Benchmarking leading AI agents against Google reCAPTCHA v2

https://research.roundtable.ai/captcha-benchmarking/

by mdahardy • 8 months ago

124 97 8 months ago

research.roundtable.ai

PaulHoule• 8 months ago

I know people were solving CAPTCHAS with neural nets (with PHP no less!) back in 2009.

golfer• 8 months ago

Indeed, captcha vs captcha bot solvers has been an ongoing war for a long time. Considering all the cybercrime and ubiquitous online fraud today, it's pretty impressive that captchas have held the line as long as they have.

mdahardy8• 8 months ago

You could definitely do better than we do here - this was just a test of how well these general-purpose systems are out-of-the-box

xnx• 8 months ago

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.

bena• 8 months ago

Makes sense, what do you think it was trained on?

mdahardy8• 8 months ago

After watching hundreds of these runs, Gemini was by far the least frustrating model to observe.

dgacmu• 8 months ago

In my admittedly limited-domain tests, Gemini did _far_ better at image recognition tasks than any of the other models. (This was about 9 months ago, though, so who knows what the current state of things). Google has one of the best internal labeled image datasets, if not the best, and I suspect this is all related.

Xenoamorphous• 8 months ago

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?

datadrivenangel• 8 months ago

That's not due to accuracy, you're getting tarpitted for not looking human enough.

sixhobbits• 8 months ago

Didn't look a lot into this but I think the fact that humans are willing to do this in the "cents per thousand" or something range means that it's really hard to get much interest in automating it

Semaphor• 8 months ago

There's a browser extension to solve them. Buster.

mdahardy8• 8 months ago

While running this I looked at hundreds and hundreds of captchas. And I still get rejected on like 20% of them when I do them. I truly don't understand their algorithm lol

Sayrus2• 8 months ago

Testing those same captcha on Google Chrome improved my accuracy by at least an order of magnitude.

Either that or it was never about the buses and fire hydrants.

ACCount37• 8 months ago

It's a known "issue" of reCaptcha, and many other systems like it. If it thinks you're a bot, it will "fail" the first few correct solves before it lets you through.

The worst offenders will just loop you forever, no matter how many solves you get right.

viccis• 8 months ago

That's because Chrome tracks so much telemetry about you that Google is satisfied with how well it has you surveilled. If you install a ton of privacy extensions like Privacy Badger, uBlock, VPN extensions with information leakage protections, etc., watch that "accuracy" plummet again as it makes you click 20 traffic signals to pass one check.

timshell3• 8 months ago

Yeah, we've looked at it in the context of reCAPTCHA v3 and 'invisible behavioral analysis': https://www.youtube.com/watch?v=UeTpCdUc4Ls

It doesn't catch OpenAI even though the mouse/click behavior is clearly pretty botlike. One hypothesis is that Google reCAPTCHA is overindexing on browser patterns rather than behavioral movement

utopman• 8 months ago

Not sure it is your case but I think I sometimes had to solve many of them when I am in my daily task rush. My hypothesis is that I solve them too fast for "average human resolving duration" recaptcha seems to expect (I think solving it too fast triggers bot fingerprint). More recently when I fall on a recaptcha to solve, I consciently do not rush it and feel have no more to solve more than one anymore. I don't think I have super powers, but as tech guy I do a lot a computing things mechanically.

IG_Semmelweiss• 8 months ago

that, and VPN.

Yes.

hnburnsy• 8 months ago

Pro tip, select a section you know is wrong, then de select it before submitting. Seems to help prove you are not a bot.

sph• 8 months ago

Shhh, you're not supposed to tell people. Now they'll patch it and I'll have to select stairs and goddamn motorcycles 4 times in a row.

pants2• 8 months ago

Another pro tip: the audio version of the captcha is usually much easier / faster to solve if you're in a quiet environment

felixfurtak• 8 months ago

I always assume that people are lazy and try and click the least amount of squares as possible to get broadly the correct answer. Therefore, if it says motorbikes just click on the body of the bike and leave out rider and tiles with hardly any bike in them.

If it says traffic lights just click on the ones you can see lit and not the posts and ignore them if they are too far in the distance. Seems to work for me.

cm21872• 8 months ago

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!)

charcircuit2• 8 months ago

It doesn't matter. Select it of you think other people would select it too.

dfxm12• 8 months ago

There's not a right or wrong answer. They're just building consensus for self driving vehicles.

jameslk2• 8 months ago

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

terminalshort2• 8 months ago

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.

Tostino2• 8 months ago

Tell them I hate them.

max512• 8 months ago

That doesn't surprise me. I find it hard to believe it's a pure coincidence that I would get stuck in the loop regularly when I'm on the university wifi but it would never happen anywhere else ever. After a dozen try, I would remote connect to my home pc and it would magically work on the first try every single time.

jgalt212• 8 months ago

Firefox and Ubuntu can reliably trigger this mode.

plingbang• 8 months ago

I have a recording of me trying to pass the captcha for straight 5 minutes and giving up. To be fair, this has only happened once.

What is the purpose of such loop? Bots can simply switch to another residential proxy when the captcha success rate gets low. For normal humans, it is literally "computer says no".

perfmode• 8 months ago

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.

nistiminic• 8 months ago

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.

nwellinghoff• 8 months ago

No you don’t select the post. No you don’t select the guy. Hence the point. Agreed they are annoying.

stephen_g• 8 months ago

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!

guluarte• 8 months ago

in other words reasoning call fill the context window with crap

flakiness• 8 months ago

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.

mdahardy8• 8 months ago

Same! As we talk about in the article, the failures were less from raw model intelligence/ability than from challenges with timing and dynamic interfaces

swyx• 8 months ago

i mean did you see the cross-tile numbers

criddell• 8 months ago

I wonder if any of the agents hit the audio button and listened to the instructions? In my experience, that can be pretty helpful.

WhereIsTheTruth2• 8 months ago

3 models only, can we really call that a benchmark?

mdahardy8• 8 months ago

yes

WhereIsTheTruth2• 8 months ago

Ignorance Is Bliss

maknee• 8 months ago

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?)

mdahardy8• 8 months ago

We have an example of a failed cross-tile result in the article - the models seem like they're much better at detecting whether something is in an image vs. identifying the boundaries of those items. This probably has to do with how they're trained - if you train on descriptions/image pairs, I'm not sure how well that does at learning boundaries.

Reload are challenging because of how the agent-action loop works. But the models were pretty good at identifying when a tile contained an item.

Youden• 8 months ago

I'm also curious what the success rates are for humans. Personally I find those two the most bothersome as well. Cross-tile because it's not always clear which parts of the object count and reload because it's so damn slow.

ajsnigrutin• 8 months ago

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.

cubefox• 8 months ago

They can also sign up and post spam/scams. There are a lot of those spam bots on YouTube, and there probably would be a lot more without any bot protection. Another issue is aggressive scrapers effectively DOSing a website. Some defense against bots is necessary.

1gn15• 8 months ago

I use manual verification on first post (note: can probably use an LLM to automate this), or just not have a comment section in the first place. That way, you moderate based on content, not identity or mental ability (which can be discriminatory and is also a losing game, as seen in TFA).

Either that, or just be honest and allow anonymous posting lol

kjok2• 8 months ago

If not today, models will get better at solving captchas in the near future. IMHO, the real concern, however, is cheap captcha solving services.

arbol2• 8 months ago

The solvers are a problem but they give themselves away when they incorrectly fake devices or run out of context. I run a bot detection SaaS and we've had some success blocking them. Their advertised solve times are also wildly inaccurate. They take ages to return a successful token, if at all. The number of companies providing bot mitigation is also growing rapidly, making it difficult for the solvers to stay on top of reverse engineering etc.

kjok2• 8 months ago

when they incorrectly fake devices

And how often does this happen? Do you have any proof? Most YC companies building browser agents have built-in captcha solvers.

mehdibl• 8 months ago

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.

Legend2440• 8 months ago

It is relevant because they are trained for the purpose of browser use and completing tasks on websites. Being able to bypass captchas is important for using many websites.

It would be nice to see comparisons to some special-purpose CAPTCHA solvers though.

bagacrap• 8 months ago

And more broadly, if an agent is supposed to do everything a human can on the web, its ability to solve a captcha is likely a decent litmus test.

daveguy2• 8 months ago

These are all multi-modal models, right? And the vision capabilities are particularly touted in Gemini.

https://ai.google.dev/gemini-api/docs/image-understanding

cedws• 8 months ago

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?

mdahardy8• 8 months ago

That's a cool idea. I bet it would work better.

jngiam1• 8 months ago

I hypothesize that these AI agents are all likely higher than human performance now.

daveguy2• 8 months ago

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%).

RobertDeNiro• 8 months ago

I think the prompt is probably at fault here. You can use LLMs for object segmentation and they do fairly well, less than 1% seems too low.

mdahardy8• 8 months ago

The cross-tile challenges were quite robust - every model struggled with them, and we tried with several iterations of the prompt. I'm sure you could improve with specialized systems, but the models out-of-the-box definitely struggle with segmentation

alexnewman2• 8 months ago

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

amirhirsch• 8 months ago

This was done on the re-captcha demo page no invisible fingerprinting, behavioral test, or user classification.

alexnewman2• 8 months ago

Ah yea probably not a good test then. Good point

rkagerer• 8 months ago

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.

timshell3• 8 months ago

One of the writers here. We believe the real Turing Test is whether your AI performs a CAPTCHA like a human would/does.

theoldgreybeard• 8 months ago

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.

throwawayu5pg• 8 months ago

static, cross-tile and reload. recaptcha call window pings LPRs.

jameslk2• 8 months ago

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

padolsey• 8 months ago

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.

TulliusCicero• 8 months ago

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".

akimbostrawman• 8 months ago

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.

tim333• 8 months ago

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.

VectorLock• 8 months ago

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?

Benchmarking leading AI agents against Google reCAPTCHA v2

Related Stories