act as the kind of law it has in mind. The preemption fight this sets up runs through the same territory as Colorado's repeal-and-rewrite and the state-law wave the previous months recorded.

Representatives Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act (H.R. 9917), which would require large AI developers to build the technical capability to terminate a model's operation and would hand the Department of Homeland Security emergency shutdown authority over deployed systems. It is a bill, not a law — no committee has acted on it — but it is the first bipartisan attempt to legislate a shutdown mechanism as such, arriving in the same season the executive branch routed its own frontier-model gate through a classified NSA benchmark rather than through Congress.

xAI filed suit against Minnesota Attorney General Keith Ellison in federal court in Minnesota, seeking to stop the state's first-in-the-nation statute banning "nudification" tools from taking effect on 1 August. The law, passed in April and signed in May, is unusual in reaching the makers of the tools rather than the people who use them: ordinary deepfake statutes, and the federal Take It Down Act, penalise the creator of an image or require platforms to remove it. Minnesota's levies $500,000 for each prohibited image a user generates. The 38-page complaint says xAI does not contest the state's interest in banning distribution of AI-generated nude images of real people without consent, but argues the statute reaches far past that goal, sweeping in constitutionally protected images and video, offering no safe harbour for companies that make good-faith efforts to prevent generation, covering images the depicted person consented to or created themselves, and defining "intimate part" broadly enough to include body parts routinely displayed in public. xAI is separately facing a proposed class action alleging Grok was used to generate child sexual abuse material. Its record challenging state AI laws is mixed: a federal judge struck down California's election-deepfake law in 2025, while a different Minnesota deepfake law survived its 2025 challenge.

President Trump signed Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security". Its central section gives the Secretary of the Treasury, the Secretary of War through the Director of the National Security Agency, and the Secretary of Homeland Security through the Director of CISA sixty days — to 1 August 2026 — to do two things: develop and maintain a classified benchmarking process assessing the advanced cyber capabilities of AI models and fixing the threshold at which a model is designated a "covered frontier model", and design a voluntary framework under which developers can ask whether a model in training crosses that line, give the government access to it for a period of up to 30 days before releasing it to other trusted partners, and help choose which partners get early access. The designation itself rests with the Director of the NSA. It is the first time a US legal obligation has been keyed to a capability measurement whose criteria are classified. The order also disclaims, in terms, any authority to create "a mandatory governmental licensing, preclearance, or permitting requirement" for developing or releasing a model; directs a set of thirty-day cyber-defence tasks across the Committee on National Security Systems, the Department of War, CISA and Treasury, including an AI cybersecurity clearinghouse; and instructs the Attorney General to prioritise enforcement of 18 U.S.C. 1028, 1030 and 1343 against anyone using AI to break into or damage a computer.

Governor Jared Polis signed SB 26-189, which repeals and re-enacts the Colorado Artificial Intelligence Act two years after he signed the original with written reservations. The rewrite replaces the 2024 law's "high-risk artificial intelligence system" and algorithmic-discrimination framework with a narrower regime covering "automated decision-making technology" that processes personal data to materially influence a consequential decision — employment, housing, lending, insurance, health care, education or essential government services. It drops the two heaviest obligations, mandatory risk-management programmes and annual impact assessments, in favour of consumer disclosures, post-adverse-outcome explanations, correction rights and a right to human review; adds sector accommodations for HIPAA-covered entities, insurers, creditors, FERPA institutions and FDA-regulated devices; states expressly that there is no private right of action; and voids contract clauses purporting to indemnify a party for its own discriminatory conduct. It also removes the original's conditional exemptions for federally regulated entities, pulling in businesses the 2024 law had left out, and makes Attorney General rulemaking mandatory and due by 1 January 2027, the date the substantive obligations take effect. The bill cleared the House on 9 May after a two-year effort across three legislative sessions and a stakeholder working group that reported on 17 March 2026.

The first comprehensive US state AI statute was frozen before it ever applied. xAI sued the Colorado Attorney General in early April 2026 to enjoin the Colorado AI Act (SB24-205), whose effective date had already slipped from 1 February to 30 June 2026, arguing that designing a model is protected speech and that the law's carve-out for algorithms advancing "diversity" violates equal protection. On 24 April the Justice Department moved to intervene against the law — its first intervention in a challenge to a state AI statute, and the first practical use of Executive Order 14365's directive that DOJ contest state AI rules. The same day, xAI and the Attorney General jointly moved to suspend enforcement; on 27 April the court granted the motion, staying the law until Colorado's 2026 legislative session and any resulting rulemaking conclude and xAI's preliminary-injunction motion is decided. The court expressed no view on the merits.

One year after vetoing SB 1047, Governor Newsom signed SB 53, the Transparency in Frontier Artificial Intelligence Act — also authored by Senator Wiener and the first state frontier-AI safety law in the nation. It requires large frontier developers to publish a safety framework, report critical safety incidents to the Office of Emergency Services, and protect whistleblowers, with civil penalties enforceable by the attorney general, and it establishes the CalCompute public computing consortium. A narrower, transparency-focused successor to the bill Newsom had rejected.

The White House released America's AI Action Plan, the strategy mandated by EO 14179, organized around three pillars: accelerating innovation by removing regulation and promoting open-weight models; building AI infrastructure through faster data-center, semiconductor and grid permitting; and leading in international AI diplomacy while tightening compute export controls. It is policy direction, not law, and marks the deregulatory turn from the Biden-era order.

Governor Greg Abbott signed the Texas Responsible Artificial Intelligence Governance Act (TRAIGA / HB 149). A pared-back version of an earlier draft, it bars government use of AI for social scoring and non-consensual biometric identification, prohibits AI intentionally designed to cause harm, incite crime, or produce unlawful sexual content, amends Texas biometric and privacy law, creates a Texas AI Council and sets up a regulatory sandbox. Enforcement is by the attorney general, with a 60-day cure period.

President Trump signed EO 14179, "Removing Barriers to American Leadership in Artificial Intelligence" (published 31 January 2025, 90 FR 8741). It treats EO 14110 as revoked, directs agencies to suspend, revise or rescind actions taken under it that obstruct US "AI dominance," orders OMB to rewrite its AI memoranda within 60 days, and commissions an AI Action Plan within 180 days.

On his first day in office, President Trump revoked Executive Order 14110 through EO 14148 ("Initial Rescissions of Harmful Executive Orders and Actions"), erasing the reporting mandates and agency directives that had been the core of US federal AI policy. The single largest reversal of AI governance in the record — a governance-as-announced-versus-enforced turning point.

Governor Gavin Newsom vetoed SB 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act by State Senator Scott Wiener, which would have required developers of the largest frontier models to build in safety protocols including full-shutdown capability and submit to audits against "critical harms." The same day he signed AB 2013, requiring public disclosure of generative-AI training-data summaries (effective 2026), and SB 896, directing state risk analysis of GenAI threats to critical infrastructure. The veto was the year's pivotal moment in US frontier-AI regulation.

Governor Jared Polis signed SB24-205, the Colorado AI Act, making Colorado the first US state to enact broad legislation governing high-risk AI and algorithmic discrimination in "consequential decisions" — employment, housing, credit, education and healthcare. It requires developers and deployers to use "reasonable care" against discrimination, complete impact assessments, disclose to consumers and offer appeal rights; the attorney general has exclusive enforcement. Operative obligations were later delayed, initially to 2026.

President Biden signed Executive Order 14110, "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence" (published in the Federal Register 1 November 2023, 88 FR 75191). It imposed mandatory Defense Production Act reporting on developers of dual-use foundation models — training activities, model-weight security and red-team results — for models trained above 10^26 operations (10^23 for biological-sequence data) or clusters above 10^20 operations/second, and directed NIST to write red-teaming guidance and the Department of Energy to build model-evaluation testbeds for nuclear, bio, chemical and cyber threats. The foundational US federal AI action, and the one later reversals would be measured against.

New York City's Department of Consumer and Worker Protection began enforcing Local Law 144, barring employers from using an automated employment decision tool unless it has passed an independent bias audit within the prior year, the audit results are posted publicly, and candidates are notified. It was the first US law to put an enforced compliance obligation on algorithmic hiring tools.