I'm inclined to go with your alternative hypothesis. The quality of the captcha response degrades the more times you solve it consecutively.

I was recently developing a web frontend that had a captcha implementation on the login form. The first couple times I solved the captcha there was no problem. Then once it got to fourth and fifth one it started making me solve a few of captcha pages in a row. Eventually it was forcing me to solve several pages of "Select the picture of the fire hydrant" over and over, often marking it as incorrect when it definitely wasn't, and then restarting back from the beginning so each run-through took multiple minutes forcing you to solve dozens of captcha pages before accepting it as correct.

So, to add to your hypothesis, they're probably doing this intentionally not only to get more data for their AI algos, but also to mitigate people training bots on CAPTCHAs.

EDIT: Also the fact that you're being downvoted is baffling to me but my tinfoil hat persona makes me wonder if it's some orchestrated downvote because you're pointing out a hypothesis that Google is trying to keep hidden.

It's both. The more consecutive CAPTCHAs you solve in a short period, the more likely Google is to assume you're not a human (or not up to any good), but you definitely can get google browsing identity (however they tag you as a combination of IP, browser fingerprint, etc) to the point that they won't every consider enough responses a good indicator. I've done it many times, and yes, it usually is fixed within a few hours or the next day.

I agree that a heuristic (machine learning or otherwise) designed to prevent spam was almost certainly the culprit. I disagree with the notion that this does not constitute a hellban, but I'm open to considering alternative nomenclature.

The fact stands that google locked me away from my documents and intentionally undermined my ability to understand what had happened or rectify the situation.

Collateral damage is still damage.

It may be a ban or an errorban, but it's not a hellban, in any sense of what "hell" means in "hellban:: https://blog.codinghorror.com/suspension-ban-or-hellban/

It also might just be a broken integration with recaptcha.

It's ironic that this comment saying about "hellbanning" was "autodead", and I had to "vouch" to revive it.

I wonder if it hit some automatic filter by virtue of being new, and talking about banning/hellbanning?

I find I can often get away with throwing in a wrong answer or two.

It's probably similar to old recaptcha where they give you a selection of known answers and a couple of unknowns which they're uncertain of. Probably known not-correct answers too so that the system works.

If you select the unknowns as well as all of the known answers (and don't select the wrong answers) they can train their AI robots that the extras you selected are also a part of the "has traffic lights" dataset