Commentary. The artificial-intelligence debate has moved past the question of whether regulation is necessary. The more useful question now is whether governments can write rules that reduce serious risks without freezing a technology whose benefits are still emerging.
Our view is that they must try—and that voluntary promises by technology companies are not enough. Recent hearings in Britain exposed a significant divide among leading firms: OpenAI and Anthropic backed legally binding safety requirements, while Meta and Google DeepMind argued for existing regulatory structures and technical standards. [1] That disagreement is not a reason to wait. It is evidence that the industry itself has not settled on a reliable system of accountability.
Why binding rules are justified
AI systems can affect employment, public services, privacy, elections and national security at a speed that outpaces ordinary legislative cycles. The risks are not limited to spectacular future scenarios. Poorly tested systems can already produce discriminatory decisions, expose sensitive information or make confident errors in settings where users lack the power to challenge them.
That is why proposals for independent testing, whistleblower protection and a public authority capable of reviewing high-risk systems deserve serious consideration. In the United States, candidates competing to represent San Francisco have broadly agreed that Washington needs guardrails, although they differ over the design of a regulator and the scale of protections for workers facing displacement. [2]
A useful model would focus on outcomes rather than on every algorithm. Developers of powerful or high-impact systems should document training and testing, disclose known limitations, provide meaningful routes for appeal and accept liability when negligence causes foreseeable harm. Independent auditors—not only company compliance teams—should verify claims in areas such as cybersecurity, safety and discrimination.
The strongest counterargument
Opponents of sweeping regulation have a legitimate concern: badly designed rules can entrench the largest firms. Compliance costs may be manageable for multinational companies but prohibitive for smaller laboratories, universities and open-source developers. Rules drafted around today’s systems could also become obsolete before they take effect.
There is a further danger in treating every AI application as equally risky. A medical diagnostic tool, a workplace scheduling program and a generative writing assistant do not warrant identical obligations. Broad licensing schemes could slow useful experimentation, raise prices and push development to jurisdictions with weaker safeguards.
Those objections argue for proportionate regulation, not regulatory surrender. A tiered framework could impose the strictest requirements on systems used in critical infrastructure, law enforcement, health care, finance and defense, while allowing low-risk applications to operate under lighter transparency rules. Small organizations could receive technical assistance and shared testing infrastructure rather than simply being handed the same bill as the largest companies.
Regulation should be a condition of trust
The industry’s preferred alternative is often a combination of voluntary commitments, technical standards and existing laws. Those tools have value. Standards can be updated faster than statutes, and companies often possess expertise governments lack. But voluntary systems depend on firms policing commercial incentives that may reward speed, secrecy and market dominance.
The debate has already reached a point where even prominent executives are calling for stronger oversight. Anthropic Chief Executive Dario Amodei has argued for binding requirements, third-party testing and government authority to block deployments judged unacceptably risky, while proposing a stronger federal role for AI standards. [3]
That does not make industry proposals automatically wise. Companies may support regulation that burdens future competitors or validates their own technical approaches. Legislators should therefore resist writing rules at the request of any single corporate bloc. Oversight must be transparent, independently funded and subject to democratic review.
Our position is straightforward: innovation should continue, but access to the public should depend on demonstrable safety. The objective is not to make AI risk-free—no technology can meet that standard. It is to ensure that the people deploying powerful systems can explain what those systems do, detect when they fail and compensate those harmed by preventable failures.
Waiting for a perfect framework would be another form of policy. The better choice is a reviewable, risk-based system that can improve as evidence accumulates. Innovation needs room to breathe, but public trust needs something firmer than promises.