Like the presumption of a fair coin in this page, his presumption is that the questions are fair. If they were trick questions, such that the seemingly improbable answer is correct more often than not, guessing based on partial knowledge may not be a benefit.
I think he's right, but I'd trust his results more than his abstract reasoning.
It seems plausible that a test might have a bias whereby the harder questions are predominately one answer or the other. How large would this effect have to be to overwhelm the 2% probability? Would you trust your grade on the presumption that this effect can be ignored? What about your life?
For the record, I agree with Yvain's advice. I did just fine on the SAT's, partly because I followed the mathematically correct advice that one should guess if one can eliminate at least one of the 4 choices. But I wouldn't proclaim that this strategy needs no testing just because the mathematics are correct.
http://en.wikipedia.org/wiki/Gamblers_fallacy
http://en.wikipedia.org/wiki/Representativeness_heuristic
Minor quibble: if a coin fell heads 29 times in a row, my credence for it falling heads a 30th time is higher than 0.5 because I now have reason to believe the coin is biased. How much higher depends on how many biased coins there are, as yummyfajitas said.
Here's another epic example of people failing to use basic probability theory: http://lesswrong.com/lw/13i/shut_up_and_guess/