The AI policy discourse has picked up significantly this year. Although advocates for common-sense AI regulation breathed a sigh of relief when California Governor Gavin Newsom vetoed his state’s controversial SB 1047 in September, there are still hundreds of AI-focused bills circulating through U.S. statehouses, and it’s unclear how the federal government will approach AI regulation in the months and years to come.What is clear, though, is that we need a better, simpler, and, ultimately, reasonable way of thinking about this very important issue. However, it’s hard to discern what a reasonable policy position would be when there are so many extreme viewpoints and general confusion.Part of the problem is that the discourse has become a free-for-all proxy battle for airing everybody’s anxieties about artificial intelligence and tech more broadly. Even narrowly focused AI policy initiatives quickly become (virtual) shouting matches among well-funded organizations concerned with existential risk, industry groups concerned with AI’s impact on jobs and copyright, and policymakers trying to remedy the perception that they missed their window to effectively regulate social media. This can drown out legitimate concerns over AI policy overreach enabling regulatory capture and negatively affecting America’s economy, innovative spirit, and global competitiveness.But despite all the hubbub and competing interests, there actually is a reasonable policy position the United States can take: to focus on marginal risk and apply our regulatory energy there. It’s a simple approach that has already been proposed by a number of the top AI academics in the industry. And it’s worth understanding.Avoiding spurious AI regulationsMarginal risk refers to a new class of risk, introduced by a new technology, that requires a paradigmatic shift in policy to handle it. We saw this with the internet where, early on, new forms of computer threats (like internet worms) emerged. On the national security front, we had to shift our posture to deal with vulnerability asymmetry, where being more reliant on computer systems made us more vulnerable than other nations.Critically, focusing on marginal risks avoids spurious regulation, improving security by focusing on the right issues instead of wasting our efforts in ineffective policy.More broadly focused policies and tactics for governing information systems have been shaped over decades, with each new epoch raising concerns to which the industry needs to respond. And every computer system built now and going forward is already subject to those policies. Overall, this policy work—such as the work of the Internet Crimes Against Children Task Force, or extending lawful intercept to compute systems—has improved the regulatory landscape for coping with new technologies. Efforts to limit access to enabling hardware, such as putting export restrictions on computer chips, have had limited but likely positive outcomes for the United States.Still, other policies have failed in every attempt to employ them—and might even weaken security. These include approaches such as attempting to regulate math or adding backdoors to phones or cryptography. Absent a material change in marginal risk, these types of approaches will fail with AI, too.AI policy based on realityWhen it comes to regulating AI, we should draw from these learnings, not ignore them. We should only depart from the existing regulatory regime, and carve new ground, once we understand the marginal risks of AI relative to existing computer systems. Thus far, however, the discussion of marginal risks with AI is still very much based on research questions and hypotheticals. This is not just my perspective—it has been clearly stated by a highly respected collection of organized experts on the matter.Focusing on evidence-based policy (i.e., real, thorough research on marginal risk) is particularly important because the litany of concerns with AI has been quite divorced from reality. For example, many decried OpenAI’s GPT-2 model as too dangerous to release, and yet we now have multiple models—many times more powerful—that have been in production for years with minimal effects on the threat landscape. Just recently, there was rampant fear-mongering that deepfakes were going to skew the U.S. presidential election, but we haven’t seen a single meaningful example of that having happened.On the contrary, AI appears to be tremendously safe. In fact, we now have cars that drive safer than humans, computer systems that diagnose better than doctors, and countless advances in areas ranging from creative endeavors to biotechnology—all because of AI. In the end, we might conclude the best policy for human welfare is to invest aggressively in AI rather than to encumber it.So, until we’ve established a reasonable understanding of its marginal risk, let’s be sure to recognize the tremendous potential for AI to have a positive impact on the world—a promise upon which, to some degree, it is already delivering.This article originally appeared on originally appeared on Fortune.com.
