I've regularly been getting prompts that I'd never seen before. I don't recall any time since joining EMC that there has been such great variety in prompts. Of course, that only makes them trickier to solve for us. But potentially trickier to automate as well. Here's one I just got: I thought it was talking about colour temperature at first, because both in Dutch and in English the prompt is talking about the image, and not about what's on the image. However, when I saw three images depicting snow I figured they did mean the content, so I didn't select the image second from the top, first from the left.
Lately this puzzle, which I'd gotten before, has been changed by reducing the contrast, applying some sort of haze effect. As often with these puzzles, I wonder if AI might have some issues with it, or if it's just to check if you're human by it taking you some time to solve it.
IDK why, but I feel like they just get easier to fail. I know what a motorcycle looks like, I have an e-bike of my own. But I miss ONE single square with like 1% of a headlight and it is an automatic "Please try again."
Totally felt, what's worse are these: I'm not kidding where I've had a few of these in a row and I get the "Try Again" message regardless.
So, I don't know if you've noticed, but there are - I think - three companies that we have to deal with now. hCaptcha - This is newer and I actually find the puzzles easier that the horrible reCaptcha ones. Ususally you can see what they want at the top in the example picture. They have gotten a bit harder, but I rarely get in a fail loop with them. Cloudflare - This is very new with minecraft-mp.com. It is super simple as it seems to determine humanness by how you check a box!? I have no idea how this works, but it rarely fails. reCaptcha - I HATE this one. They are owned by google. Traffic lights, Cars, Motorcycles, Bicycles, Busses, Stairs, Blurry pictures. They seem to have gotten SLIGHTLY easier to navigate, but I hate them with a passion. Is the person on the bicycle or motorcycle part of the darn thing!? Is the corner of the traffic light part of the traffic light? Is there a car in that blurry picture (or not)!? If a captcha is going to get me in a fail loop it is this one most of the time. They have been at it the longest, you'd think they would be able to do better! Just my 2 cents worth
for reCaptcha, never do them when it is an image partitioned, always do them when it asks to choose out of 9 images. sometimes it takes a few reloads to get an acceptable one.
I thought I'd tell you: These systems use session information to figure out if you are human. By clicking the box, you simply allow to send that information over, after which it confirms that that session information is 1) (roughly) unique, and 2) Similar enough to typical human session information. Basically, if you open a clean and new VM trough some VPN service, open the web browser, go to the website by copying the link over, and then click the box, it will fail. Everything you do that is not that makes it work. I don't know exactly what information is available to them as I don't work in frontend, but your browser gives websites more information than you might think. You seem to ask that quite a lot, and there are good reasons for them being the way they are! First of all, you should know that most of these tasks are judged not by some ground truth, but instead by what other humans do, collecting a lot more information than just what you click (see answer above). That doesn't mean things that are easy enough to solve by machine learning aren't ever used, or that capcha's aren't beatable (most of them are,) but instead, it means that, to beat it, you'd need to know not the correct labelling, but the human labeling. Most of these images are deliberately made to attack networks, and usually you solving them is you doing clickwork for some AI company that is training something. As bare image detection has gotten better, you see a lot of these "attacker" images. If I'm not mistaken, those will usually be generated by the adversary in some adversarial neural network system, where it is using human input to validate its current state, or at least that is what the images look like to me. I should also state that there is only so much "machines cannot do this" people care about. At some point, it's cheaper to just hire a warehouse of clickworkers in Kenya, who charge tenths of a cent per completed Captcha, than to do any machine learning to solve it automatically.
Thanks for your contribution, I think what you're saying is probably correct. I didn't think that these days, CAPTCHA results were still used for training (as they used to be with the original reCAPTCHA), but thinking about it, they very well might be. It's a cheap way to get a lot of (poorly) labelled data.
