Captchas are made to defeat AI logic, so sometimes it’s not the obvious thing. It could very well possibly be selecting all images that match the color tone, something a bot may not work out.

IMO the idea here is that most users are not thinking very hard, so they are going to see the word “warmer”, think “snow = cold” and leave their analysis at that. AI on the other hand is going to put more effort into interpreting the specific meaning of the request in context of the images. The primary challenge for captchas now is to defeat AI, so the captcha ideas that get through probably did so because they gave the AI trouble in testing, but did not give most users trouble.

I think that going forward, people who put thought into following specific directions accurately are going to have a lot of trouble with captchas.