Commentary. The argument over artificial-intelligence regulation has entered a more serious phase. Lawmakers in the United States are proposing mandatory shutdown mechanisms and incident reporting after testing raised concerns about autonomous systems behaving in unexpected ways. California has also moved toward independent oversight of AI developers, while technology companies remain divided over how much government intervention is necessary.[1]
Our view is straightforward: the most powerful AI systems should not be governed chiefly by promises from the companies that build them. Voluntary safeguards can be useful, but they are not a substitute for enforceable duties, transparent testing and liability when preventable harm occurs.
That position does not require treating every AI model as an existential threat. Most systems are ordinary tools, and excessive regulation could impose costs on smaller developers, discourage research and hand even more power to the largest firms. OpenAI chief executive Sam Altman has argued for a lighter framework that accepts some risk in exchange for broad access to AI’s benefits, while technology executives have backed voluntary controls, external audits and board oversight.[2]
Those concerns deserve a hearing. Regulation written in panic is likely to be vague, slow and easy to evade. A rigid approval process could delay medical, scientific and accessibility applications that offer real public value. Governments also face a practical problem: AI capabilities are advancing faster than legislation, and no single national regulator can independently manage systems that operate across borders.
But these difficulties strengthen the case for carefully designed rules rather than self-regulation. The recent push for an “AI kill switch” reflects a basic principle of safety engineering: systems capable of taking consequential actions must include a reliable way for authorized humans to stop them. Proposed U.S. legislation would require developers of AI agents to build shutdown capability and report serious incidents.[1] A switch alone, however, is not enough. It may fail, arrive too late or be difficult to use when an automated system has spread across networks.
What responsible rules should require
First, developers of high-risk systems should document the environments in which their models are tested and disclose significant failures to an independent authority. “Trust us” is not an adequate public-safety standard when the consequences may affect elections, financial markets, infrastructure or personal data.
Second, independent testing should be mandatory for systems that can act without continuous human approval. Anthropic chief executive Dario Amodei has called for binding safety requirements, third-party testing and government authority to block deployments judged unacceptably risky.[2] That approach is more credible than leaving companies to define both the risks and the acceptable remedies.
Third, liability must follow control. If a company knowingly releases an autonomous product without adequate safeguards, it should not be able to shift every consequence onto users. Clear liability would also encourage better design without requiring regulators to predict every future failure.
The counterargument is that national rules could make domestic firms less competitive. European officials are already facing pressure from major technology companies to revise the European Union’s AI Act, while critics question whether existing rules cover systems tested internally but not yet placed on the market.[3] That pressure should prompt governments to improve rules, not abandon them. International standards, mutual recognition of audits and narrow definitions of high-risk applications can reduce fragmentation.
The goal should not be to freeze innovation. It should be to make safety part of innovation’s price. Airlines, hospitals and banks operate under rules because public trust depends on more than the operator’s assurances. Advanced AI deserves the same discipline when it can act, decide or cause harm at scale.
A kill switch is a useful symbol, but the real safeguard is accountability: independent scrutiny before deployment, rapid reporting after failure and meaningful consequences when companies cut corners. The technology may be new. The public’s need for responsible institutions is not.