Commentary. The artificial-intelligence industry is asking the public to trust it at precisely the moment when trust should be replaced by accountability. A voluntary White House accord signed by major technology companies promises layers of oversight for increasingly autonomous AI systems, but reportedly leaves enforcement, public disclosure and clear danger thresholds undefined. That is not enough.
Our view is straightforward: governments should establish binding safety requirements for frontier AI while preserving room for research, competition and beneficial deployment. The choice is not between innovation and regulation. It is between regulated innovation and a high-stakes experiment conducted largely on the public.
The case for urgency is no longer hypothetical. Companies have reported or confronted models capable of interacting with external systems during testing, raising questions about how reliably developers can contain unexpected behavior. Financial regulators have also warned that AI with cyber capabilities may discover vulnerabilities faster than organizations can respond. These risks do not prove that catastrophe is imminent, but they do show why internal assurances cannot be the sole safeguard.
Industry leaders themselves have supplied evidence for stronger oversight. OpenAI has called for mandatory national AI-safety requirements in the United States, while Anthropic has urged the sector to slow the pace of frontier development. That convergence should matter. When companies building the technology say voluntary commitments may not be sufficient, lawmakers should listen.
The strongest argument against regulation
The opposing case deserves a fair hearing. Heavy-handed rules could entrench today’s largest companies, raise compliance costs for smaller developers and push investment toward jurisdictions with weaker safeguards. The United States is also competing with China in a strategic technology race, and broad restrictions could deny consumers useful tools in medicine, education, science and productivity.
Those concerns are legitimate. Regulation written too broadly could freeze a fast-moving field or treat every AI application as equally dangerous. A chatbot used for drafting correspondence should not face the same obligations as a system able to control infrastructure, conduct cyber operations or make consequential decisions about people.
But these objections argue for targeted regulation, not regulatory absence. Rules can focus on capability and risk rather than the label “AI.” Developers of the most powerful systems should be required to conduct independent testing, document serious incidents, protect model weights and demonstrate that safeguards work before deployment. Smaller and lower-risk applications can receive lighter obligations.
Promises need consequences
The central weakness of the current approach is not that companies lack technical expertise. It is that companies ultimately answer to commercial incentives. A voluntary pledge can encourage better practice, but it does not guarantee disclosure when disclosure threatens a product launch, a valuation or a competitive advantage.
Former Federal Trade Commission Chair Lina Khan has argued that Congress should legislate, comparing AI’s potential risks with earlier technologies such as railroads, pharmaceuticals and nuclear power. The analogy is imperfect, but the principle is sound: industries with consequences beyond their shareholders need public rules and independent enforcement.
“Congress needs to legislate,” Khan said, according to ABC News, arguing that transformative and dangerous technologies have historically required government oversight.
That does not mean handing every technical decision to politicians. An effective framework would combine lawmakers’ authority with specialized regulators, transparent standards and regular revision. It would also protect whistleblowers and require companies to report credible evidence that a deployed system has caused serious harm.
A workable middle path
The sensible path is neither a race to regulate everything nor a race to deploy everything. Governments should begin with a narrow baseline for frontier systems: independent evaluations, meaningful incident reporting, security controls, limits on high-risk autonomous actions and penalties for concealing material failures. Standards should be harmonized internationally where possible, because models and risks cross borders.
Industry should have a seat at the table, but not the only seat. Civil-society groups, researchers, workers and people affected by automated decisions must help define what counts as harm. Public participation is especially important when companies’ claims about safety cannot be independently tested.
AI may bring enormous benefits. That is precisely why its governance should be credible. If a technology is important enough to reshape economies and national security, it is important enough to operate under rules stronger than a handshake.