Regulating artificial intelligence means acting on a technology that changes faster than the laws expected to govern it. At the G20 technology meeting in North Carolina, the United States proposed a light-touch approach represented by the Carolina Principles: avoid broad new AI rules when existing law can address a problem, and intervene mainly when genuinely novel risks appear. The position reflects a belief that innovation needs room and that premature legislation may preserve the present rather than govern the future.
Europe has chosen another path. The AI Act organizes obligations according to risk: some applications are prohibited, others require assessment, documentation and oversight, while many ordinary uses remain relatively unrestricted. It is therefore misleading to describe the debate as freedom against prohibition. Both models set limits; they disagree about when those limits should apply, who must demonstrate safety and how explicit responsibility should be before harm occurs.
Companies often argue that fragmented rules increase costs and uncertainty. A startup may not have the legal resources to meet different obligations in every jurisdiction, while a large group can turn compliance into a competitive advantage. Yet an absence of rules does not automatically distribute power. It may leave power with the platforms that own models, data and infrastructure. Every regulatory choice therefore creates a market, including a decision presented simply as removing obstacles.
Open models make the problem more complex. Restricting available weights can reduce some dangers while concentrating research and capability among the largest companies. Allowing them to circulate supports independent evaluation, local adaptation and competition, but makes control more difficult. There is no single solution for a model used in language research and a system capable of operating in cybersecurity or biology. Governance needs to examine real capability and context rather than commercial labels.
The G20 is looking for shared principles at a time when AI has become industrial policy. The United States, China and Europe are not discussing safety alone. They compete over chips, energy, talent, standards and market access. A rule can protect a right while influencing who can afford the technology. Neutrality is impossible. Even a decision not to intervene benefits some actors and transfers the risk of failure to others.
The decisive quality is the ability to change course. Good regulation should be updateable, measurable and capable of distinguishing among uses. It should demand transparency when a decision affects people, audits when scale makes errors systemic and responsibility when a system moves from words to actions. It should also preserve room for research, artistic experimentation and low-risk tools. Flexibility need not mean absence of control; it can mean proportionate control.
REGULATED FREEDOM is not a contradiction. Every technology already operates inside infrastructures, contracts and permissions. The question is who defines them and who can challenge them. If rules are written only by platforms, freedom means their terms of service. If rules are imposed without understanding the technology, they can quickly become useless. Governing AI means building a space where innovation and protection can both be tested instead of becoming opposing slogans.