Since the identifiers are chosen randomly, it is possible that two or more sources will choose the same number. Collision occurs with the highest probability when all sources are started simultaneously, for example when triggered automatically by some session management event. If N is the number of sources and L the length of the identifier (here, 32 bits), the probability that two sources independently pick the same value can be approximated for large N  as 1 - exp(-N**2 / 2**(L+1)). For N=1000, the probability is roughly 10**-4.
The typical collision probability is much lower than the worst-case above. When one new source joins an RTP session in which all the other sources already have unique identifiers, the probability of collision is just the fraction of numbers used out of the space. Again, if N is the number of sources and L the length of the identifier, the probability of collision is N / 2**L. For N=1000, the probability is roughly 2*10**-7.
The probability of collision is further reduced by the opportunity for a new source to receive packets from other participants before sending its first packet (either data or control). If the new source keeps track of the other participants (by SSRC identifier), then before transmitting its first packet the new source can verify that its identifier does not conflict with any that have been received, or else choose again.