FOLLOW US

WASHINGTON — President Donald Trump dismissed growing AI safety warnings Monday, arguing that his leadership and existing government powers were enough to contain the technology. His administration’s own policy documents, however, describe risks requiring federal coordination, technical testing and legislation.

Trump called opposition to artificial intelligence and data centers a “SICK conspiracy” and said China would benefit from efforts to constrain development. In his Truth Social post, he also asserted that Washington already possessed substantial criminal and regulatory authority over AI companies. Reuters reported his remarks Sept. 14.

His argument went beyond defending current programs. Trump presented a “STRONG AND SMART (High IQ!) PRESIDENT” as the only safeguard the technology needed.

Yet the policy choices already before Washington are specific: whether developers must maintain emergency shutdown capabilities, whether outside investigators can examine failures and whether violations carry penalties. Those questions concern powers that officials and reviewers can exercise when a company disagrees with them.

USA Herald previously examined warnings coming from inside the AI industry. Trump’s response brings the dispute directly into government, where voluntary cooperation and enforceable obligations offer different kinds of protection.

The White House has already asked Congress for safeguards

In March, the administration publicly acknowledged concerns about AI’s effects on children, communities and household electricity bills. Its national legislative framework announcement called for Congress to strengthen the government’s ability to combat AI-enabled scams and address national security concerns.

The White House also sought protections against sexual exploitation and encouragement of self-harm on AI platforms used by minors. It called for parental account controls and policies shielding residential electricity customers from data-center costs. The announcement even used the word “guardrails” when describing its proposals concerning speech and censorship.

Those recommendations did not amount to support for every restriction proposed by AI researchers. They did, however, acknowledge problems that presidential assurances alone would not resolve. The administration said it wanted Congress to turn its framework into legislation Trump could sign.

The difference between that position and Monday’s rhetoric is consequential. An administration can oppose a particular safety proposal while identifying the protections it considers necessary. Declaring the president sufficient leaves unanswered which additional obligations, if any, the White House would accept from Congress.

AI safety warnings and the limits of federal testing.

Trump’s June 2 executive order on advanced AI innovation and security supplies a clearer account of his approach. It directed agencies to develop classified benchmarks for advanced cyber capabilities and design a voluntary arrangement for early government access to certain frontier models.

Under that arrangement, developers could provide access for up to 30 days before releasing models to other trusted partners. The order contemplated confidentiality and security protections, along with cooperation on selecting partners that could use the technology defensively.

It expressly denied authorization for mandatory government licensing, preclearance or permitting of new AI models. It also directed the attorney general to prioritize existing criminal laws against people using AI for unlawful computer access and related crimes.

Together, those provisions favor cooperation before release and enforcement against illegal conduct. The order does not establish a universal approval process under which every advanced model must receive permission before deployment.

The accompanying White House fact sheet explained the intended benefits. Federal agencies would gain cybersecurity tools, while rural hospitals, community banks and utilities could receive help protecting critical systems. A voluntary clearinghouse would coordinate efforts to identify and repair software vulnerabilities.

These programs show that the administration recognizes practical security problems and sees AI as part of the response. They also illustrate why descriptions of government oversight require precision. Giving federal experts access to technology does not automatically give them authority to block its commercial release.

What government evaluators can examine

The Center for AI Standards and Innovation, housed within the National Institute of Standards and Technology, provides another layer of scrutiny. Its published responsibilities include voluntary agreements with developers and evaluations of national security risks, including cybersecurity, biosecurity and chemical weapons capabilities.

The center also assesses foreign systems and works with other agencies on evaluation methods. Its stated mission emphasizes measurable risks and collaborative research. That is substantial technical work, although the mission statement does not establish a general licensing authority over AI developers.

In May, NIST announced new testing agreements with Google DeepMind, Microsoft and xAI. The arrangements expanded predeployment evaluations and research into frontier capabilities. NIST said earlier partnerships had also been renegotiated to reflect the center’s updated directives and the administration’s AI Action Plan.

Federal purchasing introduces another avenue for evaluation. A March agreement between the center and the General Services Administration supports assessments of security, performance and functionality for government use. That work includes methods for examining systems before deployment and measuring their performance afterward.

NIST separately describes its broader AI Risk Management Framework as voluntary. The framework helps organizations incorporate risk management into development and use; NIST says it is being revised under the White House action plan.

For the public, the unresolved enforcement question comes after an evaluation finds something dangerous. Who must receive that finding, what corrective action follows and who can compel compliance? A testing program’s existence does not, by itself, answer those questions.

Outside reviewers would gain access, not government powers

Anthropic CEO Dario Amodei’s proposal to pace frontier AI development attempts to address the information problem. He wants embedded outside evaluators with access comparable to employees conducting internal risk assessments. Anthropic has committed to pursuing that arrangement.

The proposed reviewers would examine safety practices, incidents and training processes. Amodei says their contracts should allow publication of key findings without Anthropic’s editorial control, subject to limited redactions. Those exceptions would protect matters including security-sensitive information, legal privilege and customer confidentiality. Unfavorable findings alone would not justify suppression, and reviewers could disclose when redactions affected their conclusions.

That access could expose problems outsiders would otherwise struggle to see. The proposed arrangement does not give reviewers public regulatory authority to fine a company or order a shutdown. Its effectiveness would depend on access, contractual protections and what happens when reviewers identify a serious failure.

Amodei also advocates regulation covering companies unwilling to cooperate voluntarily. His proposal seeks coordination among democracies and eventually other governments, while preserving their lead over China. He describes pacing as additional time for safety work, not an end to technical progress.

A bipartisan proposal would create shutdown authority

Congress has a more coercive option to consider. Reps. Ted Lieu, D-Calif., and Nathaniel Moran, R-Texas, announced the AI Kill Switch Act in July.

Their proposal would require covered developers to retain the ability to slow, suspend or shut down powerful systems. It would also authorize the homeland security secretary, in consultation with the commerce secretary and the director of national intelligence, to order an intervention following a qualifying incident.

The sponsors describe a graduated response, with measures calibrated to the danger. They also call for incident reporting and preservation of records so investigators can determine what happened. Those are proposed statutory duties, not powers created by the lawmakers’ announcement.

The bill text released by Lieu’s office includes civil penalties of up to $2 million per day for general violations. Violations of its emergency-authority provisions could bring penalties of up to $20 million per day, following notice and an opportunity for a hearing.

The text also provides for verification of compliance through audits, inspections or other forensic review. Companies could seek reconsideration and judicial review. A reconsideration request would not automatically suspend an emergency order.

Its definition of a covered incident reaches beyond mass casualties or economic damage. It includes interference with a lawful shutdown instruction, concealment from monitoring mechanisms and certain losses of control outside structured testing. That language would allow action on specified dangerous behavior without requiring officials to wait for a death toll.

The proposal also raises a transparency trade-off: nonpublic information submitted under its provisions would be shielded from public-records disclosure. Government investigators could obtain evidence that the public could not necessarily inspect.

In an August statement published by Moran’s office, the lawmakers argued that emergency controls would support continued innovation. They framed their proposal around preserving human intervention as systems become more capable, rather than requiring development to stop.

The bipartisan sponsorship complicates any suggestion that demands for enforceable safeguards come solely from opponents of the industry. The sponsors want AI development to continue. They want the ability to stop a dangerous system written into law.

California is building an audit system

States are moving on a separate track. On Sept. 9, California Gov. Gavin Newsom signed two measures addressing independent AI evaluation.

Senate Bill 813 establishes a framework for independent verification organizations assessing compliance with state law. Assembly Bill 1405 creates a registry for AI auditors and standards governing their independence, transparency and integrity, according to the governor’s office.

These measures address the credibility of the people examining AI systems. An audit label offers limited reassurance if the reviewer lacks independence or meaningful access. Creating an oversight framework, however, should not be confused with having completed audits or demonstrated that particular models are safe.

Newsom also called for national regulations. His position places California’s new measures within a broader argument that state action cannot, on its own, address the technology’s scale.

California’s earlier Transparency in Frontier Artificial Intelligence Act, signed in September 2025, takes a different approach. It requires large frontier developers to publish safety frameworks and establishes a mechanism for reporting potential critical incidents to the state’s emergency services office.

The law also protects whistleblowers reporting significant health and safety risks and provides for civil penalties enforced by the attorney general. Those provisions combine public disclosure with legal consequences, giving the state tools that an industry pledge alone does not supply.

Washington wants to limit the states’ reach

The administration’s effort to constrain state AI rules creates another source of uncertainty. Trump’s December 2025 executive order directed the attorney general to establish a task force challenging state laws considered inconsistent with federal policy.

It also ordered an assessment of state AI laws and explored restrictions involving certain federal funds, subject to existing legal authority. The administration argued that conflicting requirements could burden companies and weaken American competitiveness.

An instruction to challenge state laws does not automatically erase them. Particular disputes still depend on applicable law and the outcome of the relevant legal proceedings. The order itself is not a blanket judicial ruling invalidating state protections.

The White House’s March legislative recommendations similarly urge Congress to preempt state AI laws deemed unduly burdensome. The document opposes creating a new federal AI rulemaking body, favoring existing agencies and industry-led standards.

Its proposed limits include exceptions. The recommendations would preserve generally applicable state protections against fraud and harms to children, alongside state zoning authority and rules governing a state’s own AI use. At the same time, they argue that states should not regulate AI development.

These remain legislative recommendations. Their central bargain is a national approach in exchange for limits on state intervention. The strength of the federal protections would determine what Americans gain or lose from that exchange.

Across the competing plans, the same practical questions recur: who gets access, who receives warnings, who can require a correction and who answers for failure. Independent reviewers, federal researchers, state attorneys general and emergency officials would perform different roles. Each needs a defined mandate, sufficient expertise and a way to act on evidence.

Trump’s confidence does not settle those institutional choices. His own administration has documented the need for testing, coordination and additional legislation. The decisions now facing policymakers concern how much authority to give those safeguards — and whether compliance remains a company’s choice.