[dheera]

It is pretty straightforward to train a neural network to solve these -- e.g. fire hydrants, traffic lights, cars.

I would have thought ReCAPTCHA would take into account human factors (e.g. speed of clicking) as higher priority to the accuracy of the selection.

[floatingatoll]

Regularity of clicking is considered a sign of robot behavior, which is especially frustrating if you learned to perform repetitive image identification mouse tasks in a computer with rhythmic regularity (think Turk, for example).

[wyre]

AFAIK it takes into account mouse movement and the speed of clicks.

[therein]

In my experience, relatively easily defeated by `await Promise.delay(randomDelay())`

[dheera]

Sounds like a cat and mouse game.

Mouse: They could then try to analyze human delay randomness -- it's probably not uniform.

Cat: And then someone will come up with a replacement to randomDelay that mimics the above pattern.

Mouse: And then they will look for changes in the distribution itself from person to person

etc.

[MertsA]

I know back in the day for RuneScape bots using SCAR there were macros to move the mouse from one position to another on the screen with randomized acceleration, randomized curvature, overshoot, clicking in some bounding box, etc. all using a normal distribution in an effort to thwart detection. Imagine being the poor developer tasked with trying to recover some signal out of that.

[drusepth]

Just alt-tabbed in from writing runescape bots to HN and wanted to say that's still the case (for the client I use). The code is pretty complicated now but still functions much like you say. Mouse position is tracked, then any input which repositions the mouse accelerates at a "reasonable rate" and with a randomized curvature to get "close enough" before it self-corrects and gets a pixel-perfect click.

A few years ago the client stopped sending mouse data back to Jagex altogether. Luckily, I don't think there's many poor developers tasked with trying to recover any signal out of that anymore. :)