This paper explores the rules or intuitive predictions (for both category prediction and numerical prediction). Instead of using statistics, people rely on a limited number of heuristics to sometimes are reasonable and sometimes give serious errors. In fact, people seem to make predictions based on the degree which the prediction represents the given evidence.

This may work (sometimes representation conincides with likelihood), but often it has no correlation with likelihood or reliability of evidence.

Initial studies show the people predict by similarity, not be statistical likelihood. People relied upon descriptions and ignored base rates. Even when the subjects were told that the qualitative description had either low or high predictability, it had little effect on improving the statistical considerations. People still predicted likelihood based on similarity. However, when no description is available, people will use the base rate.

When specific evidence is given, people will ignore prior probabilities.

In numerical predictions people again base their prediction assuming the descriptions are highly accurate, even when told not to assume that. People fail to "regress" or move to predicting the mean even when the input data is stated to be highly unreliable.

Confidence also seems to be related to consistency of data.

In interviews, people will express great confidence in their evaluations, even when they know that interviews are notoriously fallible.

One problem with high reinforcement by praise is that we tend to praise when performance is high and punish when performance is low. By regression we see that after high performance (and praise) the most likely effect is to get worse, while the most likely effect after poor performance (punishment) is to improve.