You are correct; there are large individual differences in decline. Part of the reason for the big differences between the cross-sectional and longitudinal designs are exactly because of that. The longitudinal design controls for some of that variance by looking at individual subjects over time. Probably some of the individual subjects are actually increasing in cognition or staying the same over time. Some of the individual subjects are probably doing much worse than what the figure shows. The mean paths are useful for summarizing the data as a whole and observing the overall pattern.

When scientists look at averages, they also take note of the variability in the measures that they sample. To make a significant statement about the mean, they are careful to look at the differences between means relative to the overall variability in a sample.

There are some cases where the average would not be a good way to represent the trend. For instance, it could be that there were mostly two types of people: some people continue steadily improving their intelligence throughout their life, and some people have huge and dramatic cognitive decline as they age. The average of the two may be a steady but unimpressive decline, which fails to represent what actually happens to any individual in the whole population. There are efforts to check for such bi-modal distributions, and it doesn't seem to be the case in normal cognitive aging.