In the center of Bubba Nasvall, about 60 kilometers away from the Bay of Bengal, Namita Pradhan sat at her desk, staring at the video recorded by a hospital on the other side of the world.
The video shows the inside of someone’s colon. pradhan is looking for polyps, which are small lumps in the large intestine that may cause cancer. They look a bit like sticky acne. When she finds the polyp, she will mark it with her computer mouse and keyboard and draw a number circle around this small bump. Pradhan has no special medical training, but she is helping to train an artificial intelligence (AI) system, which can eventually complete the work of a doctor.
On the fourth floor of a small office building, dozens of young Indian men and women are working hard at their desks, and pradhan is one of them. They are trained to annotate all kinds of digital images, from stop signs and pedestrians in street scenes to factories and oil tankers in satellite photos, and they can accurately locate them.
Most people in the technology industry will tell you that AI is the future of their industry. This technology is developing rapidly, thanks to something called machine learning. However, executives of technology companies rarely discuss the labor-intensive efforts in the process of their creation. AI is learning from humans, and it is learning from many humans.
But before the AI system can learn, someone must mark the data provided to it. For example, humans must accurately locate polyps. This work is crucial to the creation of AI such as self-driving cars, monitoring systems and automated medical care. However, technology companies remain silent about this work, because they face increasing concerns from privacy activists about their storing and sharing large amounts of personal data with external enterprises.
Earlier this year, Cade Metz, a senior science and technology editor, tried to help us understand the behind-the-scenes scene of AI training, which the wizards in Silicon Valley rarely agree with. Metz took a stroll in India and visited five offices, where people were engaged in almost endless repetitive work needed to train AI systems, all of which were run by a company called iMerit.
There are intestinal surveyors like Ms. pradhan and experts who distinguish between good and bad coughs, linguists and professionals who recognize street signs. What is a pedestrian? Is that a double yellow line or a virtual white line? In the future, robot cars need to know the difference.
The scene Metz saw doesn’t look like the future we imagined, or at least the automated future you might imagine. The office can be a call center or a payment processing center. One of them is located in an old apartment building in the middle of a low-income residential area in western Kolkata, which is crowded with pedestrians, tricycles and street vendors. In Buba Nasvall, which he visited, and other cities in India, Nepal, the Philippines, East Africa and the United States, tens of thousands of office workers are devoted to training machines.
Tens of thousands of workers, independent contractors who usually work from home, also annotate data through crowdsourcing services such as Amazon Mechanical Turk, which allows anyone to assign digital tasks to independent workers in the United States and other countries, and workers earn a few cents per label.
Headquartered in India, iMerit labels data for many big-name companies in the technology and automobile industries. The company refused to disclose the names of these customers on the grounds of confidentiality agreement. But the company recently revealed that more than 2,000 employees in nine offices around the world are contributing to Amazon’s online data tagging service SageMaker Ground Truth. Previously, it also listed Microsoft as a customer.
What is certain is that AI may hollow out the job market in the future. But for now, it is creating relatively low-paid jobs. According to the data of Cognilytica, the market value of data labels exceeded 500 million dollars in 2018 and will reach 1.2 billion dollars in 2023. Research shows that this kind of work accounts for 80% of the time spent building AI technology.
Is this job exploitative? It depends on where you live and what you are doing. In India, this is the ticket to the middle class. This is a decent job in New Orleans, USA. But for people who are independent contractors, this is often a "point of no return".
Some skills must be learned, such as finding signs of disease in videos or medical scans, or keeping your hands stable when drawing digital lassos around images of cars or trees. In some cases, when the task involves medical videos, pornographic or violent images, the work will become terrible.
Kristy Milland said, "When you first see these things, you will feel deeply uneasy. If you don’t want to go back to work, you may not go back to work. " Mirande spent several years doing data tagging on Amazon Mechanical Turk, and now she has become a labor activist representing the workers of this service. "For those of us who can’t afford to lose our jobs, you just have to keep putting up with it," she said.
Before going to India, Metz tried to label pictures on crowdsourcing service, draw a number box around Nike logo and identify pictures that were "unsafe at work". He was very clumsy at that time. Before he began to work, he had to pass the test, but he failed three times in succession. Labeling images so that people can immediately search for retail goods on the website, not to mention taking the time to identify rough images of naked women and sex toys as "NSFW", is not entirely encouraging.
AI researchers hope that they can build systems that can learn from a small amount of data. But in the foreseeable future, human labor is still essential. Mary Gray, an anthropologist at Microsoft, said: "This is an expanding world hidden under technology, and it is difficult to exclude human beings from the cycle."
Bubba Nasvall is also known as the "Temple City". Ancient Hindu shrines stand in the roadside market at the southwest end of the city, including giant stone pagodas dating back to the 10th century. In the city center, many streets are unpaved. Cows and wild dogs wander between mopeds, cars and trucks.
With a population of 830,000, the city is also a fast-growing online labor center. About 15 minutes’ drive from the temple, on a paved road near the city center, a white four-story building is located behind a stone wall. There are three rooms in it. The rooms are filled with long tables, each with its own widescreen computer screen. This is where Ms. pradhan tagged the video.
Pradhan, 24, grew up outside the city and got a degree from a local university, where she studied biology and other subjects before taking the job at iMerit. This is the job recommended by her brother, who has worked in the company before. Pradhan stays in a hotel near her office on weekdays and goes home by bus every weekend.
Metz visited pradhan’s office in January this year. Many ladies wearing traditional Indian clothes and long gold earrings are sitting at long tables. Ms. pradhan is wearing a green long-sleeved shirt, black trousers and white lace-up shoes, commenting on a video for an American customer. In her usual 8-hour work every day, the shy lady watched more than a dozen colonoscopy videos, and constantly turned them upside down to get a closer look at each frame.
Every once in a while, pradhan will find what she wants, and she will put a digital "box" around it. She drew hundreds of such boxes to label polyps and other signs of disease, such as blood clots and inflammation.
Pradhan’s customer is an American company, and iMerit is not allowed to disclose its name. It will eventually input pradhan’s work into the AI system, so that it can learn to identify medical conditions by itself. The owner of the colonoscope doesn’t necessarily know the existence of the videos, and Ms. pradhan doesn’t know where the videos came from, and neither does iMerit.
Ms pradhan learned this task during a seven-day online video call with a non-intern. This doctor lives in Oakland, California, USA, and helps train many iMerit office staff. But some people question whether experienced doctors and medical students should make such labels themselves.
Dr. George Shih, a radiologist at Weill Cornell Medicine and NewYork-Presbyterian Hospital and co-founder of the startup MD.ai, said that this job needs "people with medical background and knowledge of anatomy and pathology". MD.ai helps enterprises build AI for health care.
Talking about pradhan’s work, she said it was "interesting" but very tiring. As for the graphical nature of video? She admitted: "It was disgusting at first, but then you got used to it."
The images marked by pradhan are terrible, but not as terrible as other images processed by iMerit. Their customers are also building AI, which can identify and delete unwanted pictures on social networks and other online services. This means that pornographic, violent and other harmful images need to be labeled.
This work may make practitioners feel very uneasy, and iMerit tries to limit the amount of such content they review. Liz O ‘Sullivan (Liz O‘ Sullivan) said that pornography and violence are mixed with more harmless pictures, and those pictures with terrible labels are isolated in different rooms to protect other employees. O ‘Sullivan worked closely with iMerit on such projects.
O ‘Sullivan said that other label companies will let employees make unlimited comments on these pictures. "I wouldn’t be surprised if it led to post-traumatic stress disorder or worse," she pointed out. Companies that are morally reprehensible are simply unwilling to take on such responsibilities. You must use other jobs to fill pornography and violence, so that workers don’t have to watch pornography and beheadings. "
IMerit said in a statement that it will not force employees to watch pornographic or other offensive content, and will only undertake this work if it helps to improve the monitoring system. According to a company executive, pradhan and other labelers earn between $150 and $200 a month, and can bring $800 to $1,000 for iMerit.
By American standards, wages in pradhan are ridiculously low. But for her and many others in these offices, this is about the same as the average salary of data entry jobs. Although the job is boring, it can help pay for the apartment.
Plassen Diya Bai grew up on a farm about 50 kilometers from the east coast of India and Kolkata, the largest city in West Bengal. His parents and extended family still live in his childhood home, which was a brick house built in the early 19th century. They planted rice and sunflowers in the surrounding fields and dried the seeds on the carpet covered with the roof.
He is the first person in his family to receive a college education, including computer courses. But the school didn’t teach him so much knowledge, and an average of 25 students in the classroom could be assigned to a computer. After graduating from college, he taught himself computer skills. At that time, he signed up for a training course organized by a non-profit organization called Anudip. This is recommended by a friend, and the monthly cost is equivalent to 5 dollars.
Anudip offers English and computer courses all over India, training about 22,000 people every year. This institution directly recommended students to iMerit, and its founder established iMerit as a sister business in 2013. Through Anudip, Diya Bai got a job in an iMerit office in Kolkata, as did his wife, Barnali Paik, who grew up in a nearby village.
In the past six years, iMerit has hired more than 1600 students from Anudip. At present, the total number of employees in the company is about 2,500, of which more than 80% are from families with a monthly income of less than $150.
Founded in 2012, iMerit is still a private company, which allows employees to perform digital tasks, such as transcribing audio files or recognizing objects in photos. Companies around the world pay companies, and more and more, they are helping with AI training. Radha Basu, who co-founded Anudip and iMerit with her husband deepak, said: "We want people from low-income backgrounds to enter the technology industry." Basu and Deepak have long-term cooperation with technology giants Cisco and Hewlett-Packard in Silicon Valley.
The average age of these workers is 24. Like Diya Bai, most of them come from the countryside. The company recently opened a new office in Metiabruz, a predominantly Muslim community in western Kolkata. There, it employs mostly Muslim women, and their families are unwilling to let them leave this bustling area. They were not asked to look at pornographic pictures or violent materials.
At first, iMerit focused on simple tasks, sorting out product lists for online retail websites and censoring posts on social media, but it has shifted to supporting AI. The growth of iMerit and similar companies represents a shift from crowdsourcing services like Mechanical Turk. IMerit and its customers can better control how employees are trained and how their work is completed.
Diya Bai is now the manager of iMerit, who is responsible for labeling the street scenes used by a large American company to train driverless cars. His team analyzes and marks digital photos and three-dimensional images captured by lidar. They draw bounding boxes around cars, pedestrians, stop signs and wires all day.
Diya Bai said that the job might be boring, but it gave him a life he might not have had. He and his wife recently bought an apartment in Kolkata, within walking distance of the iMerit office where she works. Diya Bai said: "My life has undergone fantastic changes, both in terms of my financial situation, personal experience and English skills. I got a chance! "
A few weeks after his trip to India, Metz took Uber through downtown New Orleans. About 18 months ago, iMerit moved into a building opposite Superdome Street. A big American technology company needs a way to tag data for the Spanish version of its home digital assistant. Therefore, it sends the data to the new iMerit office in New Orleans.
After Hurricane Katrina in 2005, hundreds of construction workers and their families moved to New Orleans to help rebuild the city, and many stayed. Many Spanish speakers came with this new workforce, and the company began to hire them.
Oscar Cabezas, 23, moved to New Orleans from Colombia with his mother. His stepfather found a job on a construction site. After graduating from college, Cabecas joined iMerit and began to develop a Spanish digital assistant.
He annotated everything from tweets to restaurant reviews, identified people and places, and found ambiguities. For example, in Guatemala, pisto means money, but in Mexico, it means beer. What he said: "There are new projects every day."
The work of this office has been extended to other fields, providing services to enterprises wishing to keep data in the United States. For legal and security purposes, some projects must stay in the United States.
Glenda Hernandez, 42, who was born in Guatemala, said she missed her previous job in the digital assistant project. She likes reading. She once commented on books online for large publishing companies, so that she could get free copies. She enjoyed the paid reading opportunities brought by reading in Spanish.
Hernandez is not so interested in image tags or projects like annotating people’s cough recordings, but it is a way to establish AI, which can identify disease symptoms by phone. She said, "Listening to coughing all day is a bit disgusting!" Gray, an anthropologist at Microsoft, said that this work is easily misunderstood. Listening to people cough all day may be disgusting, but it’s also the way doctors spend their days. She said, "We don’t think it’s a chore."
Ms. Hernandez’s job is to help doctors do their job well, or maybe one day, replace them. She is proud of it. Shortly after complaining about this project, she pointed to her colleagues in the office and said, "We are all masters of cough diagnosis."
In 2005, Kristy Milland registered her first job at Amazon Mechanical Turk. She was 26 years old and lived in Toronto with her husband, who managed a local warehouse. Amazon Mechanical Turk is a way to earn some extra money.
The first project is Amazon’s own Three photos of the store will pop up on Mirande’s laptop, and she will choose the one showing the front door. Amazon is building an online service similar to Google Street View, and the company needs help to select the best photos.
She earns $0.03 per click, or about $0.18 per minute. In 2010, Mirande’s husband lost his job and Amazon Mechanical Turk became her full-time job. For two years, she worked six or seven days a week, sometimes 17 hours a day. She earns about $50,000 a year. Ms. Mirande said, "It was enough then, but not now."
The work at that time didn’t really involve AI. For another project, Mirande would extract information from mortgage documents, or retype his name and address from business card photos, sometimes earning only $1 per hour.
Around 2010, Mirande began to label AI projects. She has tagged all kinds of data, such as bloody pictures appearing on Twitter (which helps to establish AI and delete bloody pictures from social networks), or maybe aerial shots taken somewhere in the Middle East, presumably for the AI being built by the military and its partners to identify drone targets.
Mirande said that projects from American technology giants usually pay more than ordinary jobs, about $15 per hour. But this job without medical care or paid holidays may be numb or deeply disturbing. She called it "terrible exploitation" and Amazon declined to comment.
Since 2012, Mirande, 40, has been working in an organization called TurkerNation, which aims to improve the working conditions of thousands of people engaged in this kind of work. In April this year, after working for 14 years, she resigned.
Mirande is in law school, and her husband’s income is $600 less than the rent they pay every month, not including utilities. So, they are preparing for debt. But she won’t go back and label the data. She said, "This is a dystopian future. I’ve had enough!"