That's the problem with benchmarks. For IQ tests, we measure a small subset of skills, and then extrapolate to make conclusions about other, unmeasured skills.

This works fine as long as the tested humans are unaware of the methodology, because extrapolation doesn't work when the subjects have specialized, specifically, in the subset of skills from which we extrapolate. In other words, if we keep measuring intelligence through IQ tests, sooner or later we will only be good at solving IQ tests -- and that's not very intelligent.