Follow Slashdot blog updates by subscribing to our blog RSS feed

 



Forgot your password?
typodupeerror
Robotics The Internet

Bots Are Better Than Humans At Cracking 'Are You a Robot?' Captcha Tests, Study Finds (independent.co.uk) 78

A recent comprehensive study reveals that automated bots are substantially more efficient than humans at cracking Captcha tests, a widely used security measure on over 100 popular websites. The Independent reports: In the study, scientists assessed 200 of the most popular websites and found 120 still used Captcha. They took the help of 1,000 participants online from diverse backgrounds -- varying in location, age, sex and educational level -- to take 10 captcha tests on these sites and gauge their difficulty levels. Researchers found many bots described in scientific journals could beat humans at these tests in both speed and accuracy.

Some Captcha tests took human participants between nine and 15 seconds to solve, with an accuracy of about 50 to 84 per cent, while it took the bots less than a second to crack them, with up to near perfection. "The bots' accuracy ranges from 85-100 per cent, with the majority above 96 per cent. This substantially exceeds the human accuracy range we observed (50-85 per cent)," scientists wrote in the study. They also found that the bots' solving times are "significantly lower" or nearly the same as humans in almost all cases.

This discussion has been archived. No new comments can be posted.

Bots Are Better Than Humans At Cracking 'Are You a Robot?' Captcha Tests, Study Finds

Comments Filter:
  • by jd ( 1658 ) <imipak AT yahoo DOT com> on Thursday August 10, 2023 @06:01AM (#63755426) Homepage Journal

    Ha! We've finally found a use case for LLM weak AI systems.

  • by quonset ( 4839537 ) on Thursday August 10, 2023 @06:07AM (#63755434)

    Have you seen the type of humans wandering around of late? I believe a chipmunk could do better at these tests than humans.

  • by Anonymous Coward on Thursday August 10, 2023 @06:12AM (#63755436)
    Addition or alternative to captcha make the client mine some crypto or something?

    If the spammers want to spam a million people, the costs could add up. They may use other people's computers for the spamming but the mining might make the owners more likely to notice their stuff is pwned.
  • By design (Score:5, Insightful)

    by micksam7 ( 1026240 ) * on Thursday August 10, 2023 @06:29AM (#63755458)

    Considering so many captchas are being used to feed ML datasets, it's very not surprising AI trained on those datasets can now solve them effortlessly...

  • by greytree ( 7124971 ) on Thursday August 10, 2023 @06:31AM (#63755460)
    Can the latest, most advanced AIs work out how to get through the cookies test and tell the website owners very clearly "NO I DONT WANT ANY FUCKING COOKIES FROM YOU. OF ANY SORT. EVER." ?
    • by Anonymous Coward on Thursday August 10, 2023 @06:54AM (#63755488)
      You could store that choice on the client and then it could tell the website owner everytime you visited the website...
      • by mjwx ( 966435 ) on Thursday August 10, 2023 @08:27AM (#63755604)
        If you've a browser extension that does that it'd be grand.

        Best I've got is I Don't Care About Cookies to kill the popup and Privacy Badger to castrate the cookies themselves. Gets almost all of them.

        What we really need is the ROTW to catch up to the EU and UK with making cookie skulduggery illegal (a "G" GDPR). No, you can't ask me to give up my statutory rights with a popup. Also the irony of American sites doing this is palpable. The EU knows it's own laws don't apply to entities outside it's own borders, they're just doing it to scare the uninformed and stupid against supporting GDPR laws.
        • What we really need is the ROTW to catch up to the EU and UK with making cookie skulduggery illegal (a "G" GDPR).

          If different countries and free trade areas were to adopt their own counterparts to GDPR, in how many different countries would each website operator need to register with a designated local representative pursuant to article 27?

          The EU knows it's own laws don't apply to entities outside it's own borders

          Laws of the Union apply at the border. US-based online stores without a representative pursuant to article 27 could see their shipments turned away at member states' customs.

          • by mjwx ( 966435 ) on Thursday August 10, 2023 @12:28PM (#63756194)

            What we really need is the ROTW to catch up to the EU and UK with making cookie skulduggery illegal (a "G" GDPR).

            If different countries and free trade areas were to adopt their own counterparts to GDPR, in how many different countries would each website operator need to register with a designated local representative pursuant to article 27?

            Erm, maybe they can just stop trying to store people's data and sell it off to other parties. That seems the cheapest solution

            The EU knows it's own laws don't apply to entities outside it's own borders

            Laws of the Union apply at the border. US-based online stores without a representative pursuant to article 27 could see their shipments turned away at member states' customs.

            Erm, a server in a completely different country, no physical presence in the EU... How would they enforce that. You clearly know nothing about the EU if you think that they'll block any site that doesn't register with them. Even in the worst case scenario... why am I still getting anti-EU annoyance popups when I'm in the Americas?

            Also article 27 only applies if you're holding data on EU or UK citizens, residents at al... The simple answer to avoiding A27 is don't. That's the point, don't store or use our data for whatever purpose you feel free to. Again, the EU knows this is effectively unenforceable on any entity outside the EU (the UK has more imperialist desires, but is powerless).

            You clearly have no idea about the GDPR and have fallen victim to their scare mongering. A27 isn't the GOTCHA you think it is. People who want to sell everyone's data with no regard for privacy must love you.

            • Erm, maybe they can just stop trying to store people's data and sell it off to other parties. That seems the cheapest solution

              Say a company operates a web-based store through which it sells something goods, such as toys or physical copies of a work, to customers. I fail to imagine how the company might accomplish that without collecting two pieces of customers' personal data: the billing address and the shipping address. Even a website that doesn't ship physical goods is taking billing addresses once it sells a subscription.

              Erm, a server in a completely different country, no physical presence in the EU... How would they enforce that.

              If a website neither mentions any member state nor quotes prices in euros nor offers a translation of the website in a language primarily spoken in member states, the site probably falls under the recital 23 [gdpr-info.eu] carveout for foreign businesses that don't knowingly target viewers in the Union. But once you take payment for shipment of physical goods, your shipments can in theory be stopped at the border.

              why am I still getting anti-EU annoyance popups when I'm in the Americas?

              The Americas include the U.S. state of California. California passed a privacy law called CCPA, some of whose provisions parallel provisions in the GDPR.

    • by chas.williams ( 6256556 ) on Thursday August 10, 2023 @08:55AM (#63755644)
      But if you don't accept the cookies, the website may or may not work correctly. You want to receive the best possible browsing experience, right? The Internet doesn't work without cookies.
  • Great (Score:5, Insightful)

    by ukoda ( 537183 ) on Thursday August 10, 2023 @06:32AM (#63755462) Homepage
    Can we now get rid of these stupid things? They are really only good for saying "Fuck you" to visitors and makes the site operator's view of you clear. I'm sick of clicking on traffic lights, or worse the broken unsolvable ones.
  • by sometimesblue ( 6685784 ) on Thursday August 10, 2023 @06:39AM (#63755474)
    The article in the Independent just references itself. I've not been able to find the original research on the internet. What 'bots' are being used? If they are a university level research grid, then thats not normally going to be used by the average script kiddy hacker to break a capcha. If a python script trawls the internet and has to break a capcha, then processing time above 10 seconds is going to make the enterprise unfeasible. If an anti-bot measure does require a research grid to defeat it, then that probably still a success.
  • by dfm3 ( 830843 ) on Thursday August 10, 2023 @06:57AM (#63755492) Journal
    My understanding of many current implementations is that they don't just look for a correct answer, but analyze how you solve it - by tracking things like cursor movements or the time between clicks. They also use things like browser fingerprinting. A human, for example, would not be able to click on 4 tiles with exactly 15ms between clicks. I've noticed lately that those "select the traffic lights" puzzles more often than not will accept solutions that are blatantly wrong - maybe I only clicked on one tile per light fixture, or clicked on all the frames that have slivers of poles and wires. That makes me wonder if the tiles I click on don't actually weigh much in the calculation
    • by Megane ( 129182 ) on Thursday August 10, 2023 @07:06AM (#63755506)
      They're apparently also looking for a continuous stream of mouse position data. I had a browser which stuttered in performance because of a problem with the old AdBlock Plus blocking for hundreds of milliseconds every now and then, and the browser didn't put tabs in separate threads. Google's captcha kicked into nightmare mode.
      • by iAmWaySmarterThanYou ( 10095012 ) on Thursday August 10, 2023 @07:58AM (#63755568)

        I'm pretty sure they're not collecting mouse position data on my iPad.

        I have seen times where I realized I missed one after I hit go but I suspect they simply have a "good enough" setting rather than some complex and clever algorithm.

        • by Megane ( 129182 ) on Thursday August 10, 2023 @10:50AM (#63755920)
          Before I gave up completely and went with Firefox, I realized that it only needs for you to click on three or four correct items, so just click on the four fullest squares of those super-zoom images. It's also probably better to not be fast about it, but to take an arced path at a moderate speed between each click. Never mind that you can click them faster than a bot ever could, it cares more about the path you take to click them.
          • by iAmWaySmarterThanYou ( 10095012 ) on Thursday August 10, 2023 @12:05PM (#63756128)

            Interesting idea, thanks. Like everyone, I find them frustrating and useless. Similar to the useless "check here if you're not a bot" but worse.
            My kid had one for some game she was trying to sign up for. Click the animal in the upright position. 4 animals, each successively 90 degrees off. I couldn't do it either. So they stupidly lost some number of sales because they couldn't bother the most basic testing of their captcha clone.

    • by Anonymous Coward on Thursday August 10, 2023 @08:43AM (#63755630)

      My understanding of many current implementations is that they don't just look for a correct answer, but analyze how you solve it - by tracking things like cursor movements or the time between clicks.

      Yes, but that is "part of the solution" regarding the bots solving them.

      They also use things like browser fingerprinting. A human, for example, would not be able to click on 4 tiles with exactly 15ms between clicks.

      A bot, while it would be able to do that, would intentionally not do that either.
      The bot is going to be sending "cursor move" events along with "click" events, while Not moving in a straight line between two points, while also Not moving the cursor at a consistent speed between pixels.
      A bot won't select the same exact point within a tile to click, and will vary the time between the "down button" and "up button" events.
      They also will add some randomness to the order of tiles it plans to select, as if it "missed seeing" one and had to go back for it after a much longer delay than for the others.

      All of those things are considered part of the solution to the puzzle, and are data points that far outnumber choosing which tiles should be selected.

      I've noticed lately that those "select the traffic lights" puzzles more often than not will accept solutions that are blatantly wrong - maybe I only clicked on one tile per light fixture, or clicked on all the frames that have slivers of poles and wires. That makes me wonder if the tiles I click on don't actually weigh much in the calculation

      At one time it was a popular addition to captchas to take feedback from prior solutions to adjust the weighting of the various tiles.

      So an obvious traffic light that nearly everyone clicked on would change in weight to be stronger, along with tiles no one ever clicked on weighted stronger in the other direction.
      Tiles that say half the solvers clicked on would be weighted down to basically not count anymore, to try and account for ambiguity, like the situation you describe (slivers of the described image)

      I know the methodology has long moved on, but I am sure some versions of the software currently in use likely still have that type of weight system in place.

    • by Brain-Fu ( 1274756 ) on Thursday August 10, 2023 @12:40PM (#63756238) Homepage Journal

      Anything that these systems use to try to differentiate between humans and bots can be faked by a bot. Anything you think of, including human weaknesses, can be faked by a bot. There is absolutely no way around this.

      Maybe current implementations don't do some of this faking. That's because they don't need to. The moment they need to, they will be made to.

      The enterprise of making a website try to weed out bots is fundamentally misguided. It can't be done. That ship has sailed. Give it up.

      I can hear the objections already...we just need a webcam so you can see the user's face, watch them type it! We just need a secure key based mechanism that gives the website some kind of superior access to the end-users hardware, to make sure a human is using it and not a program! We just need some more laws!

      It's folly. None of these will work. Nothing you can think of will work. We are in the AI age now. We cannot put that genie back into the bottle, and we cannot make websites determine when they are being accessed by a bot instead of a human. We must adapt our designs such that it simply doesn't matter if a bot is doing the work. We must build a world that tolerates bots, and still functions well enough. That's the shape of things to come.

      • by narcc ( 412956 ) on Thursday August 10, 2023 @04:02PM (#63756946) Journal

        We are in the AI age now.

        We're not in the "AI age". That's absurd. We're in the middle of a hype cycle, and it looks like we're already past the peak. We've been using ML methods to defeat captchas since they first appeared. It's always been an arms race and we've had bots that outperform humans on various types for almost as long as we've had captchas.

        The study [arxiv.org] is comparing human performance (from Mechanical Turk) across a variety of captcha types to bot performance reported in other, sometimes much older, studies. Whatever "advancements" you seem to think we've made in the past couple years are not at all responsible for the results they reported.

    • by Anonymous Coward on Friday August 11, 2023 @01:42AM (#63758098)

      I'm also doing my part selecting at least 1 wrong answer whenever I can. One time I got the captcha into hardcore never-ending mode. I think it identified me as bot and there was literally no way out even if I selected all the correct answers multiple times.

  • by tinkerton ( 199273 ) on Thursday August 10, 2023 @07:34AM (#63755546)

    I have this plugin which helps me fill in the captcha, even does it for me. It's only the start though, the next generation of captchas requires more advanced tools to answer them and guess who is making and selling those tools? The same guys who are making the captchas.

  • by HnT ( 306652 ) on Thursday August 10, 2023 @07:39AM (#63755554)

    Captchas have become so terrible and annoying, I am starting to question my humanity.

  • by Barny ( 103770 ) on Thursday August 10, 2023 @07:58AM (#63755566) Journal

    The linked site contains no link to the study, nor does the /. summary.

    I guess this is that faith based reporting we hear so much about. Which captchas were tested? Which bots? All the linked site has is links to their own page for keywords that I am sure boost search ranking.

  • by Luca01 ( 10301811 ) on Thursday August 10, 2023 @08:00AM (#63755572)
    We just need to negate the test outcome...if you aced it then you are a bot, if you fail badly then you are a human. FTFY!
  • by dariuscardren ( 826733 ) on Thursday August 10, 2023 @08:04AM (#63755578) Journal
    I've been using a browser extension to check I am not a robot for years now, so I am not surprised
  • by Eunomion ( 8640039 ) on Thursday August 10, 2023 @08:05AM (#63755582)
    It's just a racket to avoid paying people. So they put up these "information toll-booths" between you and your regular sites, taxing your time and patience to train their algorithms. They don't suspect you're a robot; they know for a fact you're not.
  • by Spacejock ( 727523 ) on Thursday August 10, 2023 @08:46AM (#63755634)
    I normally have no issue solving captchas, but recently went through page after page of the things, failing every one. 'Select all motorbikes' it says. Does that include the tiny piece of the wing mirror or not? A sliver of the back tyre I can clearly see in the next square? Wait, is that a moped or an ebike?
    Eventually I gave up and chose the audio option, which I solved first time.
  • by bsdetector101 ( 6345122 ) on Thursday August 10, 2023 @08:49AM (#63755640)
    Sometimes the pictures in the blocks are blurry or just slightly over lap and hard to tell for sure. Or get laggy in responding in what you clicked on ! Just give me the box that says I am human, click here. If bots can do better, what's the point !
  • by Petersko ( 564140 ) on Thursday August 10, 2023 @09:24AM (#63755714)

    I wish them bodily harm.

    A grid of cells with, "select all the cells containing motorcycles", when sometimes a sliver of a helmet is part of it and sometimes it isn't, is a dark pattern. Whomever set that up needs to be beaten with sticks.

    They absolutely offer captchas that increase in ambiguity to block content distribution. It's shady as hell.

    • by hudsucker ( 676767 ) on Thursday August 10, 2023 @11:08AM (#63755948)
      But without this captcha, Waymo's driverless cars won't know if it is seeing a sliver of a helmet of a motorcycle rider about to dart into traffic in front of the car, or is just the top of a stationary parking meter and can be ignored.
    • by thegarbz ( 1787294 ) on Thursday August 10, 2023 @12:27PM (#63756186)

      I wish them bodily harm.

      Do you routinely attack the people trying to help rather than the cause of the problem? Do you hate an ambulance instead of a terrorist? A police officer instead of the robber?

      Captchas exist to solve an underlying problem. Hating on them rather than those who cause the issue in the first place is one of the most dumbfucking stupid things I've read on the internet this week.

      • by Petersko ( 564140 ) on Thursday August 10, 2023 @01:57PM (#63756554)

        " one of the most dumbfucking stupid things I've read on the internet this week."

        Glad to help. Hope you found it entertaining.

        In case it wasn't clear, my complaint was not about trying to find a solution to the problem. It's that the solution doesn't confine itself to that. It's because what they've built is unworkable, and used for shady purposes - like making it difficult to access the thing you have - in some cases - paid for. This is a conscious choice, I guarantee it. You can ask them to make it exasperating to pierce the veil, and that's what they'll give you.

  • by Anonymous Coward on Thursday August 10, 2023 @10:20AM (#63755848)

    Websites simply need to adjust their logic so users that fail their captcha are assumed human, and those that solve their captcha are assumed bots.

  • by ze_jua ( 910531 ) <jailhNO@SPAMfree.fr> on Thursday August 10, 2023 @10:35AM (#63755880)
    We are using Friendly Captcha. It requires the client to provide a POW (1 or 2 seconds of CPU). If you are a bot, you request several puzzles, and the Captcha service send more and more complex crypto-puzzles, using all your CPU time.
  • by bill_mcgonigle ( 4333 ) * on Thursday August 10, 2023 @10:41AM (#63755894) Homepage Journal

    So if they solution takes 3 seconds it's not a human.

    That seems too obvious. I can't believe this work.

  • by WDot ( 1286728 ) on Thursday August 10, 2023 @10:58AM (#63755934)
    I gotta say, some of the most recent Captcha tests are pretty challenging for a human. I had one where I had to distinguish between greyscale pictures of vinyl records and greyscale pictures of rolled up fire hoses, and I had to think for a few seconds. Even some of the ones that seem intuitive like “pick the images that are bridges,” I have to spend time thinking whether this metal strip is a railing for a bridge or just a short strip of fence near a sidewalk. I’ve even seen a few of late where the captcha images are clearly AI-generated scenes, and I have to click the location in the image where there is a “parrot” that looks like the kind of uncanny valley AI-generated parrot image from the few months before Midjourney and Stable diffusion got really good.
  • by neBelcnU ( 663059 ) on Thursday August 10, 2023 @11:15AM (#63755960) Journal

    How is this NOT an Onion post?!

  • by Tablizer ( 95088 ) on Thursday August 10, 2023 @11:49AM (#63756046) Journal

    I'm thinking of making a T-shirt saying, "Yes, I'm a robot! Whaddya gonna do about it!?"

  • by dddux ( 3656447 ) on Thursday August 10, 2023 @11:50AM (#63756054)

    I make mistakes on purpose when solving captcha tests, to find out how many mistakes you can make for captcha to be accepted. Like, when you get 2 words "gimme shelter", I enter "giemm shetler". That can pass sometimes, but entering "gimme shetler" will always pass. Same with pictures - you don't have to click on all the pictures, just a couple and even some wrong ones, to pass the test. You should try it sometimes. It's fun. I guess this also makes machine learning harder, and us humans look more stupid than we are. :)

  • by Applehu Akbar ( 2968043 ) on Thursday August 10, 2023 @01:42PM (#63756472)

    One of the major commercial applications for today's AIs is interpreting digital scans. AIs that are getting good at finding tumors in a noisy mammogram are going to be really good at identifying fuzzy letters and numbers in a CAPTCHA.

    Meanwhile, you as puny human get shown endless grids of motorcycles and traffic signals, and you will fail every time because they CAPTCHAs are now beyond human pattern matching ability. Time to retire the entire idea.

    • by narcc ( 412956 ) on Thursday August 10, 2023 @04:47PM (#63757106) Journal

      It's always been an arms race. We've had bots capable of defeating captchas almost as long as we've had captchas. That doesn't mean we need to scrap the whole idea, only that we likely won't have a long-term solution. Remember that 'bots' can't just magically adapt to new challenges and that training takes a great deal of time and effort.

      While the study compares human performance on a few different kinds of captchas (from Mechanican Turk) to bot performance reported in other, sometimes much older, studies, that's not really what it's about. It's about human performance and perceptions, not advances in captcha solving bots. The headline completely misrepresents the purpose and results of the study.

  • by sgunhouse ( 1050564 ) on Thursday August 10, 2023 @03:26PM (#63756858)
    If humans take longer to solve the Capcha, that is a good way to idrntify them isn't it?
  • bots aren't affected by problems that aging eyesight causes in solving captcha's

  • by belg4mit ( 152620 ) on Thursday August 10, 2023 @05:53PM (#63757332) Homepage

    Is success getting though the stupid gate, or accurately performing the task? I still run into plenty of captchas were system has misinterpreted markings on the side of the road as as a crosswalk, misses half of the stop-lights in a scene, etc. And I therefore end up having to answer more of the stupid things for doing them right.

  • by manu0601 ( 2221348 ) on Thursday August 10, 2023 @05:59PM (#63757344)
    Now a site should assume the remote user is human if it fails the test.
  • by EdZep ( 114198 ) on Friday August 11, 2023 @07:24AM (#63758448)

    I have always thought that speed and pace of clicking were being evaluated along with accuracy... such that solving too quickly would indicate a 'bot, and fail the test.

  • by sudonim2 ( 2073156 ) on Friday August 11, 2023 @07:46AM (#63758488)

    1,000 participants online from diverse backgrounds

    This means they put it on Fiver and/or Mechanical Turk.

"Why can't we ever attempt to solve a problem in this country without having a 'War' on it?" -- Rich Thomson, talk.politics.misc

Working...