Export Controls Shift Overton Window Toward Stricter AI Model Regulation

The overton window has shifted on AI regulation.

For the past few years there has been a range of debates around how we might regulate AI. There are many different camps on this topic ranging from “0 regulation” to “AI should just be fully stopped and shut down.” And there are about 20 different points on the spectrum in between.

But the prevailing mode we’ve been operating under was that AI in the US was largely not regulated for models directly, but rather risks would be addressed specifically as they come up as appropriate, and general best practices around AI safety were adhered to by most labs. As a result, we saw rapid model releases and leapfrogging of the technology constantly.

Some felt this might be an approach that could continue unabated, at least until we have clearer evidence of what novel risks weren’t being accounted for. Further, it hasn’t been obvious what kind of hammer the government would even use to regulate AI. Some people felt we need an FDA or FAA of AI, and many others considered that to be too heavy handed of an approach to keep innovation going.

But now, with a single swipe of the pen, export controls have shown us the way. No one reasonably thinks this will be the long term approach to regulating AI, but we’re getting a keen taste of what the implications of AI regulation will look like and what we can expect as a result.

In practice it means that models at higher capability levels will have to be thoroughly vetted and tested by the government before they can make it out to the market. This process may take weeks or months as the risks are often either subjective, or can only be experienced under highly contrived conditions; all of which requires tremendous back and forth between companies, researchers, academics, and the government. Even within the AI community, few agree on what a real risk is and isn’t. And just wait until politicians get involved more.

Further, when the models do finally get approved, they may be staggered to some small select set of companies or countries first, then expanded over time. They may also be used as a tool on some other negotiation entirely unrelated to AI.

All of this is to say that there’s now real risk that the level of rapid pace of releases and innovation that we’ve become accustomed to (as well as broad availability of this innovation) don’t happen in the same way. This can be partially mitigated if labs do major jumps in capability all at once, or if “minor” releases on the same base model don’t require the same level of review. But we now have more risk of slowdown than we did before.

And of course this means that other countries need to start considering their own path forward for controlling their destiny in AI. So far the US has been far ahead in AI, and thus gets the implicit “design win” for the entire AI stack. But this becomes more at risk in a world where everyone needs to own their own future in AI.

Hopefully we stay on the path we’ve been on where frontier intelligence still keeps streaming out, and open models remain open and available. But no matter what, the overton window has shifted. We know what is possible with regulation now and we can easily imagine the implications.

The overton window has shifted on AI regulation.

For the past few years there has been a range of debates around how we might regulate AI. There are many different camps on this topic ranging from “0 regulation” to “AI should just be fully stopped and shut down.” And there are about 20 different points on the spectrum in between.

But the prevailing mode we’ve been operating under was that AI in the US was largely not regulated for models directly, but rather risks would be addressed specifically as they come up as appropriate, and general best practices around AI safety were adhered to by most labs. As a result, we saw rapid model releases and leapfrogging of the technology constantly.

Some felt this might be an approach that could continue unabated, at least until we have clearer evidence of what novel risks weren’t being accounted for. Further, it hasn’t been obvious what kind of hammer the government would even use to regulate AI. Some people felt we need an FDA or FAA of AI, and many others considered that to be too heavy handed of an approach to keep innovation going.

But now, with a single swipe of the pen, export controls have shown us the way. No one reasonably thinks this will be the long term approach to regulating AI, but we’re getting a keen taste of what the implications of AI regulation will look like and what we can expect as a result.

In practice it means that models at higher capability levels will have to be thoroughly vetted and tested by the government before they can make it out to the market. This process may take weeks or months as the risks are often either subjective, or can only be experienced under highly contrived conditions; all of which requires tremendous back and forth between companies, researchers, academics, and the government. Even within the AI community, few agree on what a real risk is and isn’t. And just wait until politicians get involved more.

Further, when the models do finally get approved, they may be staggered to some small select set of companies or countries first, then expanded over time. They may also be used as a tool on some other negotiation entirely unrelated to AI.

All of this is to say that there’s now real risk that the level of rapid pace of releases and innovation that we’ve become accustomed to (as well as broad availability of this innovation) don’t happen in the same way. This can be partially mitigated if labs do major jumps in capability all at once, or if “minor” releases on the same base model don’t require the same level of review. But we now have more risk of slowdown than we did before.

And of course this means that other countries need to start considering their own path forward for controlling their destiny in AI. So far the US has been far ahead in AI, and thus gets the implicit “design win” for the entire AI stack. But this becomes more at risk in a world where everyone needs to own their own future in AI.

Hopefully we stay on the path we’ve been on where frontier intelligence still keeps streaming out, and open models remain open and available. But no matter what, the overton window has shifted. We know what is possible with regulation now and we can easily imagine the implications.