I think the debate over whether AI risk should be addressed via regulation or treaties is often oversimplified, and confused. These are not substitutes. They rely on overlapping underlying capacities and address different classes of problems, and both van benefit from certain classes of research.
David Kreuger, to pick on someone whose work I largely agree with, recently posted that “Stopping AI is easier than regulating it.” I largely agree with what he says. Unfortunately, I also think it is an example[1] of advocates for a cause creating fights where they're not needed, and in this case making the discussions around AI unfortunately more rather than less contentious, and less rather than more effective.
And the reason the fights are not needed is that different risks live at different levels, and different tools are effective in different ways.
Clearly, many of the risks and harms of AI should not be addressed internationally. There is little reason or ability to harmonize domestic laws on fraud, discrimination, or liability, which would be a distraction from either reducing the harms or addressing other risks. Existing laws should be adapted and applied, and new regulations should be formulated where needed. International oversight would be unwieldy and ineffective for even most treaty compliance efforts - as other treaties show, there is a mix of national and international oversight. But domestic regulation can create liability incentives, require or standardize audits, clarify rules, and provide enforcement mechanisms and resources. All of those are at least sometimes useful for treaties as well. When Kreuger says “the way I imagine stopping AI is actually a particular form of regulating AI,” he is not talking about the harms and risks regulation could address - though given what he has said elsewhere, he agrees that many of them are worth mitigating, even if they are not his highest priority. So it should be clear that treaties will not, cannot, and should not address most prosaic risks of AI systems and misuse.
By the converse argument, which he and others have made convincingly in the past, some harms of AI systems come from racing towards capability rather than prioritizing safety. These types of risk emerge from the dynamics of international markets and from great power competition. Obviously, these dynamics aren’t well addressed by domestic regulation on the part of any single actor. It is incomprehensible to talk about regulation alone to address those risks, just like it is tendentious to talk about using international treaties to mitigate other classes of risks and harms of AI systems.
Unfortunately, many discussions put “we need a global treaty to stop AI risks” in opposition to “domestic regulation is the only realistic path.” Not only do I think this is backwards, but I’ll argue that so is the related false dichotomy of industry self-regulation versus government rules. Industries that embrace safety welcome well-built regulation. Even in areas where they don’t have strict rules, airlines have national bodies that manage risk and accident reporting. (And the AI industry leaders often claim to be the same way, wanting national or international rules - just not any specific ones.)
So, to come to my unsurprising conclusion, we actually have several different plausibly positive and at least partially complementary approaches.
- Certain classes of research produce techniques like, evals, interpretability, human oversight approaches, control methods, and operationalizable definitions of specific risks. Some of these are dual use or net negative, but the parts that are useful are complementary to both regulation and treaties.
- Regulation needs operationalized definitions of risks, measurable standards, concrete goals, auditable procedures and oversight methods, and investigatory tools. Many of these are enabled by specific forms of technical or policy safety research.
- Treaties need shared definitions, clear goals, regulatory oversight and enforcement, credible verification, and both technical and regulatory methods to distinguish compliance from defection. Some of these are enabled by regulation, some by relevant research.
So we end up with a sort of triad, where research can enable measurement and definitions, and provide tools, regulation can force adoption and enforce usage of tools, and treaties can align incentives around defection dilemmas and provide common aims.
This doesn’t imply that most safety research is net risk-reducing, that most regulation is useful, or that most possible treaties will reduce risks. But it does say that they can be complementary. Some disagreements are substantive. But others are treating complementary approaches as mutually exclusive - and I think we should instead figure out common ground, which can make the fights about these issues both more concrete, and narrower.
- ^
yet another example
Discuss