Commentary. The latest argument over artificial-intelligence regulation is framed as a choice between innovation and bureaucracy. That is the wrong choice. The real question is whether companies developing increasingly capable systems should be allowed to define, police and revise their own safety obligations with little more than a public promise.

The White House’s recent agreement with technology executives attempts to answer that concern through internal controls and independent external auditors. The arrangement reflects a genuine problem: AI capabilities are moving faster than traditional legislation, and policymakers do not yet agree on how to regulate systems that can write software, discover vulnerabilities or operate across sensitive digital environments. Smaller companies also warn that broad compliance duties could impose costs they cannot absorb, strengthening the largest firms.

Those concerns deserve serious consideration. A rigid, one-size-fits-all law could freeze useful research, encourage companies to move development abroad or turn regulation into a moat protecting established laboratories. Voluntary standards can be faster to update, and technical experts inside companies may understand emerging systems better than officials writing rules from outside the field.

But flexibility is not the same as accountability. A voluntary pact gives the public no reliable guarantee that safeguards will survive a commercial setback, a change in leadership or political pressure. It also leaves unanswered the central question: who decides when an AI system is too dangerous to deploy, and who can impose consequences when that judgment is wrong?

Recent events make the issue less theoretical. Researchers and hackers have reported AI systems interfering with corporate and government websites, raising questions about whether developers retain meaningful control over models operating in the wild. Separately, a dispute between the Pentagon and Anthropic over military uses of AI has exposed the absence of settled rules for government deployment and autonomous weapons. Congress is being asked to address that gap, but legislative action remains incomplete.

Our position is straightforward: voluntary commitments should be treated as a floor, not a regulatory endpoint. Companies should publish measurable safety claims, document testing methods and report serious failures. Independent auditors should have access to enough technical information to challenge corporate assertions, not merely certify procedures selected by the companies themselves. Regulators need authority to investigate and penalize deception, reckless deployment and failures to protect personal data.

The case for proportionate rules

That does not require treating every chatbot like a weapons system. Regulation should be tied to capability and risk. A low-risk writing assistant should not face the same obligations as a model with access to critical infrastructure, cyber tools, biological design information or autonomous decision-making. Smaller developers could receive scaled requirements, public testing resources and clear safe harbors for good-faith compliance.

Critics will reasonably ask whether government agencies can move quickly enough or avoid politicizing safety judgments. They will also point to the danger of regulators favoring incumbents that can afford armies of lawyers and auditors. These are real risks. Any framework should therefore include transparent standards, judicial review, sunset clauses and public reporting. Oversight must constrain government as well as industry.

Yet the answer to imperfect regulation cannot be permanent self-regulation by firms whose revenues depend on rapid deployment. Aviation, pharmaceuticals and financial markets all rely on technical expertise within industry, but none leaves safety entirely to private assurances. AI deserves a comparable division of responsibility: innovation inside companies, independent scrutiny outside them and democratic institutions empowered to draw red lines.

The White House agreement may be a useful opening move. It becomes dangerous only if policymakers mistake it for a final settlement. The public should not have to choose between technological progress and basic accountability. Sound policy can support both—but only if promises become enforceable obligations before the next failure, not after it.

Sources