Should AI Slow Down? Tech Leaders and Politicians Clash Over Safety

Introduction

Should frontier artificial intelligence be developed at a slower pace? Anthropic CEO Dario Amodei has triggered a broad debate with an essay arguing that the frontier should be “paced.” His proposal has drawn responses from AI executives, lawmakers, and former government officials. Supporters say increasingly capable models need more time for safety testing and oversight before deployment. Opponents worry that slowing development could cost the United States its lead in competition with China.

The disagreement is not primarily about ending model training. It is about whether companies and governments can reduce the risks of loss of control, misuse, and excessive concentration of power without surrendering technological momentum.

Amodei’s three-part proposal

Amodei outlined three measures:

• Independent third-party evaluators embedded in company processes to verify safety commitments, assess practices, and report incidents;
• Coordination among frontier AI companies in democratic countries on common standards and limits;
• Global coordination between democratic and authoritarian governments where practical.

He stressed that pacing would not mean stopping model training or technical progress. Instead, companies would take more time to align and safeguard systems, while external evaluators would have an opportunity to verify their claims. Anthropic said it would unilaterally commit to the first step.

Broad support among AI leaders

OpenAI CEO Sam Altman said he agreed with the need to pace the frontier and supported independent evaluators. He also welcomed a federal framework that would establish consistent safety requirements for frontier AI. At the same time, Altman made clear that pacing did not mean stopping: progress should be slower than it otherwise might be.

Altman identified two especially serious failure modes: humans losing control of the future to AI, and a world in which power becomes excessively concentrated. He argued that avoiding both requires a narrow middle path. Google DeepMind co-founder Demis Hassabis said Amodei was pointing in the right direction, although the details still needed to be worked out. Elon Musk also endorsed the essay.

That alignment suggests a growing willingness among some industry leaders to accept stronger safety governance. Yet the preferred approach resembles delayed deployment, additional testing, and shared standards rather than a moratorium. Former White House AI and crypto official David Sacks said the companies could proceed, while questioning whether calls for regulation were entirely altruistic. Former Meta chief AI scientist Yann LeCun took the opposite view and mocked what he considered exaggerated safety warnings.

Washington’s security-versus-safety dilemma

Donald Trump rejected the idea of slowing AI, arguing that the United States must win the AI race and that existing presidential, regulatory, and criminal powers provide sufficient control. Vice President JD Vance was skeptical of frontier companies asking the government to regulate them, suggesting that such requests could become a “Trojan horse” for limiting competition.

House Speaker Mike Johnson offered a more conditional position. He said a moratorium was unacceptable because China could overtake the United States, but also acknowledged the need to make models safe. Other political voices emphasized the risks of doing too little. Bernie Sanders backed calls for a pause, while former Vice President Kamala Harris supported responsible slowing. Former Transportation Secretary Pete Buttigieg called for safeguards, rules, limits, and a kill switch for rogue AI.

Former FTC Chair Lina Khan argued that existing laws already give authorities power to address dangerous, untested, or defective products. In her view, debate over new legal regimes should not obscure current enforcement tools.

Why the debate matters

The argument reveals three important shifts. First, frontier AI safety has moved from a corporate ethics issue into national policy. Second, regulation is no longer framed only as unrestricted development versus a complete pause. Independent evaluation, incident reporting, and shared standards offer possible middle-ground mechanisms. Third, competition with China is shaping every discussion of restraint: any slowdown must explain how it can improve safety without simply giving away strategic advantage.

The difficult question is implementation. Corporate promises need credible verification, while government rules may be pulled in different directions by security competition and public protection. The next stage of the debate will therefore be less about who supports “slowing down” in principle and more about whether transparent, enforceable rules can survive pressure to move faster.

Source: The Verge AI