The Many AI Guardrails We Already Have and the Best Way Forward
By Paul Steidler: President Donald Trump appears to have wisely decided to hold off signing an Executive Order (EO) yesterday that reportedly would have given the U.S. government up to 90 days advance notice to review AI models. Though said to be voluntary, the measure would have imposed de facto industry requirements on the AI industry and likely served as a catalyst for extensive federal regulation.
The motivation for the EO is solid. After Anthropic informed White House officials that its Mythos model could be used to identify software liabilities and launch widespread cyberattacks, the administration took a new look at AI regulation. Anthropic has wisely chosen not to release Mythos to the public.
Had Anthropic done so, the consequences would have been dire, perhaps fatal to the company while decimating its investors.
Legions of product liability attorneys would have gone after the cash-rich company for its negligence and for failing to exercise its duty of care to make sure the public is protected from defective products. The businesses, health providers, and governments harmed by Mythos would also have many attorneys willing to work on a contingency basis.
In a prescient November 2024 report on AI and the law, RAND weighed in on these issues. “Absent new law from Congress or state legislatures, tort law will be applied to harms caused by artificial intelligence (AI) when such cases are brought in court. Tort law is a form of common law, meaning that it has been formed primarily by judges ruling on specific cases and developing precedent to apply to future cases.”
The report continues, “In contrast to many other proposed regimes for governing AI, tort law is already in force and is regularly applied in state and federal court. Liability law also does not require an act by Congress, the President, or other federal or state decisionmakers to apply to AI-caused injuries.”
This does not, however, mean that new AI laws are not necessary. Sensible laws are highly preferable to executive orders, which can be repealed and changed. Sensible laws also provide much greater certainty for economic planning than executive orders.
Such laws should be bipartisan and developed in public, ideally at Congressional hearings with witnesses who testify under oath.
An excellent basis for executive orders on AI models and legislation is found in the December 17, 2024, U.S. House of Representatives Bipartisan Task Force on AI report, endorsed by Speaker Mike Johnson and Democratic Leader Hakeem Jeffires.
It calls for transparent reporting on model development that will include information about, “metrics identified in model development, plans to provide notice and explanation of the use of AI models to members of the public impacted by the models’ use, and any ongoing training, validation, and testing.”
Federal legislation is also important as it will preempt a bevy of state laws being developed in this regard, which create confusion and impose sizable regulatory costs, especially for mid-size and smaller tech companies. Model development, encompassing the testing and evaluation of AI systems, is an inherently federal responsibility.
Whatever the White House decides to do on future AI executive orders, Congressional action will be important in 2026 and beyond.