I, for one, welcome our traffic light-identifying overlords.
Anyone who has been surfing the web for a while is probably used to clicking through a CAPTCHA grid of street images, identifying everyday objects to prove that they're a human and not an automated bot. Now, though, new research claims that locally run bots using specially trained image-recognition models can match human-level performance in this style of CAPTCHA, achieving a 100 percent success rate despite being decidedly not human.
ETH Zurich PhD student Andreas Plesner and his colleagues' new research, available as a pre-print paper, focuses on Google's ReCAPTCHA v2, which challenges users to identify which street images in a grid contain items like bicycles, crosswalks, mountains, stairs, or traffic lights. Google began phasing that system out years ago in favor of an "invisible" reCAPTCHA v3 that analyzes user interactions rather than offering an explicit challenge.
What's ironic is that the main purpose of reCAPTCHA v2 is to train ML models. That's why they show you blurry images of things you might see in traffic.
AFAIK the way it works is that of the 9 images, something like 6 are images the system knows are True or False, and another 3 are ones it is being trained on. So, it shows you 9 images and says "tell me which images contain a motorcycle". It uses the 6 it knows to determine whether or not to let you pass, and then uses your choices on the other 3 to train an ML model.
Because of this, it takes me forever to get past reCAPTCHA v2, because I think it's my duty to mistrain it as much as possible.
It'd be a bit unreliable, though. Not everyone has the same reaction to the same thing, nor do they express it in a similar way.
Someone might think a snake or a spider is cute, whereas another would want to incinerate it on the spot. A third might be concerned because they seem to be injured, etc.
Not to mention that image recognition/emotional analysis has been an ongoing field of research for some time. Making the link is not overly difficult.