Hacker News Reader

35 out of 36 will be scarier to most people. In behavioral psychology they call this "denominator neglect."

Right, but surely sample size matters.

I shortened the title to fit HN limits. Why do you think "34 out of 36 were wrong" is less scary than "94% were wrong"?

Another way to think about that: %99.99 of AIs lose chess against our best human player.

When in reality 1 AI consistently beats all Human players every time.

Only if you can go back and demonstrate that the 2 AIs that didn't fail are consistent and demonstrably correct. Otherwise they just got lucky.

I think if they flipped it and said "Only 1 out of 36 AI systems evaluated was more accurate than ..." would have a higher impact

You mean 34/36

Is that less scary?

35 out of 36 is scary.