OK... from the sound of it, it isn't really a good application for SRS systems. (They're focused on data, not skills - and thinking through biased problems would seem to be a skill, and something where you don't want to memorize the answer!)

However, probably one could still do something. Mnemosyne 2.0 is slated to have extensible card types; I've proposed a card type which will be an arbitrary piece of Python code (since it's running in an interpreter anyway) outputting a question and an answer.

My example was generating random questions to learn multiplication (in pseudocode, the card would look like 'x = getRandom(); y = getRandom(); question = print x "*" y; answer = print x*y;'), but I think you could also write code that generated biased questions. For example, one could test basic probability by generating 3 random choices: 'Susan is a lawyer' 'Susan is a lawyer but not a sledge-racer' 'Susan is a lawyer from Indiana', and seeing whether the user falls prey to the conjunction fallacy.

(As a single card, it would get pushed out to the future by the SRS algorithm pretty quickly; but to get around this you could just create 5 or 10 such cards.)