People who label artificial intelligence earn only $2 an hour?

Spread the love

People who label artificial intelligence earn only $2 an hour?

Although with the popularity of  AI, we rely more and more on artificial intelligence in our lives, but the ridicule of “artificial mental retardation” has never disappeared.

 

I believe everyone knows that if we want AI to accurately identify the birds in the picture, we need to manually mark these photos as birds in the data set, and then let the algorithm and the image generate correlation judgment and recognition.

 

People who label artificial intelligence earn only $2 an hour?

If it is small-scale experimental data, it would be okay. Once it encounters such a large-scale tagging demand of millions, the time consumed is really unimaginable.

 

Where there is demand, there is a market.

 

According to a joint survey conducted by Princeton University, Cornell University, the University of Montreal, and the Institute of Statistical Science, the researchers found that most of the marking work was done by workers from all over the world outside of Europe and America.

 

The paper pointed out that digital economy companies such as Samasource will hire a large number of cheap workers from sub-Saharan Africa and Southeast Asia, let them complete these boring jobs, pay them 8 dollars a day, and these companies can earn tens of millions of dollars each year. Dollar.

 

People who label artificial intelligence earn only $2 an hour?

Workers working for ImageNet can only get $2 per hour

 

As many deep learning technologies require large amounts of data to train models, the demand for data labels is also increasing. More than 80% of the machine learning development process includes data collection, cleaning, and labeling.

 

For large technology companies such as Uber and Alphabet, these services are even more important.

 

However, when we change our perspective, although these workers are contributing to the current hottest artificial intelligence industry, these systems may not directly benefit their communities in the end. What is even more embarrassing is that many systems may Produce biased judgments about their race or gender.

 

For example, ImageNet, this large public image data set can be said to be one of the most influential data sets in the history of artificial intelligence.

 

The researcher pointed out that workers tagged with ImageNet can only get an hourly wage of $2, and only 4% of workers earn an hourly wage that exceeds the federal minimum wage of $7.25.

 

At the same time, because ImageNet uses WordNet for annotations, according to an experiment called “ImageNet Roulette”, if people submit photos to a neural network trained by ImageNet, the neural network will use the tags in the data set to describe these image.

 

But when people enter the photos they are most interested in: selfies, the software will output some racist and offensive tags to describe them.

 

People who label artificial intelligence earn only $2 an hour?

Although data labeling is not as physically demanding as traditional factory labor, many workers report that their task speed and quantity are “tiring” and “monotonous” because they have to label images strictly in accordance with customer specifications , Video and audio.

 

“Ghost Workers” have no bargaining position, and it is difficult to protect their basic rights

 

In recent years, many local technology companies have emerged in the south of the world, such as Fastagger in Kenya, Sebenz.ai in South Africa and Supahands in Malaysia. With the continuous expansion of artificial intelligence development, the expansion of these companies has also opened the door for low-skilled workers to enter the labor market, but related labor exploitation is still happening.

 

Researchers call these data-labeled workers “ghost workers” because the outside world often sees only the high recognition rate of a system, and ignores the labor work these workers do for the training data set.

Researchers suggest that in the United States, this low wage structure is largely due to time spent on uncompensated activities, such as working on a task that will eventually be rejected.

 

This leads to another problem with the power dynamics of platforms such as Amazon Mechanical Turk. Let’s take this platform as an example. On this platform, all power is concentrated on the requester of the task. The requester has the right to set the price that they expect, which can be as low as $0.01, and the requester can also Rejecting the work done by the workers and claiming that the time required for the task is far less than the time spent by the workers.

 

In the United States, marked workers in such jobs are often considered independent contractors rather than employees, so the protection measures guaranteed by the Fair Labor Standards Act cannot be applied to them.

 

The relevant situation in the United States is discussed only because these data are the easiest to obtain, and there will only be more worse labor phenomena on a global scale.

 

“Assembly line” workers are becoming the competitiveness of enterprises

 

In 2018, BBC reporter Dave Lee visited marker workers living in slums in Kenya and found the same problem.

 

According to reports, the daily work of Brenda, a single mother living in Kibera, is to process most of the image data into a form that the computer can understand.

 

For example, in an uploaded photo, Brenda needs to use the mouse to track the objects that appear, including people, vehicles, road signs, lanes, and the sky, and specify whether it is clear or hazy. Inputting millions of such pictures into the artificial intelligence system can improve the recognition accuracy of the system for products such as self-driving cars.

 

 

Brenda’s working environment is by no means friendly. She and all her colleagues are crowded in a small office. During the whole work, she must keep an eye on the display screen and magnify the image to prevent even one pixel from being mislabeled.

The superiors will check their work, and if they do not meet the requirements, they need to rework.

 

Of course, there are rewards. The names of the fastest and most accurate markers will appear on multiple TV screens in the office for encouragement. Of course, their favorite is the shopping voucher from the mall.

 

Samasource is their largest employer. According to CEO Leila Janah, the company is able to establish partnerships with tech giants such as Google. In addition to accuracy and safety, there is another reason that they have the cheapest labor in the world and local People urgently need stable work.

 

At that time, Samasource gave a daily salary of $9, and they hoped to help those who have a daily salary of less than $2 and need to engage in underground work.

 

“Indeed, it is very cost-effective,” Janah said. “But a key point in our work is that we will not provide a salary level that may disrupt the local labor market. If we give too high a salary, we will It will cause trouble to the entire society. For example, it may have a potential negative impact on the housing cost and food cost of the community where our employees live.”

 

Masakhane, another organization mentioned in the paper, is dedicated to protecting African languages ​​through AI. It is worth noting that Masakhane will not label data for AI researchers, but has established a community to label, research and build algorithms for the African continent.

 

The official website wrote: “We recommend that AI development be regarded as the way forward for economic development.” “This development activity should not focus on low-productivity activities, such as data tagging, but should focus on high-productivity activities, such as model development/deployment. And research”.

 

Finally, the paper points out that for this phenomenon, the potential solution is to simply integrate these data markers into the AI ​​development process, rather than let them make money by marking each image as a pipeline worker. In this way, workers will get a fair salary, and thanks to their life experience and professional knowledge, differences in the data collection process can also be well discovered and resolved, and the overall accuracy of the system will also be improved.


Spread the love