the bendy letter blockade

[parallaxicality]

Whenever I want to revel in human superiority over machines, I remember that no artificial intelligence yet design can read handwriting, or even, for that matter, those bendy letters and numbers that sites put up to protect themselves from bot intruders.

But it does raise an interesting question. What is it that allows us to look at a bendy a, or an a in someone else's handwriting, and yet recognise it as as a? How are we able to abstract the Platonic ur-a from all the various outward manifestations of said letter we encounter in our lives? If we could design a bot that could somehow abstract in this way, we'd have cracked a major AI issue

[jlhredshift]

Whenever I want to revel in human superiority over machines, I remember that no artificial intelligence yet design can read handwriting, or even, for that matter, those bendy letters and numbers that sites put up to protect themselves from bot intruders.

But it does raise an interesting question. What is it that allows us to look at a bendy a, or an a in someone else's handwriting, and yet recognise it as as a? How are we able to abstract the Platonic ur-a from all the various outward manifestations of said letter we encounter in our lives? If we could design a bot that could somehow abstract in this way, we'd have cracked a major AI issue

I think it is because it is not binary. It is analog and analogue thinking that allows us to see a choo choo in a cloud, a bear in an asterism, or a clock in a Salvador Dali painting.

[01101001]

But it does raise an interesting question. What is it that allows us to look at a bendy a, or an a in someone else's handwriting, and yet recognise it as as a?

Douglas R Hofstadter: on seeing A's and seeing As

To show the fearsome complexity of the task of letter recognition, I offer the following display of uppercase "A"s, all designed by professional typeface designers and used in advertising and similar functions.
[...]
Indeed, I once even proposed that the toughest challenge facing AI workers is to answer the question: "What are the letters 'A' and 'I'?"

Edit: especially for searchers: the keyword here, what a "bendy letter blockade" is probably better known as, is "CAPTCHA" awfully acronymized from "Completely Automated Public Turing test to tell Computers and Humans Apart" (Wikipedia: CAPTCHA).

[Torsten]

I am impressed by this ability too. I am really challenged to decipher some people's handwriting, and yet, by context, I can usually make sense of it. Beyond that though, I don't know why, other than it has something to do with pattern recognition.

The legibility of my handwriting varies considerably. Some of my notes will be done in block letters and they look like they they were printed with some sort of stencil - very consistent letter shapes and sizes, and possibly machine-readable.

But then there are times when I'm in a real hurry, and they look like chicken scratch. For example, my shorthand for words ending with "ation" is to simply replace those letters with a line that may be straight, or wavy.

situ____

or

situ~~~

It would be a good program that deciphers that.

ETA: Looks like an interesting article, 01~~~~

[Celestial Mechanic]

[Snip!] What is it that allows us to look at a bendy a, or an a in someone else's handwriting, and yet recognise it as as a? How are we able to abstract the Platonic ur-a from all the various outward manifestations of said letter we encounter in our lives? If we could design a bot that could somehow abstract in this way, we'd have cracked a major AI issue.

Not to mention we'd have created a tool that will assure everyone of receiving more spam. Careful what you wish for!

SPAM^TM is a trademark owned by Hormel and the name of a product that helped America win World War II in the Pacific theater.

[Celestial Mechanic]

But back to the topic, seriously! I am reminded of something that I saw recently. It was a display in a store window for a brand of pizza called Phillipe's. The first letter was a line drawing of a chef's hat. I think that after seeing the rest of the name, ...hillipe's and from the context I knew the hat had to also be the letter 'P'. That's two things that computer programs are bad at handling, context and ambiguity. The first could be taken care of with dictionary searching. The second cannot be anticipated in advance; chef's hats are not in the domain of character recognition programs. The program would have to "hypothesize" a 'P' and see if it could be found within the hat.

[Gillianren]

Spam^TM is a trademark owned by Hormel and the name of a product that helped America win World War II in the Pacific theater.

Actually, they put it in all caps. SPAM. If you don't, it is clear that you mean the irksome e-mail that you get unprompted. I believe they've sued; I believe they've lost.

[HenrikOlsen]

Whenever I want to revel in human superiority over machines, I remember that no artificial intelligence yet design can read handwriting, or even, for that matter, those bendy letters and numbers that sites put up to protect themselves from bot intruders.

There's an arms-race going between CAPTCHA designers and designers of OCR software to break it.
It looks like the current state is that if a human gets it right every time, the machine can break it.
To avoid automated bypassing, they currently have to make the text so hard to read that I fail half the time.
Don't stop to pat your back too much, the machines are not far behind you.

[01101001]

There's an arms-race going between CAPTCHA designers and designers of OCR software to break it.

Ars Technica: Gone in 60 seconds: Spambot cracks Live Hotmail CAPTCHA

A security firm is now reporting that the CAPTCHA used for Windows Live Mail can now be cracked in as little as 60 seconds.

Back in early February, a group cracked Windows Live Hotmail's CAPTCHA. A few weeks later, Gmail's version followed suit.
[...]
Microsoft, Google, and all other websites that currently use CAPTCHA, need to find a solution that puts them a step ahead of the spammers. Using better images and improving CAPTCHA will simply prolong the arms race.

[George]

Nxet, I wodner if word recgonition will be aded to the crypotgrahps. Then comes faces, no doubt.

[Neverfly]

Nxet, I wodner if word recgonition will be aded to the crypotgrahps. Then comes faces, no doubt.

Those would drive Gillianren wild.

[Gillianren]

What, the spelling errors? Other than that, I'm not sure what you're talking about.

ETA: Oh. I get it now. All they'd have to do is run spellcheck, wouldn't they? It catches those errors.

[Fazor]

Wouldn't a better "CAPTCHA" be one that gives you words to desipher, but the "key" is actually something infered by the words.

Like the image cypher is "Red Green Blue"
and the key is "The words listed are all ______" (key = colors)

I mean, even if a bot decyphered the image, then it'd have to read the logic question and make the inference.

But it doesn't matter. There's companies now that have rooms full of people that do nothing but answer the CAPTCHA's for the bots, for a fee. Spam is big business on both sides of the sword.

[Neverfly]

Wouldn't a better "CAPTCHA" be one that gives you words to desipher, but the "key" is actually something infered by the words.

Like the image cypher is "Red Green Blue"
and the key is "The words listed are all ______" (key = colors)

I mean, even if a bot decyphered the image, then it'd have to read the logic question and make the inference.

But it doesn't matter. There's companies now that have rooms full of people that do nothing but answer the CAPTCHA's for the bots, for a fee. Spam is big business on both sides of the sword.

Because dumb people like me wouldn't be able to figure it out.

[hhEb09'1]

Whenever I want to revel in human superiority over machines, I remember that no artificial intelligence yet design can read handwriting, or even, for that matter, those bendy letters and numbers that sites put up to protect themselves from bot intruders.

I've used tablet PC software (on Toshibas, HPs, Lenovas) for a few years. I can take notes in handwriting, save them, and the search feature can find all instances of where I wrote a particular (typed) word or phrase, fairly reliably, without much training. I was impressed.

[tdvance]

Wouldn't a better "CAPTCHA" be one that gives you words to desipher, but the "key" is actually something infered by the words.

Like the image cypher is "Red Green Blue"
and the key is "The words listed are all ______" (key = colors)

I mean, even if a bot decyphered the image, then it'd have to read the logic question and make the inference.

But it doesn't matter. There's companies now that have rooms full of people that do nothing but answer the CAPTCHA's for the bots, for a fee. Spam is big business on both sides of the sword.

1. captcha has to be auto-generated.
2. a bot leveraging GOOGLE would beat it (I'm thinking of the bot that could solve 80% of crossword puzzle entries using Google).

[George]

What, the spelling errors? Other than that, I'm not sure what you're talking about.

ETA: Oh. I get it now. All they'd have to do is run spellcheck, wouldn't they? It catches those errors.

A computer would be confused by the spelling errors, but a human has far superior pattern recognition, and will often not be confused.

The Galaxy Zoo folks have much to say about this since they have advanced leaps and bounds using human pattern recognition over computer pattern recognition.

[parallaxicality]

So I take it the Turing test is full of it?

[hhEb09'1]

I dunno, I failed it

[Fazor]

1. captcha has to be auto-generated.
2. a bot leveraging GOOGLE would beat it (I'm thinking of the bot that could solve 80% of crossword puzzle entries using Google).

I've seen many that are generated by a list of random, but actual, words and/or names. If you have one that's drawing from a random database of words, it's not impossible to have each word associated with certian groups. You could even have more than one group per word (e.g., "Blue" is a color, also a type of cheese, also an emotion, etc etc).

And yeah, a google-crosscheck could give it away, but I'd think it'd at least require added sophistication on the part of the bot.

IMHO, if you can program the security, you can program a way around it. The limitation is that both are products of the same resources.

[Dragon Star]

Why not make a field of shapes that form a mirage of sorts, an illusion, or form a geometric shape through the use of other shapes for a border (e.g. a bunch of triangles stacked to shape a much larger triangle, square, exc...) that most anyone would be able to identify? I should think it would take quite a while for computers to learn that bit.

[Ilya]

Douglas R Hofstadter: on seeing A's and seeing As

To show the fearsome complexity of the task of letter recognition, I offer the following display of uppercase "A"s, all designed by professional typeface designers and used in advertising and similar functions.
[...]
Indeed, I once even proposed that the toughest challenge facing AI workers is to answer the question: "What are the letters 'A' and 'I'?"

I admit that about a quarter of the squiggles in Hofstadter's example I would not recognize as "letter A" had I not known they are supposed to be letters. And several I still do not recognize as letter A.

[tdvance]

Use German Fraktur characters--THAT would stump any AI (and many people)!!! When I took German in college, I was cheap and got a used German-English dictionary for about a dollar, an old one that used Fraktur characters. Looking anything up was a real pain.

[Demigrog]

Text based CAPTCHAs are so late 90s. KittenAuth is the way of the future.

[tdvance]

Just for info--I'll take you too seriously

select several of 9 photos of "elephants" for example, to get through.

If each photo had a 50/50 chance of being an elephant, a random selection has 1 in 512 chance of being right. In reality, each photo probably has less than a 50/50 chance--so if you randomly select, according to a distribution with a mean of 3 and a moderate standard deviation, you probably have better than 1 chance in 512. So, have a botnet, and better than one out of every 512 IP addresses on your net gets you in!

[sarongsong]

Why not make a field of shapes that form a mirage of sorts, an illusion...I should think it would take quite a while for computers to learn that bit.

Is there a 'digital hallucinogen' to help them out here?