Commentary. The argument over artificial intelligence regulation has been framed as a choice between innovation and control. That is a false choice. The real question is whether governments can establish narrow, enforceable guardrails before the most powerful systems become embedded in institutions that citizens cannot realistically opt out of.
That question has become more urgent as the United States presses G20 members to avoid creating new AI rules, while major technology companies lobby lawmakers against restrictions on open-source models. The industry’s position deserves to be taken seriously: poorly designed regulation can protect incumbents, slow useful research and push development into jurisdictions with weaker transparency. But those risks do not justify a regulatory vacuum.
Open-source AI has genuine public benefits. Researchers, small companies and public-interest groups can inspect, adapt and improve models that would otherwise be controlled by a handful of corporations. Open systems can broaden competition and make it easier to identify flaws. Nvidia, Microsoft, Meta and other technology companies argue that “premature restrictions” could stifle innovation and send it overseas, advocating targeted legal and commercial frameworks instead of sweeping limits.[Source]
That warning should not be dismissed as self-serving simply because it comes from companies with enormous commercial stakes. The United States needs competition in AI, and regulation that treats every model as equally dangerous would be counterproductive. A small research model used for translation should not face the same obligations as a system deployed in medical triage, critical infrastructure or mass surveillance.
Yet the industry’s preferred distinction between “open” and “closed” systems can obscure the central issue: capability and use. A model’s weights may be freely available, but the consequences of its deployment are not necessarily diffuse or benign. A powerful system that can generate fraud, automate cyberattacks or produce convincing disinformation may create risks that are difficult to reverse once the model is widely copied.
What proportionate rules would look like
Governments should resist theatrical demands for a universal shutdown mechanism and equally theatrical promises that voluntary safeguards will solve everything. Instead, rules should attach to demonstrable risks and real-world uses.
- Require testing and incident reporting for systems above clearly defined capability thresholds.
- Place stronger duties on deployers using AI in employment, credit, health care, policing and education.
- Protect independent researchers and whistleblowers who expose failures.
- Require provenance and disclosure measures where synthetic media could materially affect elections or public safety.
- Review rules regularly so that low-risk tools are not trapped under obligations designed for frontier systems.
This approach would also address a legitimate concern raised by opponents of regulation: governments are often slow, technically inexperienced and vulnerable to lobbying. A bad law could freeze today’s assumptions into tomorrow’s market. Regulatory agencies therefore need technical expertise, public transparency and sunset clauses—not simply more powers.
The alternative, however, is not freedom without consequences. It is a transfer of decision-making from elected institutions to companies whose incentives are tied to speed, market share and investor confidence. The United States’ call for a hands-off approach at the G20 reflects a real geopolitical concern: Washington does not want competitors, particularly China, to gain an advantage through looser development rules elsewhere.[Source] But national competition cannot become an excuse to make the public absorb private risks.
There is also a democratic cost to delay. If citizens encounter AI mainly through opaque decisions about jobs, benefits, insurance or public services, trust will deteriorate—not because people oppose technology, but because they cannot challenge decisions they do not understand. Clear accountability is not an enemy of innovation; it is part of the infrastructure that allows innovation to endure.
Our position is therefore neither a ban nor a blank cheque. Open-source development should remain possible, and regulation should be narrow enough to preserve experimentation. But the most capable systems and the highest-stakes deployments require enforceable duties before harm becomes the price of learning.
The strongest argument from the technology sector is that lawmakers should avoid blunt instruments. They are right. The stronger public argument is that “avoid blunt instruments” cannot mean “avoid rules.” The task is difficult, but difficulty is not a policy.