The older captcha with words was for the digitization of books. It helped in the Google Books, and Internet Archive projects notably.

https://cogdogblog.com/2013/12/recaptcha-all-the-books/

The current one helps in identifying and navigating streets, useful for maps and autonomous driving.

https://techcrunch.com/2012/03/29/google-now-using-recaptcha...

You'll notice that recaptcha always has a direct effect on current projects of Google/Alphabet: Google Books, Google Maps, Waymo.

Incidentally, it's a mystery to me why so many of the Gutenberg/Internet Archive books have such ludicrously bad plain-text renderings, utterly unusable/unreadable, with often barely one word correct per page - since neural nets have been getting very high scores on MNIST handwriting recognition (for example) for a long time now, maybe since the 90s? It's a shame.

I believe the "read this hard-to-read text" ones are helping machines read text, the "is this a street sign" ones are helping machines drive cars.

On any level, the data represents an average product of human visual calculation. A simple use case would be comparing that with the product of machine visual calculations to better understand and optimize the systems they are designing.