The captcha to sign up shows AI-generated images and asks things like "select images of a dog" but there are zero dogs, just dog approximations. I really hope these identifications won't be used as training data.
hCaptcha has been doing this for a few months now to my knowledge - I'm guessing this is being used to train a CLIP-like system (or alternatively: it might just be that poorly generated ai images are the best candidate for a captcha in the age of modern subject recognition, it's hard to know)
I mentioned noticing it several months ago, but it sounded like others hadn't seen it yet. It's always hcaptcha (or at least the atrocious ones are, maybe the others are just super good).
I have the same thought at first, it's really amusing to think that they expect me to select silly wrong things to pass a test and mistrain their system. I like to hope.
Alas, statistically they can totally do this by showing the same images to multiple people so that they can estimate what most people like and then define a threshold outside which they predict that it's not a human. One dimension is related to their AI tuning variables or whatever, the other is related to their analog signal from human measurements of the same images. They surely do this already since humans also have noise in their responses to a survey with anything that's not obvious.
They can still apply on-off thresholding to the analog signal that is a survey over many participants. It still gives these ridiculous gas-lighting training systems though.
One thing that's probably true -- they're finding out what the lowest quality image is that people will identify as something, which is unique. Usually people vote for what they like best, but if the stupid thing forces you to click on a certain number they can adjust that threshold.
Maybe they intend to touch up subsections of images such that they're closer to these minimum quality thresholds, who knows... just something different.
