Sketchy, but in the other direction. In high school I did a project that involved a support vector machine, and over the course of this I learnt how they worked to a reasonable level of detail.

Two years later, I interviewed at a large SV company, and they asked what projects I'd been working on. I gave a description of this project (2 years prior) and my explanation was phenomenal; in order to understand SVMs two years prior (without the 'necessary' mathematical background) I needed to develop all of the intuition (up to 'is kinda a high pass filter' etc) (which you might not ordinarily do at a time-pressured university course).

The interviewer was correspondingly impressed, and the SV company gave me an internship almost directly off the back of this interview.

The kicker is that at this point, I had a rudimentary knowledge of linear algebra, and absolutely no knowledge of any other machine learning; I had no business interning in their data science team.

My point being that even as a first pass, the bookwork questions don't work fantastically. FizzBuzz is no better, but a data science alternative would have weeded me out pretty quickly.

You're pretty harsh with yourself. You showed that you were interested in the topic and that you could learn it by yourself. I don't think anyone expect an intern to already master the subject he'll be working on, picking a smart motivated person is usually what you're looking for.

Ah, but if you combine it with the fact that when I did study those things I didn't really enjoy them, you can probably see why I look back on the experience with healthy scepticism :P

The question I'm more interested in: when you worked there as an intern, did you suck?

(If you didn't, maybe the interview worked...)

Apologies in advance for the bluntness, and don't feel like you need to answer.

I didn't work there; I went somewhere else for a number of reasons. I wouldn't have sucked, but it wouldn't have played to my strengths - my work would probably have been very average.

That tests book knowledge, not actual understanding of how to think about AI, and leaves itself open to interviewer bias, the very things fizzbuzz avoids.

The idea that FizzBuzz has no interviewer bias is laughable at best.

I can guarantee some will pick and reject people because they did things in a way they didn't like

No it does not. Without practical experience one just cannot discuss these things freely, and pure-book knowledge is clearly visible instantly.