The Ultimate Tail Risk

As retirees, we spend a lot of time thinking about risks that may never happen. We diversify our investments, buy insurance and keep emergency reserves. We don’t need to know that something bad will happen before protecting ourselves against it.

I’ve been wondering whether we should think about AI the same way?

Some of the people building the world’s most powerful AI systems are loudly warning about the risks, even as their companies spend billions racing to make them more capable. I don’t blame them—if one company slows down, another may get ahead. But that’s also why I’m uncomfortable leaving the safeguards entirely to them.

So, indulging my fondness for Isaac Asimov, here’s my admittedly grand proposal: a 28th Amendment!

I’d establish three principles while we humans are still unquestionably in charge. 1) A machine isn’t a person and has no constitutional rights. 2) An AI can’t have the final say over taking a human life, starting a war or sending someone to prison. 3) And humans must always retain the ability to override or shut down an AI.

Congress could handle the details. Borrowing again from Asimov, I’d require advanced AI to have an independent safety filter it can’t disable. I’d keep it from having unmonitored access to banks, power grids, weapons and other critical systems. And I’d require an off switch if it’s hacked, corrupted or starts operating outside its boundaries.

I don’t know whether the worst predictions about AI will come true. I hope they don’t. But that seems beside the point. We don’t diversify because we know which investment will fail. We do it because we don’t know. Perhaps AI deserves the same humility.