Coughing, colonoscopies and porn: the tedium of teaching AI

Machines don’t just learn by themselves. At centres around the world, hundreds of people spend long hours tagging images. It’s dull, badly paid and often disturbing work

Namita Pradhan, second from right, works alongside colleagues at iMerit’s technology center in Bhubaneswar, India. Tech executives rarely discuss the labour-intensive process that goes into the creation of artificial intelligence. Photograph: Rebecca Conway/New York Times

Namita Pradhan, second from right, works alongside colleagues at iMerit’s technology center in Bhubaneswar, India. Tech executives rarely discuss the labour-intensive process that goes into the creation of artificial intelligence. Photograph: Rebecca Conway/New York Times

 

Namita Pradhan sat at a desk in downtown Bhubaneswar, India, about 40 miles from the Bay of Bengal, staring at a video recorded in a hospital on the other side of the world.

The video showed the inside of someone’s colon. Pradhan was looking for polyps, small growths in the large intestine that could lead to cancer. When she found one – they look a bit like a slimy, angry pimple – she marked it with her computer mouse and keyboard, drawing a digital circle around the tiny bulge.

She was not trained as a doctor, but she was helping to teach an artificial intelligence system that could eventually do the work of a doctor.

Pradhan was one of dozens of young Indian women and men lined up at desks on the fourth floor of a small office building. They were trained to annotate all kinds of digital images, pinpointing everything from stop signs and pedestrians in street scenes to factories and oil tankers in satellite photos.

AI, most people in the tech industry would tell you, is the future of their industry, and it is improving fast thanks to something called machine learning. But tech executives rarely discuss the labour-intensive process that goes into its creation. AI is learning from humans. Lots and lots of humans.

Before an AI system can learn, someone has to label the data supplied to it. Humans, for example, must pinpoint the polyps. The work is vital to the creation of artificial intelligence such as self-driving cars, surveillance systems and automated health care.

Tech companies keep quiet about this work. And they face growing concerns from privacy activists over the large amounts of personal data they are storing and sharing with outside businesses.

Earlier this year, I negotiated a look behind the curtain that Silicon Valley’s wizards rarely grant. I made a meandering trip across India, and stopped at a facility across the street from the Superdome in downtown New Orleans. In all, I visited five offices where people are doing the endlessly repetitive work needed to teach AI systems, all run by a company called iMerit.

There were intestine surveyors like Pradhan, and specialists in telling a good cough from a bad cough. There were language specialists and street scene identifiers. What is a pedestrian? Is that a double yellow line or a dotted white line? One day, a robotic car will need to know the difference.

What I saw didn’t look much like the future – or at least the automated one you might imagine. The offices could have been call centres or payment processing centres. One was a timeworn former apartment building in the middle of a low-income residential neighbourhood in western Kolkata that teemed with pedestrians, auto rickshaws and street vendors.

Punching a clock

In facilities like the one I visited in Bhubaneswar and in other cities in India, China, Nepal, the Philippines, East Africa and the United States, tens of thousands of office workers are punching a clock while they teach the machines.

Tens of thousands more workers, independent contractors usually working in their homes, also annotate data through crowdsourcing services such as Amazon Mechanical Turk, which lets anyone distribute digital tasks to independent workers in other countries. The workers earn a few pennies for each label.

Based in India, iMerit labels data for many of the biggest names in the technology and auto industries. It declined to name these clients publicly, citing confidentiality agreements. But it recently revealed that its more than 2,000 workers in nine offices around the world are contributing to an online data-labelling service from Amazon called SageMaker Ground Truth. Previously, it listed Microsoft as a client.

One day, who knows when, artificial intelligence could hollow out the job market. But for now, it is generating relatively low-paying jobs. The market for data labelling passed $500 million (€450 million) in 2018 and it will reach $1.2 billion (€1.07 billion) by 2023, according to the research firm Cognilytica. This kind of work, the study showed, accounted for 80 per cent of the time spent building AI technology.

When you first see these things, it is deeply disturbing. You don’t want to go back to the work

Is the work exploitative? It depends on where you live and what you’re working on. In India, it is a ticket to the middle class. In New Orleans, it’s a decent enough job. For someone working as an independent contractor, it is often a dead end.

There are skills that must be learned – like spotting signs of a disease in a video or medical scan or keeping a steady hand when drawing a digital lasso around the image of a car or a tree. In some cases, when the task involves medical videos, pornography or violent images, the work turns grisly.

“When you first see these things, it is deeply disturbing. You don’t want to go back to the work. You might not go back to the work,” said Kristy Milland, who spent years doing data-labelling work on Amazon Mechanical Turk and has become a labour activist on behalf of workers on the service.

“But for those of us who cannot afford to not go back to the work, you just do it,” Milland said.

AI researchers hope they can build systems that can learn from smaller amounts of data. But for the foreseeable future, human labour is essential.

“This is an expanding world, hidden beneath the technology,” said Mary Gray, an anthropologist at Microsoft and the co-author of the book Ghost Work, which explores the data-labelling market. “It is hard to take humans out of the loop.”

The city of temples

Bhubaneswar is called the City of Temples. Ancient Hindu shrines rise over roadside markets at the southwestern end of the city – giant towers of stacked stone that date to the first millennium. In the city centre, many streets are unpaved. Cows and feral dogs meander among the mopeds, cars and trucks.

The city – population: 830,000 – is also a rapidly growing hub for online labour. About a 15-minute drive from the temples, on a (paved) road near the city centre, a white, four-storey building sits behind a stone wall. Inside, there are three rooms filled with long rows of desks, each with its own widescreen computer display. This was where Namita Pradhan spent her days labelling videos when I met her.

Over the course of what was a typical eight-hour day, the shy 24-year-old watched about a dozen colonoscopy videos, constantly reversing the video for a closer look at individual frames.

Every so often, she would find what she was looking for. She would lasso it with a digital “bounding box”. She drew hundreds of these bounding boxes, labelling the polyps and other signs of illness, such as blood clots and inflammation.

Her client, a company in the United States that iMerit is not allowed to name, will eventually feed her work into an AI system so it can learn to identify medical conditions on its own. The colon’s owner is not necessarily aware the video exists. Pradhan doesn’t know where the images came from. Neither does iMerit.

Pradhan learned the task during seven days of online video calls with a non-practising doctor, based in Oakland, California, who helps train workers at many iMerit offices. But some question whether experienced doctors and medical students should do this labelling themselves.

This work requires people “who have a medical background, and the relevant knowledge in anatomy and pathology,” said Dr George Shih, a radiologist and the co-founder of the startup MD.ai, which helps organisations build artificial intelligence for health care.

I would not be surprised if this causes post-traumatic stress disorder – or worse

When we chatted about her work, Pradhan called it “quite interesting,” but tiring. As for the graphic nature of the videos? “It was disgusting at first, but then you get used to it.”

The images she labelled were grisly, but not as grisly as others handled at iMerit. Their clients are also building artificial intelligence that can identify and remove unwanted images on social networks and other online services. That means labels for pornography, graphic violence and other noxious images.

Grisly images

This work can be so upsetting to workers, iMerit tries to limit how much of it they see. Pornography and violence are mixed with more innocuous images, and those labelling the grisly images are sequestered in separate rooms to shield other workers, said Liz O’Sullivan, who oversaw data annotation at an AI startup called Clarifai and has worked closely with iMerit on such projects.

Other labelling companies will have workers annotate unlimited numbers of these images, O’Sullivan said.

“I would not be surprised if this causes post-traumatic stress disorder – or worse. It is hard to find a company that is not ethically deplorable that will take this on,” she said. “You have to pad the porn and violence with other work, so the workers don’t have to look at porn, porn, porn, beheading, beheading, beheading.”

IMerit said in a statement it does not compel workers to look at pornography or other offensive material and only takes on the work when it can help improve monitoring systems.

Pradhan and her fellow labellers earn between $150 and $200 a month, which pulls in between $800 and $1,000 of revenue for iMerit, according to one company executive. For Pradhan and many others in these offices, it is about an average salary for a data-entry job. Tedious work. But it pays for an apartment.

Founded in 2012 and still a private company, iMerit has its employees perform digital tasks such as transcribing audio files or identifying objects in photos. Businesses across the globe pay the company to use its workers and, increasingly, they assist work on artificial intelligence.

Kristy Milland, a former data labeller for Amazon Mechanical Turk. Her job didn’t come with paid vacation or health care, and the work could be mind-numbing or downright disturbing. Photograph: Arden Wray/New York Times
Kristy Milland, a former data labeller for Amazon Mechanical Turk. Her job didn’t come with paid vacation or health care, and the work could be mind-numbing or downright disturbing. Photograph: Arden Wray/New York Times

At first, iMerit focused on simple tasks: sorting product listings for online retail sites; vetting posts on social media. But it has shifted into work that feeds artificial intelligence.

The growth of iMerit and similar companies represents a shift away from crowdsourcing services such as Mechanical Turk. IMerit and its clients have greater control over how workers are trained and how the work is done.

A few weeks after my trip to India, I took an Uber through downtown New Orleans. About 18 months ago, iMerit moved into one of the buildings across the street from the Superdome.

Hernandez’s work is intended to help doctors do their jobs or maybe, one day, replace them

The office serves businesses that want to keep their data within the United States. Some projects must remain stateside, for legal and security purposes.

Glenda Hernandez, 42, who was born in Guatemala, said she missed her old work on the digital assistant project. She loved to read. She was less interested in image-tagging or projects such as the one that involved annotating recordings of people coughing; it was a way to build AI that identifies disease symptoms of illness over the phone.

“Listening to coughs all day is kind of disgusting,” she said.

‘Cough masters’

The work is easily misunderstood, said Mary Gray, the Microsoft anthropologist. Listening to people cough all day may be disgusting, but that is also how doctors spend their days. “We don’t think of that as drudgery,” she said.

Hernandez’s work is intended to help doctors do their jobs or maybe, one day, replace them. She takes pride in that. Moments after complaining about the project, she pointed to her colleagues across the office.

“We were the cough masters,” she said.

In 2005, Kristy Milland signed up for her first job on Amazon Mechanical Turk. She was 26, and living in Toronto with her husband, who managed a local warehouse. Mechanical Turk was a way of making a little extra money.

The first project was for Amazon itself. Three photos of a storefront would pop up on her laptop, and she would choose the one that showed the front door. Amazon was building an online service similar to Google Street View, and the company needed help picking the best photos.

She made three cents for each click, or about 18 cents a minute. In 2010, her husband lost his job, and “MTurk” became a full-time gig. For two years, she worked six or seven days a week, sometimes as much as 17 hours a day. She made about $50,000 a year.

“It was enough to live on then. It wouldn’t be now,” Milland said.

The work at that time didn’t really involve AI. For another project, she would pull information out of mortgage documents or retype names and addresses from photos of business cards, sometimes for as little as a dollar an hour.

Around 2010, she started labelling for AI projects. Milland tagged all sorts of data, such as gory images that showed up on Twitter (which helps build AI that can help remove gory images from the social network) or aerial footage likely taken somewhere in the Middle East (presumably for AI that the military and its partners are building to identify drone targets).

Projects from US tech giants, Milland said, typically paid more than the average job – about $15 (€13.50) an hour. But the job didn’t come with health care or paid vacation, and the work could be mind-numbing – or downright disturbing. She called it “horrifically exploitative”. Amazon declined to comment.

Since 2012, Milland, now 40, has been part of an organisation called TurkerNation, which aims to improve conditions for thousands of people who do this work. In April, after 14 years on the service, she quit.

She is in law school, and her husband makes $600 less than they pay in rent each month, which does not include utilities. So, she said, they are preparing to go into debt. But she will not go back to labelling data.

“This is a dystopian future,” she said. “And I am done.”

– 2019 The New York Times Company

The Irish Times Logo
Commenting on The Irish Times has changed. To comment you must now be an Irish Times subscriber.
SUBSCRIBE
GO BACK
Error Image
The account details entered are not currently associated with an Irish Times subscription. Please subscribe to sign in to comment.
Comment Sign In

Forgot password?
The Irish Times Logo
Thank you
You should receive instructions for resetting your password. When you have reset your password, you can Sign In.
The Irish Times Logo
Please choose a screen name. This name will appear beside any comments you post. Your screen name should follow the standards set out in our community standards.
Screen Name Selection

Hello

Please choose a screen name. This name will appear beside any comments you post. Your screen name should follow the standards set out in our community standards.

The Irish Times Logo
Commenting on The Irish Times has changed. To comment you must now be an Irish Times subscriber.
SUBSCRIBE
Forgot Password
Please enter your email address so we can send you a link to reset your password.

Sign In

Your Comments
We reserve the right to remove any content at any time from this Community, including without limitation if it violates the Community Standards. We ask that you report content that you in good faith believe violates the above rules by clicking the Flag link next to the offending comment or by filling out this form. New comments are only accepted for 3 days from the date of publication.