That is the topic of my latest Free Press column. Excerpt: The key is to create some basic safeguards, but without stifling broader AI progress. To do so, we must defy the conventional wisdom about public oversight and instead trust the AI labs to be their own primary regulators.
My version of the proposal starts with defining a private not-for-profit body for AI regulation. An ideal body would draw some features from
[FINRA (the Financial Industry Regulatory Authority)]: a consortium of financial firms that examines the trade practices of each and makes recommendations, helping the federal Securities and Exchange Commission with oversight and regulation. The AI version would include the major labs and would be authorized and overseen by Washington, perhaps through the now-fledgling[Center for AI Standards and Innovation].This body would periodically audit major AI companies and their models, judging their conduct and safety. In the short run at least, much of this would be focused on issues of cybersecurity, and whether the new models created more cyber risk than they help to solve. If a company passed the audit, it would be exempted from standard liability law, at least provided that it had shown basic, reasonable care, as opposed to extreme or deliberate negligence. That would free the AI labs from the fear that courts might derail their business by granting huge awards to plaintiffs for ill-defined harms that could not reasonably have been prevented. And it would give the labs a strong incentive to meet the safety standards of this body.
It is reasonable to wonder whether such a body, composed of industry players, would issue fair and equitable judgments of safety. Maybe not. Yet there are many upsides and no better alternative.
For one thing, each company knows that a dangerous model from another company could cause a harmful incident and damage the prospects for the entire industry. Consider the Three Mile Island meltdown in 1979, which contributed significantly to the mothballing of the entire U.S. nuclear industry. Few people can name the company (Metropolitan Edison) behind the malfunctioning plant; the reputational penalty attached to the industry as a whole. Another incentive for safety is that the top companies do not want too much competition from lower-price, lower-quality upstarts. That too will induce those companies to support fairly tough standards, perhaps excessively tough in some cases. Still, we are choosing from imperfect alternatives. The concrete truth, whether we like it or not, is that there is far more expertise within the companies for judging AI safety than we can expect to find in the federal government anytime soon.
I am indebted to some ideas from Dean Ball, noting that his proposal is somewhat different. And here are some comments from Brendan McCord.