Technically it doesn't fuck up the AI training; it just sidesteps it (still good and definitely easier)

If you input text that doesn't match the audio and they use this for training AIs, surely that will harm the training, no?

I don't think the audio challenge trains AI since it only plays a single clip. The visual challenge works because it mixes in answers it knows (to confirm you're a human) with answers it doesn't know (to train the AI)

My theory is that the audio challenge is purely for verification and they're okay with it not doing any training since a small minority of users will opt for it anyway.

But if the audio challenge is not for training and they know the right answer why are you able to pass it by typing something completely different but with some syllables in common? It seems to me either they only have a vague idea of what the right answer is or they are open to the AI being wrong and in future it will be adjusted by input they think is reliable because it's sort of similar.

Nah they accept a range of inputs because they know different humans can't consistently produce a single answer.

Maybe if this becomes more popular and gets abused they'll narrow the range.