Like Matthias, I had always assumed that assistive tech (e.g. screen readers) would have some method of translating
<em> HTML elements into emphasised content. A change in voice inflexion; a user hint; something. Apparently, they don't, and the one time NVDA added the feature, it was so universally despised that they subsequently removed it. Not hid it behind a settings flag; full-on deleted it entirely. Useful to be aware of!
On why you should still use these semantic elements anyway (not least of all, they're a better DX than arbitrary styled
<span> wrappers, and sighted users do benefit from visual emphasis):
Does this mean that you don’t need to use semantic elements to convey emphasis? Of course not. You never know where the semantics still matter. And it is also just the right way to indicate emphasis in HTML for the majority of users. But it is nevertheless interesting and important to know that screen readers don’t care by default.