Specifically, the problem is it produces _almost_ correct code, which is worse than incorrect code because it might fool you into trusting it.
To be clear: we’re in agreement that incorrect code that passes for correct at a glance is even worse than obviously-incorrect code.