That's a bit of an overstatement.

The entire field of ML is aimed at problems where deterministic code would work just fine, but the amount of cases it would need to cover is too large to be practical (note, this has nothing to do with the impossibility of its design) AND there's a sufficient corpus of data that allows plausible enough models to be trained. So we accept the occasionally questionable precision of ML models over the huge time and money costs of engineering these kinds of systems the traditional way. LLMs are no different.