35 comments

[ 5.4 ms ] story [ 80.5 ms ] thread
I know people were solving CAPTCHAS with neural nets (with PHP no less!) back in 2009.
Seems like Google Gemini is tied for the best and is the cheapest way to solve Google's reCAPTCHA.

Will be interesting to see how Gemini 3 does later this year.

I’m sure they do better than me. Sometimes I get stuck on an endless loop of buses and fire hydrants.

Also, when they ask you to identify traffic lights, do you select the post? And when it’s motor/bycicles, do you select the guy riding it?

The buses and fire hydrants are easy. It is the bicycles. If it goes a pixel over the next box, do I select the next box? Is the pole part of the traffic light? And the guy as you say. There is a special place in hell for the inventor of reCaptcha (and for all of Cloudflare staff as fas as I am concerned!)
> Also, when they ask you to identify traffic lights, do you select the post? And when it’s motor/bycicles, do you select the guy riding it?

This type of captcha is too infuriating so I always skip it until I get the ones where I’m just selecting an entire image, not parts of an image

Google’s captchas are too ambiguous and might as well be answered philosophically with an essay-length textbox

There is a doom loop mode where it doesn't matter how many you solve or even if you get them correct. My source for this works on this product at Google.
> do you select the guy riding it? do you select the post?

Just select as _you_ would. As _you_ do.

Imperfection and differing judgments are inherent to being human. The CAPTCHA also measures your mouse movement on the X and Y axes and the timing of your clicks.

Just select the audio option. It's faster and easier. Maybe it's because google doesn't care about training on speech to text. I usually write something random for one word and get the other word correct. I can even write "bzzzzt" at the beginning. They don't care because they aren't focused on training on that data.

Now I think of it, it's really a failure that AI didn't use this and went with guessing which square of an image to select.

No you don’t select the post. No you don’t select the guy. Hence the point. Agreed they are annoying.
The other fun thing is the complete lack of localisation for people not from the US. "Select the squares with crosswalks" - with what? Oh, right, the pedestrian crossings... And the fire hydrants look like we've seen in movies, it's like, oh yeah those do exist in real life!
in other words reasoning call fill the context window with crap
To be honest I'm surprised how well it holds. I expected close-to-total collapse. It'll be a matter of time I guess, but still.
interesting results. why does reload/cross-tile have worse results? would be nice to see some examples of failed results (how close did it to solving?)
So, when do we reach a level where AI is better than humans and we remove captcha from pages alltogether? If you don't want bots to read content, don't put it online, you're just inconveniencing real people now.
(comment deleted)
If not today, models will get better at solving captchas in the near future. IMHO, the real concern, however, is cheap captcha solving services.
Ok and then? Those models were not trained for this purpose.

It's like the last hype over using generative AI for trading.

You might use it for sentiment analysis, summarization and data pre-processing. But classic forecast models will outperform them if you feed them the right metrics.

Would performance improve if the tiles were stitched together and fed to a vision model, and then tiles are selected based on a bounding box?
I hypothesize that these AI agents are all likely higher than human performance now.
Wow. Cross-tile performance was 0-2%. That's the challenge where you select all of the tiles containing an item where the single item is in a subset of tiles. As opposed to all the tiles that contain the item type (static - 60% max) and the reload version (21% max). Seems to really highlight how far these things are from reasoning or human level intelligence. Although to be fair, the cross-tile is the one I perform worst on too (but more like 90+% rather than 2%).
Hcaptcha cofounder here. Enterprise users have a lot of fancy configuration behind the scenes. I wonder if they coordinated with recaptcha or just assume there sitekey in the same as others
This was done on the re-captcha demo page no invisible fingerprinting, behavioral test, or user classification.
Ah yea probably not a good test then. Good point
Forget whether humans can't distinguish your AI from another human. The real Turing test is whether your AI passes all the various flavors of captcha checks.
One of the writers here. We believe the real Turing Test is whether your AI performs a CAPTCHA like a human would/does.
I’ve used LLMs to solve captchas for shits and giggles, just taking a screenshot and pasting it into ChatGPT and having it tell me what squares to click and I think it solves them better than I do.

Can we just get rid of them now, they are so annoying and basically useless.

static, cross-tile and reload. recaptcha call window pings LPRs.
Is it assumed that humans perform 100% against this captcha? Because being one of those humans it’s been closer to 50% for me

I’m guessing Google is evaluating more than whether the answer was correct enough (ie does my browser and behavior look like a bot?), so that may be a factor

To this day I hate captchas. Back when it was genuinely helping to improve OCR for old books, I loved that in the same way I loved folding@home, but now I just see these widgets as a fundamentally exclusionary and ableist blocker. People with cognitive, sight, motor, (and many other) impairments are at a severe disadvantage (and no, audio isn't a remedy, it is just shifting to other ableisms). You can add as many aria labels as you like but if you're relying on captchas, you are not accessible. It really upsets me that these are now increasing in popularity. They are not the solution. I don't know what is, but this aint it.
> In general, all models performed best on Static challenges and worst on Cross-tile challenges.

I also perform poorly on cross-tile, I never know whether to count a tiny bit of a bicycle in a square as "a bike in that square".

At this point i am convinced all captchas almost entirely rely on ip reputation. Even on linux with hardened firefox you can get stuck in a infinite loop with one IP but then switch to another one that let's you in after 0-2 tries.
I'd have a job with the first cross-tile one shown saying select squares with motorcycles. Does the square above the handle bars appearing to maybe contain part of a rear view mirror count? I'm not surprised the LLMs were failing on those.
Is calling Browser Use and "open source framework" a bit misleading it looks like a commercial product that requires an API key to use even if you run the source?