Commentary: The United States is once again asking the technology industry to police itself. That may sound pragmatic in a fast-moving field, but it is the wrong default for artificial intelligence, a technology whose harms can spread faster than regulators can respond.
At a September meeting with technology executives, President Donald Trump backed a voluntary approach to AI oversight, while House Speaker Mike Johnson called for greater transparency and external auditing without insisting that companies be compelled to comply. The emerging message is familiar: innovation first, regulation later—or perhaps regulation never.
Our view is straightforward: voluntary commitments can be useful, but they cannot substitute for enforceable public rules. A company may voluntarily publish safety evaluations or report incidents. It may also decide that doing so puts it at a competitive disadvantage, reveals proprietary information or exposes it to litigation. Once that happens, the responsible firm is punished for being transparent while the least responsible competitor gains an advantage.
The argument for restraint
The case against heavy-handed regulation is not frivolous. AI systems evolve rapidly, and rules written today could freeze outdated assumptions into law. Smaller companies may be unable to absorb compliance costs designed with the largest laboratories in mind. Broad requirements could also push investment and talent abroad, reducing American competitiveness in a sector with major economic and scientific potential.
There is a legitimate distinction between regulating demonstrable risks and trying to control hypothetical ones. Policymakers should not demand approval for every software update, nor should they impose identical obligations on a classroom chatbot and a system used in critical infrastructure. Targeted, risk-based rules would be more defensible than a sweeping bureaucracy.
Industry-led standards can also move faster than legislation. Engineers and independent researchers often understand emerging technical failures before lawmakers do. A voluntary framework may therefore be a sensible first layer—provided it is treated as a floor for experimentation, not the final architecture of accountability.
Why voluntary promises fall short
The problem is that the incentives are not neutral. Companies benefit from deploying systems quickly, collecting more user data and minimizing disclosures that could alarm customers or invite scrutiny. The Federal Trade Commission has reportedly opened an industry-wide probe into Anthropic, OpenAI and other AI laboratories over potential consumer harms, a sign that voluntary assurances have not eliminated the need for public enforcement.
Nor are the risks limited to spectacular future scenarios. Consumers can face deceptive outputs, privacy violations, discriminatory decisions and financial losses now. Workers may be evaluated or replaced through systems they cannot inspect. Developers may deploy autonomous tools that make errors at scale. In each case, the affected person usually has less information and bargaining power than the company.
Transparency without consequences is disclosure, not accountability.
That principle does not require Washington to micromanage artificial intelligence. It requires basic obligations: truthful claims about capabilities, testing proportionate to risk, prompt reporting of serious incidents, protection for whistleblowers and meaningful remedies when systems cause harm. Independent audits should be genuinely independent, with enough access to test the claims being made. Regulators should have authority to penalize concealment and reckless deployment.
A narrower, stronger compromise
The best alternative to both laissez-faire and blanket prohibition is a focused legal framework. Low-risk applications should face light requirements. High-impact systems—those affecting employment, credit, health, education, policing or essential services—should meet higher standards before and after deployment. Rules should be revisited regularly, with sunset clauses where appropriate, and should protect open research rather than granting incumbents a permanent advantage.
Industry expertise belongs in that process. So do civil-society groups, affected workers, consumers and independent scientists. But participation is not a veto. Democratic governments are accountable to the public in a way private companies are not.
America can still lead in AI without treating guardrails as an enemy of innovation. In fact, predictable rules may help responsible firms compete by preventing a race to the bottom. The question is not whether government should control every algorithm. It is whether companies whose systems shape lives should be trusted to define, investigate and punish their own failures.
They should not. Voluntary standards can begin the conversation. Only enforceable accountability can finish it.