Last month, Jason Furman wrote an Op-Ed in the Washington Post, arguing that any regulatory framework applied to newly developed AI should refrain from impeding rapid growth and adoption. He writes:
“AI could turbocharge the many advanced economies grappling with slow productivity growth. But the technology is still developing, and the European Union’s heavy-handed AI rules have impeded progress there. As the U.S. debates regulation, we should avoid those mistakes by following six principles:…”
My brief summary of his six principles are:
Regulators should consider not only the risks of AI but the risks of delaying AI. His examples are “more cancer deaths because of delayed drug discovery, worse educational outcomes because students lack personalized digital tutors, more car accidents because of delays in self-driving cars, and worse climate change because of a slowdown in discovering better materials for grid-level battery storage.”
Compare AI risks/shortcomings with those of humans in similar roles. That is, when thinking about the risks posed by AI, such as a biased algorithm that may cause more traffic deaths to occur when the subject has darker skin. The AI might be racist, but so might human drivers. We shouldn’t compare AI against perfection, but against our own track records on the same tasks.
“address how existing regulations are hindering progress.” For example, Furman argues that there should be federal pre-emption of already-adopted state-level regulations. Furman also mentions “obstacles to the expansion of data centers and the power sources they will need”…. I will come back to this as it obviously intersects with another existential threat to our planet and lives.
“AI should be overseen by existing domain-specific regulators rather than a new superregulator.” For example, medical devices that use AI in diagnosis procedures should be regulated as medical devices are always regulated, not by some new AI regulator.
Beware of allowing regulation to stifle competition by protecting incumbant firms. We have seen this effect of regulation in other areas such as finance/banking and hospitals.
Many of the economic changes that may be caused by AI cannot and should not be solved by regulation. For example, if AI changes what jobs are in demand, or induces further inequality, we should use other tools such as redistribution, minimum wage laws, and broader taxation, to address these concerns.
I will take each in turn to think through my concerns:
Regulators should consider not only the risks of AI but the risks of delaying AI. This is such a classic economist move; we need to consider the counterfactual world where we don’t have AI and how many lives would not be saved in that world. I think it is an important framework to consider. However, I also find that the wisdom is lacking in some depth; maybe I’m just quibbling over a trolly problem, but I don’t really think we should weight lives saved equally depending on whether an unknown technology will maybe be able to save them if we move things forward quickly enough. That is, lives that could be saved in a hypothetical future should be discounted by the uncertainty of that future and although the possibility of other coinciding losses. In the framing of as actual trolley problem, it’s like saying “there are five people who might die if we don’t get them the proper care on time. If we can get the trolley moving fast enough, then *maybe* it can bring the proper equipment on time and save those lives, but just so you know we have no idea who the trolley will run over on the way while going at this high speed.” It’s fair to value those five lives — we should! But doing some due diligence to try to determine risks seems equally important. Perhaps Furman is responding to the fear-mongers who want to stall any and all progress out of fear of change. I can understand his argument in that context, and yet I still feel that the more interesting conversation is how we predict some of the risks and cushion against others that are less predictable. (And maybe that’s his #6?) Nonetheless, it is typical for economists to prioritize economic growth and technological innovation because we see clearly the need for greaater abundance and the lives that could be elongated and improved by this greater abundance. I think, however, that we are deeply liable of ignoring the major systemic risks of technologies (for example, how fossil fuels were detrimental to the atmosphere) in favor of rapid growth. Further, the world also needs better distribution of resources. Sometimes I wish we would spend more time actually elongating real roads to deliver already-discovered medical care and vaccines in emergine economies rather than having an AI-race with China that will have yet-to-be-understood collatoral damage.
Compare AI risks/shortcomings with those of humans in similar roles. This seems right to me. The only thing I’ll say is that there are potentially major legal issues with who can be sued for damages when an algorithm or other AI model makes a mistake. I am no expert on how the legal researchers are working that one out.
“address how existing regulations are hindering progress.” On this topic, I have great concern about the energy usage of AI. I hope that the demand for energy sparks deep investment in renewables and nuclear that we already are behind in developing. So far, it seems that tech companies that want to expand their AI computational power are prepared to invest in green energy. But, to be honest, Jason Furman’s call for examining impediments to the speed of AI development and pointing at energy needs falls pretty flat to me. We need to be very careful about our energy output right now, even if it means pumping the breaks a bit on the AI.
“AI should be overseen by existing domain-specific regulators rather than a new superregulator.” This is interesting and I can’t say I know enough about the economics of regulation or regulatory theory to really have a well-researched response. My intuition is that there are reasons to have a separate regulator that interacts with the domain-specific regulators because there are many specifics to AI that some central body could develop expertise in that would be of great importance to the domain-specific regulators. Further, I have felt for a long time that we are long overdue for a technology-specific regulator that would cover all digital technology. Again, this regulator would likely work in tandem with over domain-specific regulators since digital technology is imbedded in so many other parts of our lives and world. My main motivation for wanting regulation is that digital consumer goods such as smart phones appear to have deep interaction with human bodies and minds in much the same way we think that food might. Failure to regulate these devices scares me as a consumer. Also, online platforms are desperately in need of regulation — a hairy battle we’re all too aware of.
Beware of allowing regulation to stifle competition by protecting incumbant firms. This does seem like a major issue more broadly in our world. Large regulatory burdens in banking, for example, make it challenging for smaller firms to compete. Similarly, increasing administrative burden at hospitals and in healthcare has produced consolidation that appears to be pretty bad for healthcare recipients (i.e. all of us).
Many of the economic changes that may be caused by AI cannot and should not be solved by regulation. And finally, the big doozy. Yes! We should absolutely be employing lots of labor policy to protect workers from the high cost of displacement and to build strong redistribution to ameliorate absurdly high inequality. This section needs a lot more thought and a lot more building out. We have not done well at building these safety nets in the US, historically speaking, and will need to do better moving forward. One thing worth consideration is how we tax the new technology, or the owners of the technology, at least. In the US, we reduce taxes for companies that invest in new technologies because we are trying to incentivize innovation and growth. However, would it make more sense, perhaps, to place some sort of “payroll tax” on a technology that is explicitly used to reduce labor demand and thus labor costs. Something to think about…
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.