I don't think the study supports many of the conclusions being drawn in the comments here.

What the study measured was that older people tended to be worse than younger people at adjusting their predictions after making small errors (theoretically because of lower uncertainty levels), but better than young people at adjusting predictions after making large errors (theoretically because of higher surprise sensitivity).

What the study did was fit each person with one of the curves in figure 2, the curves being functions of how much a person was willing adjust their new prediction ("learning rate" value on Y axis) in response to seeing different amounts of error on their previous prediction ("relative error" value on X axis).

- In the "normative" case (best case), a person has a perfect S curve, where they don't adjust their predictions too much after small errors, but greatly increase willingness to swing their predictions after errors reach a threshold.

- In the "surprise insensitive" case, where a person is unable to be surprised by large swings in data and update their predictions accordingly, the steep rise in their S curve is flattened out. These people are bad at learning after large errors.

- In the "low hazard rate" case, where a person is able to be surprised by large swings in data, but their threshold for surprise is too high, their S curve is shifted to the right. These people are bad at predicting after moderate errors, but fine at predicting after small and large errors.

- In the "low uncertainty case", where a person is too sure of themselves at low error levels, the S curve is depressed at the left end. These people are bad at learning after small errors, but good at predicting after medium and high errors.

- In the "reduced PE" case, where a person isn't good at understanding magnitudes of prediction errors at all, their S curve is vertically compressed, and their predictions are worse across the board, at low, medium, and high error levels.

Figure 6 shows outcome of the experiment, with age being correlated with higher "uncertainty underestimation" ("Unc") and higher "surprise sensitivity" ("SS") in the fitted curves.

This happens because older people are worse at learning from small changes in data, but can compensate somewhat be being more willing to change their predictions after large swings in data.

"Insufficient uncertainty" might be a reasonable explanation for this, but it's easy to imagine other possible explanations as well. Maybe "insufficient attention" or "insufficient caring" could be factors, with older people maybe being more willing to stick to rough predictions without sweating the details. It would have been interesting if the study tried to measure self-confidence / certainty levels more directly, instead of just relying on fitting to a theoretical model.