Commentary. The argument over artificial-intelligence regulation is increasingly framed as a choice between two bad extremes: pause the development of the most powerful systems, or accept that regulation will always trail innovation. That is a false choice. The better course is disciplined, internationally comparable testing of high-risk AI, paired with clear responsibility when companies cause harm.
The pressure is now unmistakable. In the United States, lawmakers have proposed a temporary halt on the development of certain advanced systems until a new federal authority can establish safety rules. The European Union, meanwhile, is implementing its AI Act in stages, while the United Kingdom is still debating whether its largely sector-based approach is strong enough. The result is a regulatory landscape that is neither fully permissive nor coherent.
Our view is that governments should resist both panic and complacency. A blanket pause sounds attractive because it promises time to understand technologies that may affect employment, security and democratic life. But a moratorium would be difficult to define and enforce. Developers could move research across borders, smaller firms could be shut out while established incumbents consolidate their advantage, and governments could lose access to systems that might improve medicine, public services or scientific research.
The opposing argument deserves equal weight. AI companies have repeatedly demonstrated that deployment can outpace safeguards. Systems can generate convincing falsehoods, expose sensitive information, discriminate through automated decisions or enable new forms of fraud. The problem is not merely that a model may produce an embarrassing answer. In high-impact settings, an error can deny someone a loan, distort a medical decision or amplify a security threat.
That is why safety cannot remain a voluntary promise. A recent proposal in Washington calls for continuous testing, independent stress tests, audits, transparency and legal accountability for developers of frontier systems. Those ideas point toward a workable principle: obligations should rise with demonstrated capability and foreseeable risk, rather than attach indiscriminately to every algorithm.
Common tests, local rules
International coordination is essential, but a single global rulebook is neither realistic nor necessarily desirable. Countries will differ on privacy, labor rights, speech and national security. Democratic governments should retain room to make those choices openly. What they should not accept is a world in which the same powerful model receives radically different safety scrutiny depending on where it is trained or sold.
Common evaluation methods could provide that floor. Governments and qualified independent laboratories should be able to test advanced systems for cyber misuse, biological risks, deception, privacy violations, dangerous autonomy and discriminatory performance. Results should be reported in a form that regulators can compare, while proprietary details remain protected where disclosure would create security risks.
Critics will warn that mandatory testing could slow innovation and burden smaller developers. That concern is legitimate. Compliance rules designed for the largest laboratories should not automatically be imposed on a university project or a small business using an off-the-shelf tool. Regulation should therefore distinguish between building a frontier model and deploying an existing one, provide shared testing infrastructure, and reserve the heaviest duties for systems with the greatest potential impact.
Industry also has a point when it argues that rigid technical rules can become obsolete quickly. But flexibility is not the same as discretion without consequences. Outcome-based standards can require developers to demonstrate that a system meets measurable safety thresholds, while allowing different technical methods to reach them. Independent oversight matters because self-assessment alone cannot resolve conflicts between commercial speed and public protection.
Accountability is the missing link
The essential question is not whether AI should exist. It is who bears responsibility when a company knowingly releases a system whose risks it failed to test, disclose or control. Liability rules should reflect that responsibility, with stronger penalties for concealment and reckless deployment than for good-faith mistakes that were promptly corrected.
The emerging debate is therefore an opportunity to replace slogans with institutions. A pause may be justified for a narrowly defined capability if evidence shows an immediate, extraordinary danger. In general, however, indefinite freezes and unregulated acceleration are both failures of governance.
Archange Shadows supports a middle path: enforceable safety requirements, independent evaluations, transparent incident reporting and international cooperation on common tests. That approach will not eliminate risk. No serious policy can promise that. It can, however, make progress conditional on evidence—and make those who profit from powerful systems answerable to the public they affect.
Sources: Eurasia Review analysis; European Parliament Research Service briefing; PCWorld report on proposed US safeguards; The Tennessean overview of US and EU regulation.