A plausible hypothesis is that presenting frequency information simply makes algorithmic calculation of the result easier, and so subjects are no longer reliant on fallible heuristics in order to arrive at the conclusion.

There's only room for making it easier when the word "probable" is not synonymous with "larger N out of 100". So I maintain that alternate understanding of the word "probable" (and perhaps also an invalid idea of what one should bet on) are relevant. edit: to clarify, I can easily imagine an alternate cultural context where "blerg" is always, universally, invariably, a shorthand for "N out of 100". In such context, asking about "N out of 100" or about "blerg" should produce nearly identical results.

Also, in your study, about half of the questions were answered correctly.

The claim of the heuristics and biases program is that the conjunction fallacy is a manifestation of the representativeness heuristic.

I guess that's fair enough, albeit its not clear how that works on Linda-like examples.

In my opinion its just that through their life people are exposed to a training dataset which consists of

1. Detailed accounts of real events.

2. Speculative guesses.

and (1) is much more commonly correct than (2) even though (1) is more conjunctive. So people get mis-trained through a biased training set. A very wide class of learning AIs would get mis-trained by this sort of thing too.

I'm not familiar with the effect of variable string length difference, and quick Googling isn't helping. If you could direct me to some research on this, I'd appreciate it.

The point is that you can't pull the representativeness trick with e.g. R vs RGGRRGRRRGG . All research I ever seen had strings with small % difference in their length. I am assuming that the research is strongly biased towards researching something un-obvious, while it is fairly obvious that R is more probable than RGGRRGRRRGG and frankly we do not expect to find anyone who thinks that RGGRRGRRRGG is more probable than R.