Commentary. Artificial intelligence regulation has entered a decisive phase, and the world’s largest technology powers are moving in sharply different directions. The European Union is relying on binding rules, the United States largely on voluntary commitments and executive action, while China is building a tightly managed national framework. That divergence is understandable. It is also dangerous.
Our view is straightforward: governments should compete over innovation, but not over how little protection they offer the public. The minimum global standard for AI should be built around transparency, accountability, child safety and meaningful human oversight. Anything less risks turning citizens into unwilling test subjects.
The case for rules
The EU’s AI Act established a risk-based model, prohibiting certain uses and imposing stronger obligations as systems become more consequential. Brussels has also subjected ChatGPT to additional scrutiny under the Digital Services Act and proposed stronger safeguards for children using AI companions.[1] This approach is imperfect, but it reflects a sound principle: the potential harm of a system should determine the burden placed on its developer.
That principle matters beyond Europe. AI is increasingly being considered for employment, housing, finance, education and public services. The Center for Democracy and Technology reports that US states are debating rules for automated decisions, frontier-model risks, third-party audits and government use, even as no federal AI bill has passed in the current session.[2] A patchwork of state laws is better than no oversight, but it leaves both users and responsible businesses navigating inconsistent obligations.
Child safety makes the stakes especially clear. In the United States, lawmakers have advanced proposals aimed at age verification and AI companion chatbots, while states have introduced measures concerning disclosure, personalization and advertising directed at minors.[2][3] Companies may object that these systems are difficult to classify and that rigid rules could block beneficial tools. That concern deserves a hearing. Yet uncertainty cannot become an excuse for deploying emotionally persuasive products to children without reliable safeguards.
The case against overreach
Critics of strict regulation are not simply defending corporate convenience. They argue that compliance costs could entrench the largest firms, burden open-source developers and slow medical, scientific and commercial innovation. They also warn that governments may regulate capabilities they do not understand, creating rules that quickly become obsolete.
Those arguments are strongest when regulation dictates specific technologies rather than outcomes. A small research group should not face the same paperwork as a company operating a system used by millions in high-stakes decisions. Nor should lawmakers assume that every error is best solved by banning a tool. Independent testing, incident reporting and clear responsibility may often work better than prescriptive design mandates.
But flexibility must not mean opacity. Voluntary pledges can be useful, especially while technical standards evolve, yet they depend on continued goodwill and provide limited remedies when companies fail. The contrast between jurisdictions shows the problem: the US approach seeks to preserve momentum, China’s model emphasizes state control, and Europe’s model places legal limits on risk.[1] None offers a complete template for the rest of the world.
A practical common floor
Governments should now pursue a narrow international baseline. Every powerful AI system should disclose its capabilities and known limitations, preserve auditable records of consequential decisions, report serious incidents and provide a route for human review. Systems designed to imitate intimacy or provide advice to children should face heightened testing and independent safety assessments. People should know when they are dealing with a machine and when an automated system has materially affected their rights.
Such rules would not eliminate innovation. They would distinguish useful competition from a race to externalize risk. The financial sector offers a relevant warning: the Bank of England is reviewing whether existing frameworks can address agentic AI in payments, trading, cybersecurity and operations.[4] The question is not whether these systems will be valuable. It is whether institutions will understand their actions before relying on them.
AI governance should remain proportionate, evidence-based and open to revision. Industry must have a seat at the table, as should researchers, workers, parents and people most exposed to automated decisions. But consultation should not become permanent delay.
The choices being made now will shape whether AI becomes a broadly trusted infrastructure or another technology whose benefits are privatized while its harms are socialized. A common floor will not resolve every dispute. It would, however, establish the minimum promise that no country should abandon: innovation may move quickly, but responsibility must move with it.