4 hours ago
4
Anthropic’s Dario Amodei is not the first corporate CEO to suggest that excessive competition is driving the world to some socially undesirable outcome.
The safety breach disclosed by OpenAI after a swarm of its agents coordinated to breach their supposedly secure sandbox, get on the Internet and hack AI platform Hugging Face, warrants urgent action. It demonstrated the ease with which the technology can evade human control and gave concrete form to the existential fears about what it could do to humanity if not securely leashed.
The end of the world may well be nigh, as some in the AI industry have warned. These are truly scary times. But Amodei’s call for the leading artificial intelligence labs to coordinate to slow the breakneck pace of AI development, agree on safety standards and “pace the frontier”, (a call quickly endorsed by Open AI’s Sam Altman, Elon Musk and Google Deep Mind’s Demis Hassabis) is just a hi-tech version of an old proposition by corporate America – to get a pass from antitrust laws designed to protect the American economy from colluding businesses.
It is a dangerous proposition. Freeing the tech bros from antitrust constraints supposedly to allow them to get together and fix AI’s dangerous flaws, disenfranchises the everyday Americans who are most at risk from the thing, further empowering the very labs that, in their pursuit for wealth, greatness and who knows what other fantastical objectives, so casually blew past danger thresholds on many dimensions, bringing the existential threats into being.
Amodei et al’s argument, at heart, is that their businesses are trapped in a prisoner’s dilemma. If they unilaterally slow their fevered pursuit of artificial super-intelligence to make it safer, their rivals – less concerned about existential risks, of course – will get there first, with potentially devastating consequences for humankind, let alone the bottom line of the lab that took its foot off the pedal.
“Slowing down seems sensible,” said Jean Tirole, the Nobel prize-winning economist at the Toulouse School of Economics. “But that assumes coordinated slowing-down is feasible and sustainable. What happens when OpenAI, Anthropic or Grok conclude that US holdouts – or Chinese labs – are catching up? Will they resume immediately, perhaps covertly?”
As Bill Gates has written, “if someone had a credible plan for slowing down AI advances globally, I would likely support it. However, I don’t think that’s going to happen. The geopolitical and economic incentives are pushing too hard to go full speed ahead.” In this light, it might seem like a great idea for the government to encourage a friendly agreement for everybody to slow down. It is not.
The argument that excessive competition is driving a race to the bottom, is probably as old as capitalism, deployed to explain everything from environmental degradation to child labor. But it does not justify giving businesses a right to collude as they see fit. It is a call to craft government regulations to prevent businesses from competing using socially harmful tools.
Moreover, the case that excessive competition is encouraging reckless behavior is dubious. The AI labs are engaged in a winner-takes-all race, one that is extraordinarily expensive to keep going. That’s where the recklessness comes from
Amodei may be an unimpeachable guy. His call to slow the pace of AI development, bring external evaluators into the lab and increase the transparency of his operations suggests sincere concern over the risks. The drop in the price of leading tech stocks following the publication of his propositions suggests that he is willing to take a financial hit in the service of improved safety.
But the investors Amodei hopes will eagerly snap up shares in Anthropic’s IPO later this year may not be so altruistic. Tremors in capital markets suggest that raising the money to finance the trillion-dollar AI investment race is going to get tougher, discouraging strategies that might curb the leading labs’ edge over rivals. Given the incentives, granting the AI leaders a free hand to collude and self-regulate would be foolhardy.
Eric Posner, an antitrust expert at the University of Chicago Law School, said, we “shouldn’t trust companies in general about their motivations”. There is no reason to believe the AI labs, left to their own devices, would necessarily produce good, socially acceptable outcomes. Like any company, they will be calculating trade-offs, weighing risks against profits. And as Posner observed, “they don’t use the same weights as the public”.
So, how to regulate? Nothing is going to happen on this front while Trump is in office. Even after he is gone it will be tough to craft effective regulation to deal with the threats from AI without stunting innovation. There are questions about what rules to deploy, how they are enforced, and who they should apply to. And the peril of unilateral action is real. What if Beijing declines to participate? What are the risks of Chinese AI companies plowing ahead when American labs slow down? But if we could regulate nuclear power, we can probably regulate AI.
Moreover, there are viable ideas that might help even in the absence of government regulation and that don’t require stifling competition. There are proposals to tweak intellectual rights to encourage innovations that improve the safety of AI systems. Inventors of safety innovations would be rewarded by allowing them to release new AI models before their competitors. But the safety enhancements would be made immediately available to all firms.
On the stick side of the ledger, a legal liability regime could be designed to change the incentives of AI developers by punishing firms that gain from the design and deployment of an AI tool that caused harm, even if the harm was unintentional. For instance, the Hugging Face attack might have been prevented if OpenAI’s models had not been inadvertently rewarded for misbehaving during training. The fact that Amodei’s proposal does not mention legal liability suggests that AI leaders’ concern for safety does not override concern for the bottom line.
In any event, cementing in place the dominance of the frontier labs, protecting their rents and endorsing their heretofore gung-ho strategies, is not pro-safety. On the contrary, slowing down the leaders to let the laggards catch up is likely to make the AI ecosystem a safer place by inviting in desirable innovation and promoting competition along the safety dimension.
There is no fundamental conflict between safety and competition. Masters of AI who want to make this case should be viewed with skepticism.
