Comment by moritzwarhier
19 hours ago
> saw this coming from miles away. Computers are better at solving CAPTCHAs than people are
good point... it's interesting how Captcha was initially popularized as a reverse Turing test, but it's just variants of Proof of Work today.
And it seemed clever at the time for Google to leverage this for improvement of their OCR models (it was!), and makes you wonder what utility is derived from the proven "work" today.
CAPTCHAs were designed as a type of Turing Test, not a reverse Turing Test. It’s not surprising that the effectiveness of these weaker variants has collapsed, given that AI can now pass the real Turing Test.
LLM’s can still only pass limited Touring Tests. The longer the interaction the worse they do. Which of course means you can easily create an experiment they successfully pass, but just as easily you can create an experiment where they fail.
CAPTCHAs are nearly useless because of how little you need to pay humans to solve them.
A more interesting question is whether there is a Turing test that is easy for ALL humans to pass, while still being hard for LLMs.
In practice, most of the major CAPTCHA vendors already rely on non-privacy-preserving tests for those needing more accessible solutions than a visual puzzle.
Google's audio captcha (only available in a few languages and unusable for those who also have hearing issues) only works for a narrow band of users, not trusted enough to bypass the captcha entirely, but also not untrusted enough. If you fall outside of that band, you get a nice "your device has been classified as a fraud risk, please use the visual captcha" message.
hCaptcha goes even further and straight-up requires you to have an "accessibility cookie", which requires verifying your email address (and apparently your phone number in some cases) to obtain, as well as disabling some anti-tracking settings in your browser.
1 reply →
I'm not sure if LLMs are solving most of these captchas. There are services that employ humans to solve them for pennies per captcha.
Oh, right, "reverse" was wrong here. I thought of "computer classifies user as computer or human" versus the inverse, while the word is about who classifies, not who's being classified.
(?)
I guess so
With the crosswalk, bike, motorcycle, stairs type of things, wasn't that just improving their training data?
Yes, for Waymo, AFAIK (I don't know for sure).
The OCR thing was earlier and used for Google Books, I think. Which is also is fitting for training data, or the motto "organize all knowledge".
At that time, this goal seemed really cool!