I looked this up and realised it’s the page I’d seen briefly on a range of websites lately. It’s not annoyed me at all. Not nearly as much as having to complete captchas with slow refreshing tiles.
A person's testimony is a source. I can add mine: you can tell when you're truly blocked because if you click for the accessibility audio-based captcha it will actually tell you you're blocked (but, if you did the visual captcha, would simply loop forever while telling you you did it wrong).
I don't think you'll find an article by Google saying "yes, we sometimes completely block users while making it look like they're not blocked and wasting their time".
The captcha itself (matching pictures to text) is mostly for ML training data. I think pass/fail is mostly based on heuristics like how you moved your mouse which could get analyzed before you complete the captcha. https://www.techradar.com/news/captcha-if-you-can-how-youve-...
reason why is 1. Google and others really needed the training data, and 2. it probably helped justify the cost of providing the captcha service for free worldwide (old free tier was 1M/mo)
That was Louis von Ahn’s stated goal when he announced it - i think it was his NIPS (now NeurIPS) talk where he basically outlined the training data paying for the infrastructure as a fair exchange.
I implemented something similar for my bot defences. If headless chrome is detected you still get the same anubis-style PoW but even if you submit the right answer you get rejected.