President Donald Trump’s preference for voluntary AI safeguards puts a central question in focus: can companies developing powerful technologies also be trusted to police their risks?
Reuters reports that his administration has secured a safety agreement with six technology companies, including OpenAI, Google and Nvidia. It calls for internal controls and independent audits but specifies no penalties for noncompliance. Meanwhile, a Reuters/Ipsos poll found that three quarters of Americans worry AI companies have not done enough to prevent serious societal harm.
The political appeal of voluntary commitments is clear. They allow governments to demonstrate concern while giving businesses flexibility. In a rapidly changing industry, rigid requirements can become outdated, and smaller developers may struggle with expensive compliance obligations. Cooperation between government and industry can also help officials understand technical risks.
Yet flexibility becomes a weakness when promises cannot be independently checked or enforced. Companies face pressure to launch products, attract investment and outperform rivals. Those incentives may conflict with delaying deployment when safety remains uncertain. An audit provides limited reassurance if its findings stay private or its recommendations can be ignored.
The debate should therefore move beyond whether regulation helps or hurts innovation. The more useful questions concern what safeguards accomplish: who evaluates dangerous capabilities, who receives incident reports, and who has authority to demand corrective action?
A credible framework could combine adaptable technical standards with public reporting, independent scrutiny and clear accountability. Requirements should reflect the scale of potential harm, avoiding unnecessary burdens on low risk applications while demanding stronger protections for more powerful systems.
Trump’s approach may encourage cooperation, but cooperation alone cannot establish public trust. That requires evidence that safeguards work and consequences when they fail. As AI becomes more influential, the durability of voluntary promises will depend on whether they deliver protection people can actually see.
): Applies to nonresidential property used in manufacturing, production, or refining of qualified products. Construction must begin after January 19, 2025, and before January 1, 2029; property must be placed in service by January 1, 2031.