Editorial: “The rise of the reverse cyborg“
Computers just don’t have the brains to solve some basic problems – unless we lend them ours
I PICK up the phone and leave a voicemail message. “Can you give me a call back, it’s urgent,” I say. I’m not actually trying to reach anybody – I am calling myself to test Google Voice, an internet telephony tool that, among other things, uses artificial intelligence to transcribe messages left on your phone. In a minute, it will send the result via email.
Advertisement
While I wait, I go to , a language translation website run by Yahoo. I type out the same message, and translate it to Russian and back again. Next, I snap a photo of the notepad and pen on my desk using my smartphone, and ask the , a visual search engine, to find similar images online. I’m testing all of these tools to get a feel for the artificial intelligence that underpins them – it’s among the best that exists today.
The results are hardly mind-blowing. Google Voice’s transcript of my voice message reads: “Can you give me a call back in session.” Babel Fish’s double translation isn’t much better: “You can give to me bell, it’s is urgent,” it says. And the images offered by Google Goggles are positively odd. No sign of the office stationery images that the app might be expected to provide, but I am offered a drawing by a Renaissance artist and a picture of a fish.
Today’s artificial intelligence still struggles to complete tasks we humans can do with ease. But an ingenious solution to these shortcomings is emerging and it is set to transform our relationship with machines. If an artificial intelligence (AI) encounters a problem it cannot solve, now it can turn to the internet, where thousands of human workers are waiting to assist it. Some people do it for money, while for others it’s fun. In fact, chances are you have already helped a machine do its job and not even realised it. We are creating hybrid intelligences, and they can outperform standard AI in all sorts of ways. So perhaps it’s time to reconsider the science-fiction vision of autonomous assistants with fully artificial minds. The smartest machines in the future may instead be “reverse cyborgs” – artificial intelligences augmented by us – a mix of programming, transistors and the donated brainpower of an army of human workers.
AI is all around us, but it is far from perfect. Computers can recognise well-lit common objects and understand the enunciated words of a newsreader, but real life is not so neat. Our language is packed with puns, metaphors and idioms. We all speak in different accents. And we continually encounter unfamiliar environments and objects. Until machines can deal with such challenges, AI goals such as robot companions will remain a fantasy. If a machine cannot understand what you are saying or move around safely, how can you expect it to load your dishwasher?
The idea that humans could make up for the shortcomings of machines can be traced, at least in part, to a plane journey. It was 2001 and , a computer scientist at Carnegie Mellon University in Pittsburgh, Pennsylvania, was travelling to see his girlfriend. Looking down the row of seats on his plane, von Ahn saw passengers engrossed by a difficult task: newspaper crosswords. “I saw that people were willing to do something that computers can’t,” he says. It turned out von Ahn wasn’t quite right – computers can learn to solve some types of crossword. But the idea stuck. If he could design tasks that humans want to do, he could get people to help machines.
Von Ahn’s first projects were “games with a purpose”. For example, as part of his , players are partnered anonymously and score points if both generate the same one-word caption for images they are presented with. They do it for fun, but they are also unwittingly training vision software to recognise the same types of objects. The word “crowdsourcing” didn’t then exist, but that was what von Ahn was doing.
He next used the crowd for a major effort to digitise books and newspapers. To convert printed letters into digital text, computers use a technique called optical character recognition (OCR), but this can be easily confused if the words are blurred or distorted. But we decipher blurred letters all the time when we browse online. When you visit a website to sign up for services like email, you’re often asked to read a Captcha, a short section of distorted text that is designed to confound the automated software used by spammers. This labour, von Ahn figured, could make up for AI’s deficiencies.
The result, dubbed , was launched in 2007. The software scans pages using OCR, but when it cannot recognise a fragment of a text, it asks for help. It sends the fragment to special Captchas on a vast array of websites ( included). These ask users to decipher both the fragment and a piece of distorted text that has already been digitised. If the user correctly identifies the latter, the software assumes that the former is also correct. “The scale is humongous,” says von Ahn. “We’re doing 100 million Captchas a day.” It’s quite possible that you’ve already completed one, adding your own brainpower to a much bigger hybrid intelligence.
While von Ahn was working on reCaptcha – a project that is now hailed as milestone in human computation – Amazon launched a service that would make possible scores of other cyborg-like systems. By the early 2000s, the online retail giant had grown so large that it needed a better way to spot duplicate entries in its product listings. The task is hard to automate, so the company farms out chunks of the process to online freelancers. Inspired by their success, in 2005 Amazon launched Mechanical Turk, a website that allows anyone to access the skills of a huge human workforce.
It was an almost immediate boon for companies involved in tasks like transcription, which have found that they can get jobs done remarkably cheaply (see “A digital sweatshop?”). But more ambitious uses have emerged since. Researchers interested in human computation now have an army of “Turkers” to call upon. The result is a series of human-machine hybrids that build upon the pioneering work of von Ahn and others.
Some of the projects integrate human and machine in clever new ways. Take the counter-intuitive idea of doing translation without bilingual workers. The idea, known as , is the work of and at the University of Maryland in College Park. Imagine a Russian and a Spanish speaker, neither of whom speaks the other’s language. MonoTrans software translates the sentence back and forth between the two languages, inevitably imperfectly. But after each translation, the Russian or Spanish speaker edits the text to make it clearer, and it is translated back again. Three round trips are usually enough for the translations to reach high quality, say Resnik and Bederson. A pair of workers should eventually be able to translate 1000 words a day, they add.
Mechanical Turk has another advantage: speed. The army is usually thousands-strong, so jobs can sometimes be completed in seconds. This swift response proved particularly useful for a group of neuroscientists in Berkeley, California, which in 2008 set out to build intelligent software for recognising objects. They had established a company called and their initial hope had been to automate object recognition using some of the same tricks as the brain’s visual system, which processes the basic features of a scene, such as the edges of object, before more complex details. “We soon realised that computer vision is not good enough to make a product,” says Pierre Garrigues, the company’s director of research.
IQ Engines’s eventual product – an app called , launched last year – marries the company’s object-recognition technology with the power of the crowd. Users take a picture with an app on their smartphone. The software then takes a first stab at identifying it and, if the system fails, passes the image to Mechanical Turk or the company’s own pool of workers. Response times vary, but workers are able to return about half of all queries in less than 25 seconds, says Garrigues.
The app is good enough to aid people with poor sight. “Close your eyes and imagine how you would identify all the things in your kitchen,” says Darrell Sandrow, a student at Arizona State University in Phoenix, who is almost totally blind. “I can feel the various boxes, cans, jars and other packaging and, in many cases, I can often determine what I have. It isn’t always possible, though. What kind of cereal is in that box? How about all the cans? I’ve used oMoby to save myself a lot of trouble by identifying cans in the kitchen. Simply snap a pic, wait 30 seconds or maybe a minute and get the answer.”
Similar technology could help intelligent machines navigate the world. For example, at , a robotics company in Palo Alto, California, engineers working with a robot called PR2 have experimented with having the device ask workers on Mechanical Turk for help when it confronts an object it does not recognise.
For AIs to really benefit, though, responses from the crowd will have to get faster. Researchers are already reducing the lag. Take the work of at the Massachusetts Institute of Technology. He’s developing a crowd-powered photography app for cellphones. The aim is to help people capture fleeting moments from fast-moving scenes, such as sports events. Rather than wait with shutter finger poised, users continually record video until the crucial moment occurs. The footage then goes to workers on Mechanical Turk, who select the best frame.
To ensure that he has employees on hand, Bernstein signs up workers in advance and pays them 0.5 cents per minute to wait, for up to 5 minutes, until the video is sent. In initial tests, he found that the first workers started analysing the video within 2 seconds of the footage arriving.
In a few years, latencies for some other types of task could be just as brief, and quality control will improve. And as smartphones proliferate, more and more people will have AI apps that can tap into human crowdsourcing from any location. As long as the motivation of money or fun continues to fuel labour (see charts), somebody somewhere will always be on hand to assist a machine.
Ultimate taskmaster
There’s no reason why computers could not ask other machines for labour, too. Eventually, there may be a pool of workers – some human, some silicon – available 24/7. Last year, Dafna Shahaf at Carnegie Mellon University and Eric Horvitz at Microsoft’s research labs in Redmond, Washington, described how such a pool might work. They call it the Generalised Task Market (GTM). The idea is to create software that can execute complex tasks by automatically dividing up jobs between myriad machines and humans, taking into account the skills and cost of each worker.
As a proof of principle, the pair have built a GTM prototype named Lingua Mechanica that does language translation tasks. But Horvitz has more radical long-term ambitions. Imagine that a child fails to return home. A GTM’s missing-person algorithm would be activated. It might scour the web for recent images of the child’s route home. Or recruit volunteers who would act as a network of human sensors across the search region, reporting a sighting via their smartphones. Another search algorithm might read news reports and Twitter for evidence of accidents. All without a human coordinator.
GTMs could have an extraordinary range of uses. In the event of a disaster, such as an oil spill, a GTM might assemble experts alongside data, sensors and robots. It could even manage mega-translation projects, such as converting the whole of Wikipedia into new languages. “It will consider a large set of intelligences and weave together a mesh of people and machines to solve problems,” says Horvitz.
Of course, the changes brought by donating our brainpower to a hybrid intelligence might not be all good. The advent of crowdsourcing platforms may divide the global workforce into two tiers: those who program the computers that post jobs to the crowd and those who, in effect, are the programmed. These workers, who see only the tasks assigned to them, are often ignorant of their employers’ wider goals. So what happens if the GTM or some other algorithm is used to commit a crime? Are the workers culpable? Already, spammers have designed malware that, when faced with a Captcha it cannot read, emails workers in developing countries for the answer. They are probably witting participants, but they needn’t be.
We now have the chance to debate the ethics of human computation, but it may not be long before this new breed of human-augmented reverse cyborgs becomes ubiquitous. Mechanical Turk is just 6 years old and internet crowdsourcing hasn’t been around much longer. In this short time, machine-human hybrids have shown tantalising potential. Soon you may find yourself peering into a pair of robotic eyes, and a million people will be staring right back at you.
A digital sweatshop?
Would you work for $1 an hour? Thousands of people online choose to do just that.
If you want a job done nowadays, it’s easy to turn to workers online. On the website for Amazon’s Mechanical Turk, you can advertise a task and, in exchange for a small sum, a worker somewhere in the world will complete it for you.
found that the average pay on the platform is around $1.40 per hour. In the US, where roughly half of all “Turkers” are based, the federal minimum wage is $7.25 per hour.
Yet the service is not as exploitative as it may first appear. show that most use the site in their spare time. Many are well-educated. They report finding the tasks reasonably enjoyable and a nice way of earning a little extra money. And a third of Turkers are based in India, where many people in conventional jobs earn less than $1.40 an hour.
Still, the online tasks do not come with the safeguards that advocates of employment rights fought for so long to enshrine in law. Most crowdsource workers have minimal defences against rogue employers and no opportunity for promotion or wage negotiations.
