Hacker News Reader

Comment by boomboomsubban | Hacker News Reader

Not saying these are good, but it's been ~10 years and I still have no idea what I'm supposed to click when Google says "click the traffic lights." Just the light sections? What about the edge that takes up 10% of the adjoining box? What about the back or side of a different traffic light?

I've tried them all, and my success rate doesn't noticeably change.

View on HN

it gets worse when you factor in translation problems.

lets take google/recaptcha/hcaptcha image captchas. english: "click on the images with bikes" this could mean motorbikes or bicycles in english.

in german it says "klicken sie die bilder mit den fahrrädern". fahrrad meaning bicycle exklusively! a motorbike would be "motorrad". then the images will show no bicycles so you skip - wrong - so you click on the motorbikes - wrong - oh there is one image that shows a (german) "motoroller" (a scooter) maybe they mean that? click - correct

this goes on an on and on. they have so many problems with their translations its infurating. i stopped filling them out when not absolutely necessary

wonder how this affects the models that are trained on the data generated by these capcha's? will german motor cyclists be less safe because of this

It's mostly trained on the answer of other humans. The best advice is literally "just be yourself, be human."

I had some fun poisoning the well on the older text based captchas by answering the first word correctly and putting "penis" where the last word would go. It always accepted it for some reason.

The reason is that those captchas were helping to OCR-transcribe books. The first word would be known to the system and serve as the actual verification, and the second word would be unknown and serve as you performing free MTurk work.

Honestly I don't think it matters which boxes you click

I think it's based on mouse movements, time delay between clicks, and browser fingerprinting tricks that detect headless browsers

In pretty sure it does matter. Click things that are clearly wrong and it will very rarely accept it.

However I think it is quite flexible. That box that is almost completely traffic light probably needs to be clicked, but as long as you pick at least one of those two that contain a corner you will likely pass. I would guess that there is some sort of accuracy score that is mixed in with the bot fingerprinting score.

The worst part about those captchas is that you have to click each box individually. So unless you're on a touch screen or mouse, clicking each box that contains (part of) the object, takes ages. At least let my human hands drag-select the boxes, ffs.

I think that’s all accounted for.

I always get it wrong on the first try, so I guess they're accounting for me being a robot.

It's so random. Sometimes I click submit thinking it would fail because I forgot to select a metal pole of a traffic light, or confuse bike for a motorcycle.

Reply on news.ycombinator.com