Success rate of 89.9%?
That's better than I manage with these bloody things. And don't get me started on looking for shop signs or vehicles in some random high street picture.
Computer software that mimics how the human visual cortex works can solve text-based CAPTCHA challenges, the image recognition tasks often used by websites to differentiate human visitors from spam bots. A paper describing the code was published on Thursday in the journal Science. A team from Vicarious – a California startup …
Wait, there will be an app?!? PLEASE, you can have all my money / first-born child / kingdom I don't even have, ANYTHING, just let me have it! These days most things pester me with multiple Captchas daily and I keep near-failing them all! The fairly recent ones (looking like torn scraps of a grainy photo of something that may or may not have been letters viewed in a mirror shattered into a thousand pieces on the bottom of a stormy lake - absolutely NOTHING like the examples accompanying this article) were only moderately maddening, but the current "tile" based ones that replaced them have well and truly pushed me off the deep end. I swear these bastards are allowed to keep throwing up new image after image after image after image after image after image and hold you hostage to indefinite lengths that Gitmo wardens can only dream of, without ever declaring you either human or machine (and in the end they declare you a bot anyway and from then on there's flat out nothing you can do to log in again in the foreseeable future - well, except clear all cookies and PPPoE redial to change your IP). Sensei, please, I'll do anything! ANYTHING...!
These days most things pester me with multiple Captchas daily and I keep near-failing them all!
I find blocking the scripts for google, clodfool etc tends to stop those ones coming up. I can go for days without seeing on on El Reg, but if I allow google for a moment for some other page I seem to get their horrible ones again. Over and over and over and over. I think if you make a few mistakes, after 100,000 images they decide a decent AI would get 100% but a human would only manage 99.999%, but they have to run a decent enough sample size.
These days after the 3rd of 4th run I decide I'm no longer interested in the site and find something else to do for the day.
--> Mr capthca maker, would you like to pay me a visit? I have a lovely blowtorch I'd love you to meet.
I seem very good at those. But then I have a very visual memory, near photographic and have spent much of my scientific career looking for tiny differences between very tiny things.
Also with those you are actually helping a deep learning system by categorising scene objects for them.
That system which can create novel near photo-realistic street scenes was trained on such CAPTCHAs. Thus I prefer them to the munged letters. At least you are doing something useful.
I seem very good at those.
Then slip this leash on, and you can be my captcha hound.
However, from a personal perspective I have a slight image problem with a Muscleguy on my dogleash. I could more comfortably manage a Musclebabe on a leash. Or better, I could be on a Musclebabe's leash.......no, I didn't write that, forget you saw that.
Move along, move along, nothing to read here.
A) This is old news
B) "Training" (Really adding a to a database) with loads of examples isn't real AI. Real humans don't need loads of examples. Three year old human vs computer "learning" what a hot dog is.
C) OCR once needed special letters, then by 1976 Ray Kurzweil had got OCR to work with almost any font (omnifont). Then he proceeded to build a machine that with synthetic voice could read books to the blind.
As computers have got better at captchas (a stupid name), they made them harder to the point where humans find them difficult. The ones on eBay are horrid.
It was absolutely inevitable that further development of OCR would make the captcha useless. It's not AI. Just incremental development of specialist algorithm with a larger database. Google has been using the captcha system* to improve their OCR on their Google Books scanning program (which initially was very very poor compared to human proof read OCR teams of Project Gutenberg).
So nothing surprising about this.
[* Google only offers something if they get benefit, in this case improvements in OCR of real books]
Found it XKCD Constructive Approx 2013
So regarded as inevitable 4 years ago?
Less relevant, but related XKCD Suspicion
Note test is called VK
Real humans don't need loads of examples. Three year old human vs computer "learning" what a hot dog is.
Yeah, they kinda do. Letters are especially hard compared to physical objects.
No matter how you orient a hotdog in space it's still a hotdog but letters are not like that.
Turn the letter "b" upside down and now it's a "p" - mirror it and it's a "d" or "q". It's part of why children learning to write will sometimes render letters backwards or in other strange ways.
I hate those stupid damn Captcha boxes. It seems like if bots are such a problem, then they could figure out something better than a Captcha box to discover it. I have seen Captcha boxes that were so bad, that NONE of 30 people could figure out what it was.
And the street scenes? How many do I have to look at? That dot of a sign that barely registers 8 pixels by 8 pixels might well be obvious to someone,but not me.
This post has been deleted by its author
I keep hearing about how Captchas are broken. But still they keep using the wretched things.
If the site gets so many visitors they can't evaluate a new user (clue: it's only necessary to moderate the first few posts), can't they afford a few more staff ?
And the street scenes .,, they're crap, too. Who evaluates them to find the correct answer ? A bot, I guess. Well, duh.
You clearly don't run any public sites. It's NOT just about protecting user registration after the fact, there is a lot more to it than that. Consider just the two scenarios:
Your website has a public registration form. Without a captcha any vaguely known about website will easily receive hundreds of spam/malware/whatever bot account registrations every day. A remotely popular website will receive considerably more. As a result bot accounts will be created, advert/spam/malware links will be posted all over your website forums (or whatever else it has). Fine, you could remove these after the fact but the reputation of your website will rapidly drop to Daily Mail levels with even a few malware infections and spam/advert/link bot attacks generated by these bots. You could choose to manually approve new user accounts but then you have the entertaining task of trying to identify the gennuine account requests among the hundreds of spam-bot account requests.
Your website has a "contact us" form for visitors to send messages to you using. Without a captcha, for a vaguely popular site, you will receive many bot-messages every day. This is particularly the case if they find that the content of the message is CC'd to the "sender" - in which case congratulations, you are sending spam on their behalf. Even if the content of the message is not CC'd to the "sender", it's a fine way to clutter up inboxes, perform email floods and to flatten the reputation of your organisation's email servers.
If you cc the message back to the sender you deserve to be penalised.
Why? If I wish to contact a company for something, and am stuck with their contact form rather than a proper email, why should I not get a copy of whatever I sent them?
Why should they be penalised for giving me a copy of what I sent them so I have a record of things?
Because you could put anybody's email address in the "your email address" field and effectively send them an anonymous (spam) message just using the website's "contact us" form. This is why spam-bots target these things because they want to see if they can use them as an anonymous spam relay.
Because you could put anybody's email address in the "your email address" field and effectively send them an anonymous (spam) message just using the website's "contact us" form.
The better ones have a captcha of some sort to defeat (or at least significantly rate-limit) the spammers.
So. Why should sites be punished for giving me a way to use their contact form and still have a record of what I sent?
Because they hate you? :) Seriously though, I can't think of any reason for many of those other than to try and prevent automated submissions after a manual login.
The ones I tend to fail on are the picture ones where you have to identify all of the damn squares that have a smallest part of a mountain, street sign, store front or whatever in them.
math puzzle...
well, if you want to confuse the AI, list dog breeds and one cat breed, like
"Mark has a dachsund, Jean has two border collies, Bill has a rottweiler, Ron has a German shepherd named Killer, and Velma has a tabby. How many dogs are present?"
The answer will be "zero, they all ran after the cat".
"Yup, all we need is replace this stupid s##t with even stupider s##t like "30 cows in a field, 28 chickens how many didn't?" then wonder why the correct answer is "10"...:"
This went around my office a few weeks ago. I happened to be the only one who got it right. Doesn't bode well for puzzle-challenges.
Computers will eventually be better than humans at ANY of this, and inevitably the challenges will become unusable to humans but easily solved by machines. So what's left? Something involving body fluids? "Press the Auto-Pierce pad to collect a drop of blood and verify you are a human. We're sorry, your challenge failed, please re-sample and try again."
Meantime 'Feedback' when you're progressing through multiple captcha stages would be nice. Its like typing in an asterick-obscured password field that's 20+ characters long. Why don't all sites offer a 'show typing' option? Even the guy who invented obscuring thinks it was all a mistake!
Anyone familiar with the alphabet can solve one. I switched to a Q&A on my board and the number of would-be spammer registrations plummeted. I moderate all new registrations anyway, but the switch greatly reduced my workload.
The ones on Google (which you see if you click through search results too quickly), where you have to ID street signs, cars or buildings are particularly annoying. FFS Google, so what if it's a robot? It's not as if Google doesn't use bots to scrape my site.
Everything done by technology is now called "AI" I expect that applies to my washing machine adjusting the amount of water it uses for a smaller load, upwards.
Do you recall in th 90's when everything was "Turbo" and even computers had a useless button to prove it? This is history repeating itself with the next hype bubble.
Nevertheless, thumb up for not becoming assimilated :)
Once could argue that Natural Intelligence is nothing more than pattern recognition with probabilistic generative models. One of the reasons why people see faces in so many things.
The core of NI is that we can recognize patterns of patterns (like noticing how many "faces" we, as a species, see in things that have no faces) and make predictions based on this recognition, including recognizing the pattern of how our predictions fail, and updating our models accordingly.
Am I the only one who deliberately does a few of the street sign things wrong?
In protest at how irritating they are (poor quality images, made worse viewed on a tiny phone screen) - and make them realize they are not fit for purpose - with the added bonus of their self driving car "AI" gets less useful data to work with.
I run my browser as locked-down as I can manage it. I block scripts, spying, Faceborg "like" spy-buttons, etc. It might be my imagination, but I feel like more web sites are tying their CAPTCHA challenges to the basic site scripts. Block anything and you've blocked the gatekeeper challenge as well and can't use the site at all. Could be coincidence, could be my paranoia, could be a real thing.
"Prove you're not a robot, and let us track you for profit."