All the proofs that I know of allow one to get lucky with probability about .5 in each round. When you do an interactive proof with 100 rounds, you have a 2^-100 chance of getting away with cheating.
When you go non-interactive with 100 rounds, an adversary could cheat by trying about 2^100 proofs. So you replace a stronger guarantee with a weaker one, but 2^100 is a pretty big barrier.
(I just looked and the Wikipedia page and it’s very confusing fwiw)
But the number of questions you need to compensate grows only a little.
For example, interactively if you ask for Merkle tree proof that selected leaf values have a particular property, you only have to ask for about k leaves to get probability 1-(2^-k) that you'd catch a dishonest prover who had committed a Merkle root with less than half the leaves having the property.
Non-interactively, a dishonest prover could secretly grind attempts, say 2^g times, and then you'd have a lower probability of catching them, approximately 1-(2^(g-k)). But g can't be all that large, so you can increase k to compensate without making the proof much larger.
You.can also require certain hashes to have a fixed prefix, like Bitcoin mining, forcing every prover to have to grind 2^p times. This reduces the effective g that a dishonest prover can achieve, allowing k to be smaller for the same security, so allowing the non-interactive proof to be smaller. At the cost of honest provers having to grind.