Is it a bicycle?
-
?
-
@sockpuppet7 Are you a fan of Magritte?
-
@sockpuppet7 I’ve also seen at least one that had a picture of a tricycle in it, which as I recall, I needed to click else it didn’t think I had clicked on enough bicycles.
-
@Gurth said in Is it a bicycle?:
@sockpuppet7 I’ve also seen at least one that had a picture of a tricycle in it, which as I recall, I needed to click else it didn’t think I had clicked on enough bicycles.
Good, feed the disinformation inherent in the AI training system.
-
I once got this one ("Please click on every picture depicting a bear"):
Of course, they didn't account for the fact that some languages don't consider pandas to be bears.
-
@sockpuppet7 it's not an image with bicycles
-
I think it's a motorcycle?
-
/ 
Since it's widely accepted/rumoured that these things are more for AI training than for captcha'ing, has anyone tried to give completely wrong or random answers and see if the system lets you continue or not?
-
@remi There would have to be a coordinated majority of people willing to trick the system. Doable, but unlikely.
-
@remi said in Is it a bicycle?:
/ Since it's widely accepted/rumoured that these things are more for AI training than for captcha'ing, has anyone tried to give completely wrong or random answers and see if the system lets you continue or not?
So they actually have a control mechanism for this. Some of the time you get shown an image for recognition where it knows the answer to a reasonable degree of accuracy and is asking you to validate that answer (= improving your reputation with the system)
Sometimes it doesn’t know the answer to it to a high degree of accuracy. It might have an answer and it will take your answer as additional weighting. If you keep giving it answers it doesn’t agree with, it will increasingly trend towards giving you challenges it has higher confidence in so that instead of you feeding it information, you have to rebuild its confidence level of you first.
I may have discovered some of this with extended fuck-around-and-find-out sessions, in terms of trying to find out how protective it really was in terms of using it as a generic anti spam method (answer: not nearly as effective as you’d hope and certainly less than they’d have you believe)
There was a point it would let me through if I just ticked all the boxes.
-
@Zecc no, I'm not saying this would break AIs, but rather whether the website that presents you with these things will let you log in (or continue whatever you're doing) if you give the wrong answers? This would "prove" whether the website actually uses this for captchas, or for some other reason.
(quotes around "proves" because a single person on a single website wouldn't prove anything, sites sometimes give you several of those to solve, and since captchas can actually be just weird "I'm not a robot" checkboxes, clearly the actual human detection can happen based on things other than the actual pictures you're picking (like the time it takes you to do it, or the mouse movements or whatever))
-
Thanks, that's quite interesting and more evolved that I would have thought.
@Arantor said in Is it a bicycle?:
how protective it really was in terms of using it as a generic anti spam method (answer: not nearly as effective as you’d hope and certainly less than they’d have you believe)
I'm not really surprised here, especially since I read that story years ago about how you can literally mechanical-turk these things. Some guy found that you could buy for a few pennies real-time captcha validation from people in China or similar. You send them the captcha and (statistically) a human will respond in the next few seconds. This is real humans (though in a shitty place and with a shitty jobs!) mass-answering captchas, so obviously at that point any anti-spam system relying on "spam is from not-humans" will fail.
-
@remi this is still a thing, though the going rate is nearer $3 USD/1000 solved CAPTCHAs now.
The control mechanism isn’t new either - even back when reCAPTCHA was still a project with CMU before Google bought it, they used to show you two bits of text to identify, one was the control, the other was the crowd-sourced part, and as more and more people gave it the same answer for the one it didn’t know for sure, it would move it into the pile of “successfully sourced”.
These days though there’s been demonstrations of using both Google Image Search and various of the AI tools to solve reCAPTCHA.
Though both v2 and v3 do more than just have you solve puzzles. There’s an amount of system profiling going on with stuff happening in what they refer to as a virtual machine as though this stops spammers. (Reality is they just spin up instances of headless Chrome, or have small farms of devices.)
-
-
@PleegWat But that text is missing on the image!
-
@BernieTheBernie Editing the image sounds like work.
-
@remi said in Is it a bicycle?:
has anyone tried to give completely wrong ... answers and see if the system lets you continue or not?
Not intentionally
-
-
@Arantor with a little bit of (human) training, you could quite confidently detect which was the control word and correspondingly answer with “correct-word cuss-word”.
-
@remi said in Is it a bicycle?:
/ Since it's widely accepted/rumoured that these things are more for AI training than for captcha'ing, has anyone tried to give completely wrong or random answers and see if the system lets you continue or not?
https://www.reddit.com/r/4chan/comments/sa4nx/operation_renigger/
-
@topspin said in Is it a bicycle?:
@Arantor with a little bit of (human) training, you could quite confidently detect which was the control word and correspondingly answer with “correct-word cuss-word”.
Back in ye olden reCAPTCHA times it was fairly trivial to work it out which was which because for a while it wasn’t even random, control was first, to be learned was second. Then they figured out that people had figured it out and randomised it, but even then it was possible to get a feel for it.
It all went to shit after Google bought it, of course.
