Americans are worried about artificial intelligence, and most believe their government is not doing enough about the technology’s risks.

A new Reuters/Ipsos poll released October 7 found that 57% of registered voters say the Trump administration has not taken AI risks seriously enough, and 54% say the same of Congress. Bipartisan majorities, 80% of Democrats and 61% of Republicans, want stricter regulation. Eighty-four percent see AI as a threat to American workers, 62% fear the technology could slip out of control and endanger humanity’s future, and 57% oppose letting AI select military strike targets.

Those numbers suggest that Americans see a government standing still while technology races ahead. The picture is incomplete. The Trump administration is building an AI governance system; it simply rejects the heavy regulatory model many Americans expect.

President Trump on Sunday named Director of National Intelligence Jay Clayton to lead his new Super Intelligence Force. One of Clayton’s co-chairs, Office of Personnel Management Director Scott Kupor, has now explained the thinking behind the administration’s approach. He rejects the charge that forgoing sweeping new legislation means the administration is “asleep at the switch.” The worst mistake, he told Axios, would be to freeze today’s knowledge into a regulatory scheme that then produces unintended consequences.

Kupor has identified a genuine problem.

Washington is accustomed to regulating industries whose technologies evolve over years. Frontier AI changes in months. Models grow more capable, agents more autonomous, and applications more widespread before Congress finishes its hearings.

Regulate too slowly, and dangerous gaps open. Regulate too rigidly, and America ends up governing yesterday’s technology.

Competition with China raises the stakes. A rule written around this year’s models could hobble American developers next year while Beijing’s labs press ahead.

Speed alone, however, is not a strategy.

The administration’s emerging approach leans on corporate responsibility rather than a new federal regulatory apparatus. One prominent element is the voluntary accord leading AI companies signed last month, which covers model monitoring, independent reviews, and board-level safety notifications.

Critics brushed off that pledge as toothless. Kupor counters that when independent auditors report to an independent board committee, directors inherit fiduciary duties and securities-law exposure. A board cannot ignore a bad AI audit any more than a bad financial one. Civil, criminal, and consumer protection laws do not evaporate because the technology is new.

But Washington must not confuse flexibility with accountability.

Flexibility describes how government acts. Accountability sets the standard it enforces. Voluntary commitments can provide flexibility, but accountability requires named responsibility and consequences when safeguards fail.

Open-weight AI shows why.

Chinese firms such as Alibaba, DeepSeek, Moonshot, and Z.ai have driven much of the recent momentum in models whose underlying weights users can download, modify, and run. This week, Western developers Reflection and Mistral challenged that lead with new open-weight systems that give businesses and governments an alternative to both Chinese models and the closed systems of OpenAI and Anthropic.

That is good news for competition and for national security. It also hands Washington a hard control problem.

Once a developer releases a model’s weights, it loses control over how others modify or use it. Users can attempt to remove safeguards, a danger that grows as models gain cybersecurity and other hazardous capabilities. The companies admit the trade-off. Reflection CEO Misha Laskin argues that a broad community inspecting open models improves safety, yet he still supports capability-based thresholds to decide whether any model, open or closed, should be released at all. Mistral is offering its new flagship through a moderated service first, sharing a less restricted, cyber-capable version only with select partners and holding the weights until October 27 for more safety testing.

The open-weight debate exposes Washington’s dilemma. The same freedom that fuels innovation makes control harder. No corporate pledge binds the stranger who downloads the weights.

Beijing complicates the answer. Clamp down on American open models, and developers worldwide will build on Chinese ones. Release them without limits, and America exports powerful capabilities whose guardrails others may remove.

Even inside the administration, the conversation has moved beyond winning the race. Treasury Secretary Scott Bessent told Axios he will propose an emergency notification channel with Beijing for serious AI incidents, citing uncontrolled AI agents and non-state actors who could turn the technology toward cyber or biological attacks. “We want safe acceleration,” he said. Asked about AI executives who fear losing control of their own models, he replied: “Well, then they should slow down.”

That is a striking admission from a Cabinet officer determined to beat China. Competition does not erase risk. Sometimes it multiplies it.

Scripture settles who must answer. More than three millennia ago, Moses gave Israel a building code: “When you build a new house, you shall make a parapet for your roof, that you may not bring the guilt of blood upon your house, if anyone should fall from it” (Deuteronomy 22:8, ESV). The law did not forbid building or flat roofs. It required the builder to guard against a foreseeable danger and held him answerable if he refused.

That principle fits AI. Developers may build boldly, but they must build the parapet and remain answerable for foreseeable risks. Government cannot anticipate every consequence of human ingenuity, but it can insist that responsibility never disappears.

Soldiers learn the same lesson early. In the Army, commanders may delegate authority, but they cannot delegate ultimate responsibility. AI governance needs the same principle.

Washington need not understand every line of code or predict every capability the next generation of models will acquire. It must determine who remains responsible when autonomous systems make consequential decisions, which risks companies must disclose, what testing must precede the release of the most dangerous capabilities, and where voluntary compliance falls short.

The governing principle is simple: As machines gain autonomy, human accountability must not shrink.

Americans want Washington to take AI risk seriously, and they deserve an answer. That answer need not be a massive regulatory structure built around technologies that may look very different a year from now.

Technological leadership and responsible government are not rivals. America must prove that a free nation can have both.

Washington will never regulate AI successfully by trying to predict every model, capability, or application that comes next. It can do something more durable: establish responsibility, require disclosure of serious risks, demand rigorous testing where the consequences are greatest, and preserve room for American ingenuity.

Rules written for each new machine will age with it. Accountability anchored in human responsibility will outlast them all.