Want to Ensure AI Never Threatens Humanity? Make It Be Good

That's Yann LeCun's solution to AI doom — here's why it makes sense

Yann LeCun is an island of optimism in a sea of worry and doomsaying.

Perhaps more than any other high-profile researcher, he seems confident that we’ll solve the alignment problem before AI becomes too intelligent, that is, we’ll manage to ensure it remains under our control — or, at the very least, that it remains harmless to humans and humanity — even after it vastly surpasses our cognitive abilities.

Most others have paralyzing doubts about our ability to achieve such a feat. Not without motive. a human controlling a superintelligence is akin to a monkey controlling a human. Try to imagine that for a second. It is weird, to put it lightly, to picture the bizarre scene — you can immediately think of multiple ways to trick (or harm, if needed) a monkey with your superior intellect so that it ends up subjected to your desires and not the other way around.

How could a mentally inferior animal control you? Why would you ever accept being controlled by it?

A crazy world: Humans controlled by monkeys

An intelligent being can always outplay a less intelligent one, so long as it wants to, or even if it just doesn’t care. This gives way to LeCun’s proposal: Make it not want to do that. Make it want to stay under control by design — or the easier version (the one I will focus on here) let it be good for us.

The monkey analogy isn’t a good one because this wouldn’t happen between a monkey and a human — humans haven’t naturally evolved to be good to inferior beings by default. Humans have drives and motivations encoded biologically into our brains. We are objective-driven. That makes us pursue certain goals that would eventually diverge from those of monkeys. And they do.

What if humans were, instead, intelligent but purposeless creations, only behaving freely to the extent that our actions — and our greater ability and intellect — helped monkeys achieve their own goals? That’s a different thought experiment, isn’t it? It could work. But it’s hard to imagine a purposeless human — we can’t just decouple our intelligence from our goals; they go together.

The question is then, is AI distinct from humans in a way that allows for this design choice — universal goodness by default — to be successfully implemented? That’s LeCun’s proposal and that’s what we are going to find out today.

(This is a long one, but probably the most important essay I’ve ever written on AI risk and the explanation of why I have hope not as an irrational faith or due to denial toward the unprecedented but as a conscious act grounded in logic and reason.)