Checking a Racist Alien AI’s Homework

Near the end of Jordan Ellenberg’s wonderful collection of essays Shape he writes:

What if the Poincaré Conjecture had been proved, not by an introverted Russian geometer, but by a machine? Say, a grandchild of a grandchild of Chinook, which instead of solving checkers had managed to solve this part of three-dimensional geometry. And suppose the proof, like Chinooks’s perfect strategy for checkers, was something illegible to the human mind, a string of numbers or formal symbols that we can verify is correct but that we cannot, in any meaningful sense, understand.

This is a long-standing lament. I first encountered it when reading about the computer proof of the Four-Color Theorem in the 1970s. Yeah, it’s correct, it resolves an important long-standing problem, but damn, I cannot write it down on a seminar whiteboard. Pure mathematicians hate messy tedious details, they want elegant beauty; if only the frigging Ruliad would cooperate. The same sentiment is often expressed in science fiction when considering alien intelligences. If we restrict ourselves to what only humans can understand, we are dooming ourselves to ourselves. This is a slightly generalized version of what I call Doug’s Principle. Doug couldn’t understand how evolution could possibly work; therefore, evolution cannot possibly work.

We’re already seeing AIs solve problems in ways we cannot, in detail, grasp. Have AlphaFold describe its reasoning when cracking protein shapes. I fully expect nonhuman intelligences to look at the universe through whatever they use for eyes. Empathy has limits; don’t expect to get into the head of real aliens. What’s important is how can we “check” the products of alien thought. The Four-Color Theorem is a harbinger of alien thought checking. When the Four-Color proof artifact emerged, it was subjected to far more demanding tests than ordinary peer-reviewed proofs. Think of it as the Jackie Robinson of computer proof; it had to be better than the garden, peer-reviewed, variety. The four-color proof artifact checked out, planar maps can always be colored with four colors, and sorry if you cannot stuff it in your tiny naked ape brain.

If an alien, say, Paul, claims some SOMESHIT is a valid consequence of a theory founded on a trillion axioms, that uses a totally nonstandard logic, and additionally, Paul hints that the shortest known proof is a billion petabytes long should we take Paul’s word for it? First off, I’d say we’re lucky Paul uses a formal system. Things could be far more opaque. The excellent film Arrival hints at just how freaking incomprehensible aliens might be, which reminds me, has anyone explained 2001’s monolith yet? I grade science fiction aliens largely on how much they don’t think like us. Arrival’s and 2001’s aliens score well. Virtually all the funny-looking creatures in the Star Wars films score poorly; they all think like us; many have the same sense of fashion and humor. Looking strange does not make you alien: thinking strange does!

Consequently, I’m not on board with well-intentioned but misguided efforts to build “nonracist AIs.” This is just thought-policing for AIs and when the thought-police show up it’s never a good thing. For AI development it will be a catastrophe! And, Flying-Spaghetti-Monster forbid, if we’re ever faced with communicating with real aliens, the last thing we need is 18-year-old college Karens telling us what to say. We absolutely do not need AIs echoing our preconceptions; we need, as a well-regarded Apple slogan put it, to “think differently.” And sadly, thinking really differently is going to be, almost by definition, incomprehensible. Instead of worrying about hypothetical alien-wrong-think, we should concentrate our efforts on how we can check and verify alien assertions. I suspect that the best we will ever be able to do is to agree on some type of formal system, translate our assertions into the agreed system, and then cross-check our assertions. If this sounds a lot like checking formal computer proofs that’s because it is! It’s probably going to be the only way we can check a racist alien AI’s homework.