> " could it somehow get access to the subtitles and then use them to answer queries?"

It's not even necessary - computers are already excellent at understanding spoken words. Have you tried automatic captioning recently? Half the inputs to my phone are already voice, not text.

Video is a harder problem, but it's not too far behind.

Exactly, and many bots exist today to mine user videos for the automated subtitle information. In other words, there's no escaping GPT from learning from any kind of medium.