When two heads are better than one and when they can be worse: The amplification hypothesis

According to the self-consistency model (Koriat, 2012a), confidence judgments in the responses to 2-alternative forced-choice items are correlated with the consensuality of the responses rather than with their correctness: For consensually correct (CC) items, for which the majority response is correct, accuracy is higher for the correct answer than for the wrong answer, whereas for consensually wrong items (CW), confidence is higher for the wrong answer. Assuming that group decisions are dominated by the more confident members, a maximum confidence slating (MCS) algorithm that was applied to virtual dyads outperformed the better member for CC items, but yielded worse performance than the worse member for CW items (Koriat, 2012b). We examined whether group deliberation also amplifies the tendencies that are exhibited by individual decisions, or rather improves performance for both CC and CW items. A perceptual task and a general-information task yielded very similar results. MCS applied to the individual decisions yielded a similar amplification as in Koriat (2012b), but dyadic interaction accentuated this amplification further. Thus, group deliberation had an added effect over confidence-based judgments, possibly due to the exchange of arguments within a dyad, but both confidence slating and group deliberation affected performance in the same direction, improving accuracy when individual accuracy was better than chance, but impairing it when individual accuracy was below chance. Notably, for CW items, group interaction not only impaired accuracy but also enhanced confidence in the erroneous decisions. The mechanisms underlying consensual amplifications were discussed.