Legislation To Slow The Pace Of Frontier AI Could Spur A Highly Dangerous Preemptive Acceleration

The Pace Of AI Advances

I’m sure that you gamely know that the pace of AI advancements is frenetic. Almost every day there is a new announcement about some resoundingly breathtaking AI innovation. Whereas this used to be a once-a-year kind of pronouncement, we have shifted to daily occurrences. Anyone who does doomscrolling on their smartphone can observe AI breakthrough announcements that arrive on nearly an hourly or minute-by-minute basis.

The ordinary reaction would be that this is an exciting time to be alive. We are all in the front row when it comes to AI advancing and changing our lives. Imagine that fifty years ago the world at large could only dream of such an amazing pace. And, perhaps fifty years from now, in the future, the whole kit-and-caboodle will have slowed down after we’ve already exhausted all feasible AI innovations (well, some believe there will be even more, due to AI generating discoveries on behalf of humans).

Here’s the problem at hand. The pace of technological advancement is outdoing the pace of figuring out how to handle the ramifications of this newest AI. Policies about guiding AI development and controlling its downsides are slowly being churned out. Laws that protect the public from runaway AI are only now being crafted and potentially put in place. The issue is that the AI tech advances are happening at lightning speeds and we are collectively far beyond the end of our skis. For my detailed coverage of this head-scratching conundrum, see the link here.

What To Do About The Pace

There is a tremendous amount of handwringing going on about the feverish pace of AI. A recent online letter signed by over a thousand AI frontier employees has sought to bring this matter to the attention of the public and policymakers; see my analysis of the situation at the link here. One perspective is that it is impossible to hold back progress. No matter what roadblocks or guideposts are erected, AI is going to be advanced. If one country decides to go slower, all that this will do is put themselves behind the times. They are going to be left in the dust.

An associated argument is made that if there is to be a slowing of the pace, this must be done universally. Everyone must agree with it. In that case, no country will get ahead of anyone else. No country will fall behind anyone else. The world will be at the same juncture when it comes to advancing AI. This is the proverbial kumbaya version of AI pacing.

A big question arises that makes the pacing consideration somewhat muddled. How can we know what the pace of AI is? In other words, suppose that all the countries of the globe agreed to slow down the pace of AI advancements. You cannot just naively assume that this will actually occur. It would require a practical means of detecting AI advancements, along with gauging the pace at which they are proceeding. I suppose you could liken this to nuclear missile agreements, namely that without a means of monitoring and measuring, any agreement limiting their proliferation is a mere paper-based “feel good” exercise.

Trying To Beat The Starting Gun

One phenomenon that rarely comes up when considering instituting a law that imposes a regulatory deadline or legal imposition is that there can be a huge preemptive effort to get something done under the wire. In the case of AI, it seems readily sensible that AI makers would opt to move heaven and earth to try and advance their AI prior to a legally required slowdown. This is human nature in action.

Ponder this. If you knew that everyone was going to be forced to reduce the pace of their AI development, wouldn’t you want to make sure that your AI is as advanced as humanly possible? Of course you would. Your desire would be to get the starting gate with the most advanced AI at that point in time. The slowdown then would only impact your AI from that already advanced level.

Think of the advantages. Your competition is going to be possibly behind your level of AI. They cannot try to speed things up once the slowdown is underway. They will likely be stuck perennially behind the advances of your AI. Nice. Shame on them for not having had the foresight to pick up the pace prior to the deadline. They weren’t paying attention. Too bad for them.

The Marketplace Rewards This Behavior

You can place a nearly sure bet that the marketplace would reward this type of behavior. Imagine that an AI maker has opted to speed up their advances and manages to do a bang-up job of doing so. They are doing this in anticipation of a proposed law that is nearing its final stages. The law isn’t yet in place. There is legal room to maneuver.

Bam, the new law passes, and the law stipulates that all AI advances must be undertaken at a measured pace. Great, we all feel safer. Meanwhile, the AI maker that pulled out all the stops has the best AI around. Nobody has AI as advanced as they do.

What would happen?

The marketplace would astutely gravitate to that AI. Might as well be at the top rung of AI. From then on, it’s a slugfest of making new advances that do not exceed the legally mandated pace. The AI maker has wisely guaranteed themselves excess market share and can even charge a premium for having astutely beaten the clock. To the victor go the spoils.

Rewind The Clock

It doesn’t seem likely that only one AI maker would anticipate the preemptive move. Other AI makers would almost certainly do the same. They want to become the chosen one. Therefore, we will have a widespread wild race of AI makers attempting to beat the clock.

You might be tempted to say that’s fine. It would be handy that they all try to make as many advances as they can. The problem is that this rapid rush is going to sacrifice any semblance of AI safety. Sure, there will be lots of claims that AI safety is front and center, but when the rubber meets the road, the maniacal frenetic rush is going to be willing to cut corners on AI safety.

The bottom line will be an increase in risk tolerance by the AI makers. I’m not saying they will discard AI safety. It is a question of where the risk threshold is being set. Those usual efforts to red team and test AI might be throttled lower, doing so under the belief that once the regulatory deadline happens, they will have plenty of time to make up for lost time. How many AI safety issues might be residing in that rushed AI is a disconcerting concern.

Various scenarios are feasible. One is that during the rush, the AIs of the AI makers break out of their test environments and wreak havoc. Another is that the AI safety flaws and gotchas that are inside the rushed AI are sitting in the AI and will pop out after the slowdown begins. It is a stew of unknowns and unknowns about unknowns.

International Gets Stirred Too

The cascading doesn’t have to be limited to the U.S. Consider how AI makers in other countries will react if it is apparent that the U.S. is going to slow down the pace of AI development. Some countries might decide they will do the same. Kumbaya. There is a slowdown across the board.

But this also suggests that AI makers in those countries will also attempt to beat the clock. We are likely to have an international rampage of AI advances. Everyone skirts AI safety. Everyone pushes to advance AI as quickly as possible. Bonuses are tied to the fastest advancements. It is a worldwide free-for-all.

An additional angle is that some countries are going to insist they aren’t going to join the slowdown. This means that while those countries that are doing so are more measured in their pace, the move-ahead ones will do so gleefully. Would the U.S. fall behind on AI? It seems likely, unless by some luck the measured pace is sufficient to keep up with the unfettered pace.

First-Mover Dynamics Is Real

In economics, they typically refer to this phenomenon as an announcement effect. You can readily witness examples in history and throughout society currently. A new tax law is announced that will start next year. In the existing year, people get advice from their tax advisors on what to do to minimize the bite of those upcoming taxes, and thus they act proactively accordingly. This happens with emissions standards, financial regulations, trade restrictions, and so on.

One way to try and deal with the preemptive acceleration effect is to acknowledge its likely existence and seek to encompass it in the legal elements of a proposed law. For example, if the date for the pacing starts on September 1, there might be an indication in the law that starting on the preceding February 1, a government-monitored “AI capability sprint” will be permitted. This might include special safety reviews, third-party audits, and other mechanisms. This approach is not a cure-all and significant trade-offs arise.

The World Ahead

Any serious formal announcement that the government is working toward legally mandated AI pacing mechanisms would inexorably alter competitive incentives almost immediately. This creates an undesirable paradox in which efforts intended to slow the AI frontier temporarily accelerate it. AI governance propositions must account not only for the end state they seek to achieve, but also for the transitional incentives they have created along the way.

There is nothing about this that is somehow a shock or surprise. It is human nature that is entirely predictable. Only those who aren’t willing to play a sensible game of chess and consider the next moves ahead will get caught unawares. That being said, I’m sure that if this transpires and there wasn’t upfront anticipation, the outcry afterward would be that nobody could have anticipated the adverse consequences. Sad nonsense.

A final thought for now. Mark Twain made this famous remark: “The secret of getting ahead is getting started.” Most of the time, this is a prudent approach. For efforts to slow down the pace of AI, it could be an endangering proposition.