As a correction to the title, they said four, but then said, "There is a fifth, undefined group, representing 10%, which the algorithm is unable to classify in relation to a clear type of behavior. The researchers argue that this allows them to infer the existence of a wide range of subgroups made up of individuals who do not respond in a determined way to any of the outlined models."

So based on the study: http://advances.sciencemag.org/content/2/8/e1600451.full

and looking at the results in: http://d3a5ak6v9sb99l.cloudfront.net/content/advances/2/8/e1...

the summarization of "Undefined" as "Decides randomly" seems a little wrong- there's an obvious tendency in PD (Prisoner's Dilemma) for some results: http://d3a5ak6v9sb99l.cloudfront.net/content/advances/2/8/e1...

Also, I can't help but wonder whether there is bias effect, as humans seem to tend to use a small number of groups or factors for personalities. For example, in literature, J.K. Rowling's sorting hat chose between Gryffindor, Hufflepuff, Ravenclaw, and Slytherin, and Veronica Roth had citizens of Chicago choose between Abnegation, Amity, Candor, Dauntless, and Erudite. Even though there is a mix between races in humans, we tend to categorize- like in the U.S., typically you must choose from Caucausian, Asian-american, African-american, Pacific-islander, or Hispanic, even though color and genetic makeup vary. Then MBTI has four factors (E/I, S/N, T/F, J/P). Friedman and Rosenman came up with A and B types then later Denollet added the D type. We tend to categorize like this.

Also, they chose just four games for the study. That could have affected the outcome.