The asymmetries cut both ways. The site can detect probing attempts, and if that happens switch into a mode where the captcha results are obfuscated. Let's use blog comments as an example:

The normal mode of operation is to give the user a clear error message if they failed the captcha, and have the post go through if they passed the captcha.

If the number of failures is higher than e.g. 10 in the last hour, successful captchas cause the comment to be put in a hold queue for an hour. Failed captchas cause the comment to be rejected. The attacker can't know if they passed the captcha until an hour has passed, which slows down their iteration a lot.

You can then slow it down further with some randomization, while keeping the experience of real users the same. E.g. successful captchas go to the hold queue. Failed captchas are rejected 99% of time time. The remaining 1% of the time they go to the hold queue, but are auto-deleted after a random delay of 2h-12h.

So you accept a small amount of temporarily visible spam as the price of obfuscating the signals.