Hacker News Reader

Comment by xapata | Hacker News Reader

hCaptcha makes money by having humans label things to teach machines. This suggests that at some point, the machines will be nearly as good as the humans at labeling. At this point, until the humans are tasked with a different training exercise, a bot will be effectively indistinguishable from a human via hCaptcha.

If there's value in it, it sounds like a spammer could train an hCaptcha-defeating bot via hCaptcha.

View on HN

I think it's likely they'll add new types of exercises as this starts to happen.

Wait they actually use the labeling for something other than admittance? Well that takes care of my concern about this tech then, gotta go back and change some of my other comments...

Google uses reCAPTCHA labels exclusively for themselves, and is extracting hundreds of free person years of labor from internet users every single day via this. hCaptcha lets anyone access this type of service, provided they follow certain ethical AI guidelines

How long has the content of the recaptcha puzzles been entirely unchanged, despite being shown to billions of users? Like five years? It should be painfully obvious that there is no actual labeling going on at this point, they have all the training data they'll ever need for traffic lights...

And by a corollary, since they haven't started labeling different kind of data, it's clear that either they no longer need any kind of labels at all, or this is actually not a cost-effective way of doing it.

For me at least, the images on recaptcha have been getting much much worse, to the point of often being almost indistinguishable. So although they are still typically asking for the same things (cars, buses, stop signs, traffic lights) they do seem to be actually still making progress on the labelling effort.

Anecdotally I don’t think I’ve ever seen the same image twice on recaptcha.

How does the recaptcha bottom help train AI?

This is how Skynet starts.

Reply on news.ycombinator.com