FEW EVENTS have brought together Bernie Sanders and Steve Bannon, political firebrands from opposite ends of the spectrum, in common cause. But the Pro-Human Assembly in Washington, DC on September 15th, a gathering of people alarmed by the rapid advance of artificial intelligence, managed it.
Both men warned of the unchecked power of tech “oligarchs” and called for urgent legislation to slow down the technology.
Three months later Kathy Hochul, New York’s governor, signed a law demanding faster disclosure and imposing larger fines. The Pro-Human Assembly and Mr Trump’s flurry of posts on the subject came less than a week after the public resignation of Jacob Coxon, a researcher at Anthropic, an AI firm. In general, polls show that Americans disagree: eight in ten say they support regulation, even if it slows down innovation. The trickier question is what that regulation should be. To date most activity has been in states. Both laws would have been bolder, with higher penalties and exposing firms to more liability, were it not for opposition from the likes of Leading the Future, a super PAC whose supporters include Greg Brockman, a founder of OpenAI, another firm. Clear gaps in state laws remain. If proponents of federal regulation want to secure bipartisan support, they may have to compromise in two areas, argues Brad Carson, the co-founder of Public First Action, which advocates for AI regulation and is backed by Anthropic. One is how easy it is for the government to shut down an errant model. A second is whether bills should cover only existential risks, such as AI models’ ability to create biological weapons, or also include issues such as liability and children’s use of chatbots.
Another top scientist at the firm put the probability of extinction at more than 10%. Last September Gavin Newsom, California’s governor, signed a law requiring AI firms to report safety incidents within 15 days, on pain of up to $1m in penalties.
Mr Coxon said that he and his colleagues believed that AI could “kill us all” by the end of the decade, a message that he promptly delivered across prime-time television. California’s law did not apply to the now-notorious Hugging Face incident, when OpenAI’s agents went rogue, because it occurred in an “evaluation phase” designed to elicit bad behaviour.
On paper, many AI firms back some sort of regulation. Google has suggested a body setting industry standards, modelled on the Financial Industry Regulatory Authority. But no lab has thrown its heft behind any of the big state or federal proposals in full, with the exception of Anthropic’s support for California’s law.
Anthropic, the most enthusiastic, has cited the Federal Aviation Administration and Food and Drug Administration as models; in the meantime, it has said it would embed third-party evaluators in its lab, a practice that competitors claim they will follow.
