Don’t we already have a good measure of intelligence by which to test artificial intelligence (AI) systems—IQ?
The short answer is: no.
The longer answer comprises of two reasons, as posed by John Haugeland in Artificial Intelligence:The Very Idea, for why performance on an IQ test is not a good measurement for intelligence in AI:
1.IQ tests measure degree of intelligence on the assumption that the entity being tested is already intelligent.
In other words, for the IQ test to measure degree of intelligence, there needs to be something to measure.
IQ tests operate on the assumption that the test-take is intelligent, but for AI, the assumption is precisely what is in question: we need to know whether it even makes sense to attribute intelligence to AI in the first place.
2. IQ tests are designed for people.
It might be that in people, the ability to solve a logic puzzle correlates to intelligence, but that correlation may not hold for machines.
Machines could do many tasks, which as Minsky states would require intelligence if done by humans, but does that make those machines intelligent?
As Haugeland writes, “The fact that many people can’t remember how to extract square roots doesn’t mean that pocket calculators are more intelligent than they.”
But what is intelligence, and why does it matter?
On that, a future post.