ONLY a few weeks into graduate school and aged just 22, Luis von Ahn
helped crack one of the thorniest problems bedevilling the web. It was
the year 2000 and free, web-based e-mail services were booming. But
spammers were creating thousands of accounts automatically and using
them to blast out messages. When the accounts were shut down, they
simply created new ones. At the same time, sites selling tickets to
concerts and sporting events were being besieged by programs that
bombarded them with orders, snapping up the best seats for resale at a
higher price. Websites needed a way to distinguish between human
visitors and automated ones.
Mr von Ahn had just arrived at Carnegie Mellon University in
Pittsburgh when he and his PhD adviser, Manuel Blum, came up with just
such a method. The solution had three requirements: it had to be a test
that humans could pass easily and computers could not—but could use
computers to determine whether the response was correct. The original
idea was to show web users an image, for example of a cat or a roller
coaster, and ask them to identify it. A correct answer would indicate
that the entity at the other end of the internet connection was indeed
human, granting access to the web-mail service or ticketing site. But it
turned out that people were not very good at identifying images
reliably.
So the pair came up with another idea: displaying a distorted
sequence of letters and asking people to read them and type them into a
box. This proved to be a much more reliable test of whether a visitor to
a website was human or not (something that is known, in
computer-science terminology, as a Turing test, in honour of Alan
Turing, a British computer scientist). The result was the CAPTCHA, which
stands for “Completely Automated Public Turing test to tell Computers
and Humans Apart”. Yahoo and other web-mail providers implemented the
system, and it immediately made life harder for spammers.
Mr von Ahn went on to get his doctorate—and a phone call from Bill
Gates of Microsoft offering him a job, which he turned down. He has
since created a series of internet-based systems that bring many people
together to perform useful work by dividing tasks into tiny pieces,
often presented as a simple test or game, and aggregating the results. A
decade ago Mr von Ahn called his approach “human computation” (the
title of his thesis) and “games with a purpose”—precursors to the modern
techniques of “crowdsourcing” and “gamification”.
For example, he noticed that search engines were bad at finding
images, because pictures on web pages are rarely labelled with neat
captions. So he created the ESP Game, in which two players in different
locations are simultaneously shown the same image in their web browsers,
and asked to type words describing what is in it. Each round of the
game ends when both players use the same word, so the aim is to use the
most obvious descriptive terms. In so doing, the players tag each image,
and signal which words best describe it. The technology was acquired by
Google in 2005 to help label images for its search engine.
Shall we play a game?
Mr von Ahn grew up in Guatemala, the son of two doctors. He stumbled
into computers indirectly. In the mid-1980s, at the age of eight, he
wanted to play video games. But instead of giving him a Nintendo console
his mother bought a PC. To play games on it, he resorted to typing in
programs from computer magazines and working out how to crack the
copy-protection schemes on games sold on floppy discs. The young Luis
also spent some time at a confectionery factory owned by his family. He
was fascinated by the machines that made and wrapped the sweets, and was
soon taking some of them apart and reassembling them. His love of
engineering endured, but not his sweet tooth. “I got to play around the
whole time—but now I can’t stand the taste of mint,” he says.
In Guatemala, nearly all students are tested before entering high
school. The top 20 nationwide, of whom Mr von Ahn was one, are sent to a
special school. He went on to study mathematics in America, at Duke
University, switching to computer science for his postgraduate studies
because it was more practical. “You talk to a mathematician, and he
tells you that he’s one in three people in the world who understand the
problem and it’s not been solved for 200 years,” says Mr von Ahn. “A
computer scientist says: ‘I solved an open problem yesterday’.”
In late 2006 Mr von Ahn had just started teaching at Carnegie Mellon
when he received a call from the MacArthur Foundation, saying that he
was being awarded one of its coveted “genius” grants of $500,000. Around
the same time, he did a back-of-the-envelope calculation to get a sense
of CAPTCHA’s popularity, and realised that about 200m squiggly words
were being recognised and typed into computers every day by internet
users around the world. At about ten seconds apiece, that amounted to
around half a million hours daily. This improved the security of the
internet, but at the cost of making people perform a task whose results
were immediately discarded. Surely the recipient of a genius grant ought
to be able to find a way to make more productive use of their efforts?
Driving home from a meeting in Washington, DC, in his blue Volkswagen
Golf, he was struck by an idea. Instead of showing users random
letters, why not show them words from scans of old printed texts that
automated document-digitisation systems, based on optical character
recognition, could not understand? Such words were, by definition,
incomprehensible to computers, but might be legible to humans. They
could be shown to people as part of a modified CAPTCHA test, based on
two words. One, the control word, is a known word; the other is an
illegible word from a scanned document. The user reads and types in the
two words, and is granted access provided the control word is correctly
identified. And when a few users separately provide the same
interpretation of the scanned word, it is fed back to the digitisation
system.
People performing online security checks could thus be put to work
digitising old books and newspapers, without even realising that they
were doing so. Mr von Ahn called his new idea reCAPTCHA, and when the
New York Times
began to use the technology to digitise its archive, he span it out
into a separate company. In 2009 it too was acquired by Google, for use
in its ambitious book-digitisation project. (The slogan for reCAPTCHA is
“Stop spam, read books”.) Mr von Ahn went to work at the internet giant
for a year. Paradoxically one of his tasks while at Google was to shut
down the ESP Game. It had served its purpose, labelling enough images to
train an image-recognition system based on artificial intelligence,
which could then perform the task automatically.
"More than 1 billion people have helped digitise the printed word by using reCAPTCHA.”
Other similar projects followed. Verbosity, for example, got players
to create a compendium of common-sense facts, such as “milk is white”,
which people know but computers do not. But none of Mr von Ahn’s other
projects have come close to reCAPTCHA when it comes to doing useful
work. The system now handles 100m words a day, equivalent to 200m books a
year. If Google were to pay people America’s minimum wage to read and
type in those illegible words, it would cost it around $500m a year.
Although still in his early 30s, Mr von Ahn has already made a unique
contribution to computer science and artificial intelligence, by
harnessing what he calls “the combined power of humans and computers to
solve problems that would be impossible for either to solve alone.” Put
simply, the idea is to “take something that already happens and try to
get something else out of it,” he says. His work exploits the internet’s
ability to reduce co-ordination and transaction costs, so that the
efforts of hundreds of millions of people can be aggregated effectively.
Mr von Ahn estimates that more than 1 billion people have helped
digitise the printed word by using reCAPTCHA.
Found in translation
The notion of “human computation” has spawned its own academic field.
But did Mr von Ahn really set out to generate useful results from
mundane tasks, or did he stumble on the concept with CAPTCHA and then
apply it in other domains? “It is a combination of both,” he says. He
recently came across a plan, devised when he was 13 and saved by his
mother, for a power company that would operate a free gym and generate
electricity from people lifting weights, cycling and so forth. This was,
he now realises, a precursor of his computer-science work, which does
the same for mental activity.
Mr von Ahn has won many prizes, including a presidential award for
excellence in science. He has a gaggle of patents to his name. His old
blue Volkswagen has been replaced by a blue Porsche. And he continues to
apply his distinctive approach to new problems. His latest project is a
company, which he co-founded last year, called Duolingo. It helps
people learn a foreign language (the game), while also providing a
translation service (the useful work). People are shown a word or phrase
which they do their best to translate; others then vote on the best
translation. Duolingo already has 3m users, who use it, on average, for
30 minutes a day.
Although it is outwardly similar to reCAPTCHA, Duolingo marks a
further development of Mr von Ahn’s model, because it also exploits “big
data”. The firm collects huge volumes of data and performs experiments
to determine what works best when learning a new language. For example,
should one teach adjectives before adverbs? Even experts do not know,
because there has never been a large-scale, empirical study. Thanks to
Duolingo, there now is.
Among its early findings is that the best way to teach a language
varies according to the student’s mother tongue. When teaching English,
for example, pronouns such as “him”, “her” and “it” are usually
introduced early on. But this can confuse Spanish speakers, who do not
have an equivalent for “it”. The answer is to delay the introduction of
the word “it”, which makes Duolingo users less likely to give up in
frustration. Mr von Ahn hopes to apply this technique—using data to
improve pedagogy—in other disciplines.
“Education seems like it should be an equaliser, but really it is
not,” he says. “If you have money, you can get a good education; if not,
you don’t.” He has seen at first hand what access to high-quality
education can do, and wants to use technology to make it more widely
available. This is, he realises, a far more ambitious aim than blocking
spam or scanning books, useful as those are. But there is a clear thread
running through his work, even as he tries to apply his approach to
tackling bigger societal problems, not just technical ones. Whether
people are in the gym, logging on to e-mail or learning a new language,
he wants to enable them to “do something useful—and harness the power
that they generate.”