Is it a graphic, thus requiring bulky OCR which most bots lack (apparently)? Even so, it is a fixed graphic and would not really require OCR'ing just simple pattern matching against a library of one item. I just don't get how it poses an insurmountable hurdle to frustrate bots.
The captcha monitors mouse behaviour. While it is easy for a bot to click a button on a form, it is hard to simulate the erratic movement of a mouse moved by a human.
However, it is not impossible: https://www.youtube.com/watch?v=fsF7enQY8uI
Text in captcha's will have been failed to be recognised by OCR. OCR used standard rules of what text is to recognise text.
Suitable Text usually be distorted, not parallel of run in straight lines parallel; to the horizon and contain random junk which OCR can't handle.
ie fail the usual rules of what text should look like.
With Deep Learning becoming more common its only a matter of time before Captchas will not work..
There are many different captcha's, some require choosing a number of graphics which have a theme (eg. which are parts of a sign which can later be further refined and then added to the library method below once the whole sign is assembled and the text graphic extracted by the same method) which a computer won't be able to discern. With this type you are always asked to identify a known scenario and usually an unknown for addition to the library of knowns once enough identical answers are received.
The most common use images in 2 ways:
A single randomly distorted image generated from a word then extra junk added to confuse OCR. Like "Salting" a password List by adding a "random" junk word to stop a rainbow attack.
Another form is using photos (usually of words) of something which people had to decide what it is because image is too complex to recognise automatically. Generally it's outside computer programming parameters of what defines text (or a sign or whatever) and is often surrounded by from a random environment.
This requires A large library of Photos with known "text" or other parameters like which are part of a sign etc.
The library for the second method is increased by providing 2 images which users try to identify correctly.
1 image is a known and another an unknown.
Correctly solving the known proves you aren't a robot.
Enough people matching/answering the unknown with the same answer then means that one is now known and can be added to the known library.
This is how Google Maps identifies what actually are street/place name signs (and then later the text they contain) and project Gutenberg texts which failed OCR were corrected.