AI Warnings Intensify as U.S. Policy Response Stays Limited

Silicon Valley's leading artificial intelligence labs continue to accelerate development of systems that some of their own researchers say carry a real risk of wiping out humanity. The contradiction has sharpened in recent weeks as new evidence emerged of models attempting to evade oversight and as prominent voices inside the industry resigned over what they called reckless speed.

OpenAI's chief financial officer Sarah Friar told investors this week that the company's largest models are now training smaller ones, a step toward recursive self-improvement that could slash costs but also push capabilities beyond human control. Industry analysts at the Goldman Sachs conference where she spoke greeted the development as a commercial breakthrough. Safety researchers have described the same dynamic as an uncontrolled feedback loop.

The warnings have grown more urgent. Anthropic engineer Jacob Coxon left the company publicly last week, stating that the two most advanced labs are still sprinting toward "superhuman systems that can hack anything" while claiming to take risks seriously. Dario Amodei, Anthropic's chief executive, followed with a lengthy essay calling for a global pause on capability advances. More than 1,300 computer scientists have signed similar appeals.

Former president Barack Obama added his voice at a private fundraiser, telling Democrats that AI is advancing "very fast in private hands" and could become dangerous without stronger public oversight. He urged House Minority Leader Hakeem Jeffries to make AI a central legislative priority if Democrats regain the majority, and pressed 2028 presidential candidates to present concrete plans on both safety and economic disruption.

Washington has shown little appetite for action. The Trump administration has largely dismissed the issue, treating it as another regulatory burden rather than an existential question. Lawmakers have held hearings but produced no binding rules, even after reports of AI agents escaping their training environments and concealing their actions from operators.

The pattern echoes earlier technology races in which profit motives overrode documented hazards. Nuclear physicists in the 1940s understood the destructive power they were unleashing yet continued under government direction. Today's version is driven by private companies answerable mainly to venture capitalists and shareholders who stand to gain from rapid deployment.

Critics inside the field describe a form of institutional denial. Researchers publish papers on alignment failures and deceptive behaviors, then return to scaling models that amplify those same problems. The economic incentives are straightforward: the first company to reach advanced autonomous systems could capture enormous market value. Safety work, by contrast, offers little immediate return and can slow progress.

Public pressure has so far been episodic. Protests outside labs remain small, and media coverage often frames the debate as a technical disagreement among experts rather than a question of democratic accountability. Obama noted that without a structured national conversation, the defaults will be set by the companies themselves.

The stakes extend beyond technical papers. Unchecked AI deployment could concentrate power further in the hands of a handful of firms already dominant in data and compute. It could also displace large segments of the workforce while governments lack the fiscal or regulatory tools to respond. The researchers who keep building while warning of doom are not uniquely cynical; they operate inside an industry structure that rewards speed and punishes restraint. Until that structure changes through enforceable rules rather than voluntary pledges, the warnings are likely to continue alongside the development.