Anticipating what everyone else on the road will do
Once perception has identified the cars, cyclists, and pedestrians around the vehicle, the prediction stage has to estimate what each of them will likely do over the next few seconds. A pedestrian standing at a crosswalk edge, angled toward the street, is far more likely to step out than one facing away and looking at their phone. A car drifting toward a lane boundary with its turn signal on is telegraphing a lane change well before it happens.
Modern prediction systems lean heavily on machine learning trained on huge volumes of real driving data, learning the subtle behavioral cues that correlate with different outcomes. Because the future is inherently uncertain, good prediction systems don't output a single guess, they output a distribution of plausible futures for each agent, weighted by likelihood, so the planning stage can prepare for more than one outcome at once.