I think we tend to intuitively "normalize" the likelihood of a complex statement. Our prior is probably Kolmogorov complexity, so if A is a 2-bit statement and B is a 3-bit statement, we would "expect" the probabilities to be P(A)=1/4, P(B)=1/8, P(A&B)=1/32. If our evidence leads us to adjust to say P(A)=1/3, P(A&B)=1/4, then while A&B is still less likely than A, there is some sense in which A&B is "higher above baseline".

Coming from the other end, predictions, this sort of makes sense. Theories that are more specific are more useful. If we have a theory that this sequence consists of odd numbers, that lets us make some prediction about the next number. If our theory is that the numbers are all primes, we can make a more specific, and therefore more useful, prediction about the next number. So even though the theory that the sequence is odd is more likely than the theory that the sequence is prime, the latter is more useful. I think that's where the idea that specific theories are better than vague theories comes from.