When people talk about AI, they’re often referring to software that gets very good at a particular task via a technique called deep learning. With this method, AI systems are given vast amounts of labeled data, and as they run through it, they learn to draw conclusions. For example, an AI tasked with classifying photos of animals would look at millions of images labeled cat, dog, hedgehog, and so forth, and would learn on its own which features define each animal.
Getting a properly labeled data set, however, can be a real stumbling block for researchers and companies designing new AI systems.
The Texas vets are looking at aerial photographs and classifying each vehicle in the shot as car, pickup truck, semitruck, and so forth. Gates says they’re doing this task for a defense contractor, though he can’t say what kind of AI the company is training. But it’s easy to imagine intelligence analysts running smarter searches in an image database by asking an AI tool to identify all aerial images taken in location X with semitrucks in the frame.
Alegion provides a platform that farms out this kind of data-prep work to willing workers all over the world. “Our customers need large-scale data preparation at a high competency level,” Gates says. “They need to know it’s right so when they train their machine learning model, they can trust that it’s trained correctly.”
Everyone wants well-trained AI. In years to come, AI systems may scan your medical X-rays, drive your car, and decide whether or not you get a mortgage. AI promises vast gains in efficiencies, and, when combined with robotics, may replace human workers in many dull and repetitive jobs.
The humans currently doing those jobs will likely not be happy about being replaced. Many experts believe that AI will also make inroads on white-collar jobs, and say the transition to an AI economy will put vast swaths of people out of work. Kai-Fu Lee, the former head of Google China, recently predicted that 50 percent of the world’s jobs are in danger.
Techno-optimists argue that as AI does away with some jobs, it will create new ones, as has happened before in history—for example, in the transition from an agrarian to an industrial economy.
But the question remains: Will humans want these new jobs?
Prepping data for AI is certainly a growth area, says Darrell West, director of the Brookings Institution’s center for technological innovation and author of the recent book The Future of Work: Robots, AI, and Automation. “These are the new-style jobs that will be created by AI and automation: part-time, paid based on tasks completed, without benefits,” he says. “We do have to think about the quality of jobs that are being created.”
West argues that workers are about to suffer from a one-two punch, with increasing levels of AI and automation coming at the same time as a transition to a less-secure gig economy. “It’s the combination of technological revolution and the change in business models that will produce the long-term impact,” he says.
The vets program in Texas is a pilot project that Alegion is running with Operation: Warrior’s Path, a nonprofit organization that Gates encountered as a board member for the Digital Education and Work Initiative of Texas (DEWIT). That initiative aims to help Texans find “meaningful work” in the new digital gig economy.
DEWIT’s website says that cloud labor and microtasking jobs are a boon for people with limited income opportunities in their area, as such computer-based jobs can be done from anywhere. It also argues that workers can learn computer skills that will help them advance and prosper.
“This is supplemental income—people aren’t going to switch to a career of drone image tagging,” Gates says. “But it is real work, and it can bridge the gap until they can receive skills and training to re-enter the workforce.”
Alegion farms out a variety of data-prep tasks to aid in AI training. Workers annotate photos and videos to help computer vision systems understand what they’re looking at; annotate text and audio to help natural-language-processing systems understand what they’re reading or hearing; and generally clean up data to ensure, for example, that the same patient isn’t represented four separate times in a data set of medical records.
The company’s special sauce is a software platform that enables companies to easily distribute such tasks to dispersed workers, and that also evaluates the performance of those workers. “Very quickly we will spit out the poor workers and give good workers training, we’ll help them if they’re missing a certain concept,” Gates says. “So we improve that workforce and deliver the training data with very high confidence.”
About 20 percent of Alegion’s workers are on Amazon’s Mechanical Turk platform, the most well-known crowdsourcing marketplace. Another 40 percent are specialized workers certified to work with certain kinds of sensitive data, such as medical or financial records. The final 40 percent are in Southeast Asia. Gates notes that the Malaysian government has an initiative to bring digital jobs to low-skilled workers. “Malaysia is positioning itself to take advantage of this AI boom,” he says.
Other countries are similarly becoming destinations for this new kind of outsourcing. The BBC recently reported on Kenyan workers who are prepping data for self-driving cars.
Alegion’s workers are paid in two ways. If the company is working with a group like Operation: Warrior’s Path, that group pays the workers an hourly wage. If the company is hiring workers directly, as they do with Mechanical Turkers and people in Malaysia, it pays per completed task.
Gates says that either way, U.S. workers usually earn between $7 and $15 per hour, depending on their efficiency and skill. Foreign workers generally earn between $3 and $6 per hour, he says.
That income for U.S. workers is higher than the incomes that most crowd laborers reportedly make. A recent study found that Mechanical Turkers earn a median wage of $2 per hour, and only 4 percent of workers earned more than $7.25—the current federal minimum wage.
Some workers in the crowd labor marketplace are beginning to speak out about their paltry income. Some Mechanical Turkers have banded together to campaign for higher payments per microtask, and have started an email campaign to Jeff Bezos asking, in part, for a way to represent themselves collectively.
Not everyone agrees that the move to an AI economy is going to cause an employment crisis. “It’s not going to be an Armageddon of people storming the barricades and demanding jobs,” says Robert Atkinson, president of the Information Technology and Innovation Foundation (ITIF), a nonprofit public policy group.
Atkinson argues that while there will be job losses, growth in existing occupations will more than make up the difference. In a 2017 ITIF report, he made the case that productivity gains will lead to more purchasing power for businesses and consumers. “People will spend more money on going out to eat, going to concerts, sending their kids to college,” he says.
As for the new jobs created on the shoulders of AI, such as the jobs the Texas vets are doing for Alegion, he doesn’t worry much about whether those are good jobs. “I look at this as a temporary phenomenon,” he says. “Right now we’re working with a corpus of old data that wasn’t meant for [training AI systems]. At some point we’ll be generating data that’s more clean and more ready for analytics. So some of this will go away as AI becomes a more integrated part of the economy.”
If we humans no longer have to prep data for AI, maybe we’ll find other ways to help it with its labors. For example, the machines doing all that computation need to keep cool. Perhaps humans can fan them with peacock feathers.