A widening argument over AI risk
A new round of warnings from artificial-intelligence insiders has intensified a split in Silicon Valley over whether increasingly capable systems pose an existential danger or whether the most dramatic claims overstate what today’s technology can do. The dispute moved into view around a Goldman Sachs conference in San Francisco, where executives and investors were asked about the abrupt resignation of Anthropic researcher Jacob Coxon.
Coxon, who previously worked at OpenAI, said people building advanced AI were gambling with public safety and warned that future systems could become highly capable hackers. His departure followed earlier resignations from Anthropic and OpenAI connected to safety concerns. Evan Hubinger, an Anthropic team lead, also publicly assigned a greater-than-10% probability to AI killing humanity within a decade, according to the BBC’s account. Those are personal risk estimates, not measured forecasts.
The warnings drew blunt resistance. Grindr chief executive George Arison told the BBC that the statements reflected what he regarded as a dangerous worldview and said he had asked some engineers to stop using Anthropic technology. Nvidia chief Jensen Huang was also described by attendees as rejecting Coxon’s claims. Nvidia sells the chips used to train and operate many AI systems, giving it a direct commercial interest in continued industry growth.
Critics have also questioned whether extreme warnings may help leading AI companies justify high valuations or encourage regulation that smaller competitors would struggle to satisfy. That allegation remains contested. Coxon explicitly denied that his statements were marketing, while Anthropic chief executive Dario Amodei published an essay calling for slower model development and international regulation because he considers the risks serious.
Policy pressure grows
The argument is no longer confined to technology companies. US Senator Bernie Sanders has co-sponsored proposed legislation that would temporarily pause development of artificial superintelligence. President Donald Trump, by contrast, told reporters he was not concerned about human extinction from AI and emphasized competition with China. Those positions show the political distance between advocates of precaution and an administration focused on accelerating American capabilities.
Recent security disclosures have added urgency without resolving the larger debate. Anthropic said it detected and stopped actors attempting to use its systems in activities involving bioweapons and cyber-espionage. The company has also reported that one of its tools performed strongly on some cybersecurity tasks. Such findings can demonstrate misuse potential, but they do not by themselves establish that autonomous systems are close to escaping human control.
The immediate development is therefore a struggle over evidence, incentives and governance. Researchers raising alarms are asking policymakers to act before capabilities become unmanageable. Skeptics argue that extraordinary predictions demand stronger proof and may distort investment and regulation. With companies pursuing larger models and lawmakers considering limits, the disagreement is becoming a practical question about who sets the pace and what evidence should justify slowing it.
