Our understanding of artificial intelligence (AI) is largely determined by pop culture, by blockbuster movies like “Terminator” and their catastrophic scenarios of machines going crazy and destroying humanity. This type of narrative about AI is also what attracts media attention: one of the most talked about AI-related news stories in recent months was that of a Google engineer affirming that his chatbot was sensitive, even being seen by the millions of viewers of the show Stephen Colbert. But the idea of superintelligent machines with their own agency and decision-making power is not only far from reality, it distracts us from the real risks to human lives surrounding the development and deployment of AI systems. While audiences are distracted by the non-existent spectrum of sentient machines, an army of precarious workers is behind the supposed achievements of current artificial intelligence systems.

Many of these systems are developed by multinational corporations located in Silicon Valley, which have been consolidating their power on a scale that, as points out journalist Gideon Lewis-Kraus, is probably unprecedented in human history. In Silicon Valley, they are striving to create autonomous systems that will one day have the capacity to perform all the tasks that people can do and more, without the salaries, benefits or other “costs” associated with formal employment. Although this utopia of corporate executives is far from reality, the race to try to make it a reality has created a global underclass, doing what anthropologist Mary L. Gray and computational social scientist Siddharth Suri call ghost job[2]: the invisible human work that drives “AI”.

Technology companies that have proclaimed themselves “AI First”[3] depend on the outsourcing and constant surveillance of workers, such as data labelers, distributors and content moderators. Startups even hire people to impersonate AI systems, such as chatbots, due to pressure from investors to incorporate “artificial intelligence” into their products. In fact, the London-based investment fund MMC Ventures surveyed 2,830 startups[4] of the EU and found that the 40% of them did not use AI in any significant way.

Far from the sophisticated and sensitive machines presented by the media and pop culture, “artificial intelligence” systems[5] are nourished by millions of workers from all over the world, who perform repetitive tasks in precarious and poorly paid working conditions. And unlike “AI researchers” who earn six-figure salaries at Silicon Valley companies, these sweatshops are often recruited from impoverished populations and are paid as little as $1.46 an hour after removing taxes. However, labor exploitation does not occupy a central place in the discourse on ethics in the development and implementation of AI systems. In this article, we give examples of labor exploitation driven by “artificial intelligence” systems and argue that supporting workers' transnational organizing and unionization efforts should be a priority in debates regarding the ethics of AI.

We write this article as people closely involved in the work around artificial intelligence. Adrienne is a former Amazon delivery driver and warehouse employee who has experienced the harms of surveillance and unrealistic work goals set by automated systems. Milagros is a researcher who has worked closely with people who work with the data, especially data annotators in Syria, Bulgaria, Germany and Argentina. And Timnit is a researcher who has faced retaliation for revealing and publicizing the harms of AI systems.

Treat workers like machines

Much of what is currently described as AI are systems based on statistical machine learning and, more specifically, deep learning (deep learning) through artificial neural networks, a methodology that requires enormous amounts of data to “learn” from them. But about 15 years ago, before the proliferation of gig work[6] or precarious platform work (as we have chosen to translate it in this article), deep learning systems were considered a mere academic curiosity, reduced to a small interested group of researchers. 

However, in 2009, Jia Deng and his team they published the data set ImageNet, the largest labeled image dataset of the time, made up of images pulled from the Internet and labeled via the newly introduced platform mechanical turk from Amazon. Amazon Mechanical Turk, whose motto is “artificial artificial intelligence,” popularized the phenomenon of crowd work[7] which involves large volumes of time-consuming work, broken down into smaller tasks that can be completed quickly by millions of people around the world. With the emergence of Mechanical Turk, difficult tasks to tackle suddenly became possible; For example, hand-tagging a million images can be performed automatically by a thousand anonymous people working in parallel, each labeling only a thousand images. And what's more, at a price affordable even for a university: at, them and crowdworkers They are paid per task completed, which means earning a few cents for each order made.

“Los sistemas de la llamada “inteligencia artificial” son sostenidos por millones de trabajadoras y trabajadores mal pagos de todo el mundo, que realizan tareas repetitivas en condiciones laborales precarias”.

The ImageNet dataset was followed by the ImageNet Large Scale Visual Recognition Challenge, where researchers used the dataset to train and test models that performed a variety of tasks, including image recognition. For example, associate an image with the type of object in the image, such as a tree or a cat. While models not based on deep learning performed these tasks with the greatest precision at that time, in 2012 a architecture based on deep learning, informally nicknamed as AlexNet, outperformed all other models by a wide margin. This catapulted deep learning models and put them center stage. Currently, multinational corporations proliferate models that require a lot of data, labeled by precarious workers on digital platforms around the world. In addition to labeling data extracted from the Internet, in some of these jobs employees[8] are asked to generate data: upload it selfies, photos of their friends and family or images of the objects that surround them.

Unlike 2009, when the main platform for crowdwork was Amazon's Mechanical Turk, we are currently in a time of expansion, where there are countless data labeling companies. These companies are collecting tens to hundreds millions in investment funds, while it is estimated that people who work as data labelers earn an average of $1.77 per task. Data labeling interfaces have evolved to treat the, the and the crowdworkers like machines, often prescribing highly repetitive tasks, monitoring their movements using automated tools and penalizing their mistakes. Today, far from being an academic challenge, large corporations that claim to be “AI First” are nourished by this army of precarious platform workers composed of: data-entry, content moderators, warehouse workers and people who do delivery work through platforms, among other functions.

To illustrate this we can see what happens with content moderators, who are responsible for finding and “flagging” content considered inappropriate for a given platform. Not only is their work essential, without which social media platforms would be completely unusable, but their work flagging different types of content is also used to train automated systems whose goal is to identify text and images that contain hate speech, news false content, violence or other types of content that violate the platforms' policies. Despite the crucial role they play in both keeping online communities safe and training artificial intelligence systems, they often They are paid miserable salaries. This situation occurs despite the fact that their employers are technology giants. As if that were not enough, this group of workers is forced to carry out tasks that can be traumatic while they are closely monitored.

All the videos of murders, suicides, sexual assaults or abuses of children and adolescents that do not reach the platforms have been seen and marked by the content moderators or by an automated system trained from data most likely provided by one of these people. The employees who perform these tasks can reach suffer anxiety, depression and post-traumatic stress disorder due to constant exposure to these disturbing contents.

In addition to experiencing a traumatic work environment with no or insufficient mental health care, these people are monitored and punished if they deviate from their prescribed repetitive tasks. This is the case of those who moderate content in the Sama company hired by Meta in Kenya, who are controlled, monitored and controlled by surveillance software to ensure that they make decisions about the violence in the videos within 50 seconds, regardless of the length of the video or how disturbing it is. Some of these people they fear If they don't, they could be fired after a few mistakes or delays. “By prioritizing speed and efficiency above all else,” reported Time Magazine, “this policy could explain why videos containing incitement to hate and violence have remained on Facebook's platform in Ethiopia.”

Like social media platforms, which would not function without content moderators, electronic commerce (e-commerce) conglomerates like Amazon are managed by armies of warehouse and delivery employees, among others. In line with what happens with those who work in content moderation, this group of workers keep the platforms running and provide data to the AI systems that Amazon could one day use to replace these people with technology: robots that They store packages in warehouses and autonomous vehicles that deliver them to their customers. In the meantime, these people must perform repetitive tasks under the pressure of constant surveillance, tasks that sometimes put their lives at risk and often provoke serious injuries.

“Data labeling interfaces have evolved to treat crowdworkers like machines, often prescribing highly repetitive tasks, monitoring their movements, and punishing distractions through automated tools.”

The Amazon warehouse employees are monitored through cameras and inventory scanners (readers), given that their performance it is measured based on the times that managers determine that each task should take, based on aggregate data from all those who work in the same facilities. They control the time they spend outside of assigned tasks and if they exceed them according to the company's criteria, are sanctioned and disciplined.

Like those who work in warehouses, those who do Amazon delivery work receive controls through automated surveillance systems, through an application called Mentor that calculate the score of its workers based on detected “infractions”. Amazon's unrealistic expectations for delivery times prompt many, many, and many to take risky steps to ensure the delivery of the total number of packages assigned to them for the day. For example, the time it takes for someone to fasten and unfasten their seat belt about 90-300 times a day is enough so that it is delayed on its route. Adrienne and several of her classmates they fastened the seat belt behind their back, so that surveillance systems recorded that they were driving with their seat belts on, without them being slowed down by actually driving with their seat belts on.

In 2020, Amazon drivers in the United States suffered injuries at a higher rate (almost a 50%) to that of United Parcel Service (UPS) workers[9]. And, in 2021, they were injured at a rate of 18.3 per 100 drivers, almost 40% more than the previous year. These conditions are not only dangerous for those who work: pedestrians and car passengers have turned out wounds or have died due to accidents involving Amazon drivers. Recently, a group of delivery girls and boys in Japan resigned and protested because Amazon's software sent them on “impossible routes,” which led to “unreasonable demands and long work hours.” However, despite these clear risks, Amazon continues to treat its employees like machines.

In addition to tracking its workers using scanners and cameras, last year the company required those who work as delivery drivers in the United States to sign a “biometric consent“, which authorized Amazon to use AI cameras to monitor your movements, supposedly to reduce distracted driving or speeding and ensure the use of seat belts. It is reasonable for workers to fear that facial recognition and other biometric data could be used to improve surveillance tools or to continue training artificial intelligence that could one day replace them. The ambiguous wording of these forms consent form leaves its exact purpose open to interpretation, and who must sign them is already they have suspected of unwanted uses of your data (although Amazon denied this).

The “AI” industry operates at the expense of low-paid workers, who are kept in precarious positions. The lack of unionization and precarious hiring conditions make it difficult for them to organize to oppose unethical practices or demand better working conditions, for fear of dismissal from a job they cannot afford to lose. Companies make sure to hire people from vulnerable and underserved communities, such as refugees, people who have been deprived of their liberty and other groups with few job opportunities. They are often hired through third parties such as contractors temporary service providers and not in a dependency relationship. Although it is desirable that more employers hire people from vulnerable groups like these, it is unacceptable that companies do so in a predatory or abusive manner, without any type of protection for them.

“Ethical AI researchers must analyze harmful AI systems as both causes and consequences of unfair working conditions in the industry.”

Data labeling jobs are often carried out far from Silicon Valley headquarters and far from the multinationals that apply the “AI First” logic: from Venezuela, where data is labeled for image recognition systems in autonomous vehicles, up to Bulgaria where refugees from Syria feed facial recognition systems with selfies labeled according to categories of race, sex and age. These tasks usually subcontract to precarious workers from countries such as India, Kenya, the Philippines or Mexico. These people often do not speak English, but receive instructions in English, and face dismissal or to the expulsion from the platforms of crowdwork if they don't understand the rules well.

These corporations know that greater worker power would slow their race toward the proliferation of “AI” systems that require enormous amounts of data and are implemented without adequately studying or mitigating their harm. Talking about sentient machines only distracts us from holding them accountable for the exploitative labor practices that drive the “AI” industry.

An urgent priority for AI ethics

While researchers of ethical AI, AI for social good or human-centered AI have focused primarily on developments to eliminate bias in data[10], promoting transparency and model fairness[11], here we argue that ending labor exploitation in the AI industry should be the focus of such initiatives. For example, if companies were not allowed to exploit labor from Kenya to the United States, they would not be able to proliferate harmful technologies so quickly: their market estimates would simply deter them from doing so.

We therefore advocate for funding research and public initiatives that aim to reveal problems at the intersection of labor and AI systems. AI ethics researchers should analyze harmful AI systems as causes and consequences of unfair working conditions in the industry. Likewise, it would be desirable for them to reflect on the use they make of crowdworkers to advance their own careers, while these people continue working in precarious conditions. Instead, the AI ethics community should work on initiatives that transfer power at the hands of the workers. Some examples of this include co-creating research agendas with those working based on their needs, supporting union organizing (with the challenge of doing so across countries), and ensuring that research results are easily accessible rather than limit yourself to publishing in academic spaces. A great example of this is the platform Turkopticon created by Lilly Irani and M. Six Silberman, “an activist system that allows workers to disclose and evaluate their relationships with their employers.”

Journalists, artists, scientists, and scientists can help by making visible the connection between labor exploitation and harmful AI products in our daily lives, promoting solidarity and support for precarious platform workers and other vulnerable populations. . Journalists and commentators can show the general public why they should care about those who record data in Syria or conduct hyper-supervised Amazon deliveries in the US. Shame works in certain circumstances and, for companies, the public sentiment ofyou should be ashamed” can sometimes equate to a loss of revenue and help move the needle toward accountability.

Supporting the transnational organization of workers must be at the center of the fight for “ethical AI.” Although each workplace and geographic context has its own idiosyncrasies, knowing how similar problems were navigated elsewhere can inspire local organizing and unionization efforts. For example, data taggers in Argentina could learn from recent union efforts of those who work in content moderation in Kenya, or of the Amazon Mechanical Turk workers who they organize in the United States, and vice versa. In addition, unionized people in a geographic location can advocate for others in more precarious situations, as is the case of the Alphabet Workers Union, which includes both well-paid employees in Silicon Valley, as well as low-paid subcontractors in more rural areas.

“This type of solidarity between well-paid technology employees and their lower-paid peers – who vastly outnumber them – is a tech CEO's nightmare.”

This type of solidarity between well-paid technology workers and their lower-paid counterparts - who vastly outnumber them - is the nightmare of any technology CEO. While companies often view their low-income workers as disposable, they are more reluctant to lose high-income employees[12], who can quickly get jobs to competitors. Thus, those with better salaries have more freedom to organize, unionize, and express their disappointment with the company's culture and policies. They can use this situation of greater job (and corporate) security to defend their peers in worse conditions who work in warehouses, delivering packages or tagging data. As a result, companies appear to use every tool at their disposal to isolate these groups from each other.

Emily Cunningham and Maren Costa developed this kind of peer solidarity that scares tech company CEOs. Both worked as user experience designers at Amazon headquarters in Seattle for 21 years. Along with others, others and other Amazon corporate employees, they co-founded Amazon Employees for Climate Justice (AECJ). In 2019, more than 8,700 Amazon workers publicly signed their names to an open letter to Jeff Bezos and the company's board of directors demanding climate leadership and concrete measures the company needed to implement to align with climate science and protect to the workers. That same year, AECJ organized the first strike in Amazon history. According to reports, more than 3,000 Amazon workers demonstrated around the world in solidarity with the Global Climate Strike led by young people.

Amazon responded by announcing its Climate Commitment (The Climate Pledge), a commitment to achieve net zero carbon by 2040, 10 years before the Paris Climate Agreement. Cunningham and Costa claim that they were sanctioned and threatened with dismissal after the climate strike, but they were not fired until AECJ organized actions to promote solidarity with those with lower salaries. Hours after another AECJ member sent a calendar inviting a group of warehouse workers to talk about the terrible working conditions they faced at the beginning of the pandemic, Amazon fired Costa and Cunningham. The National Labor Relations Board determined that their layoffs were illegal, and the company later reached a monetary settlement with both women for undisclosed amounts. This case illustrates where executives' fears lie: the unwavering solidarity of high earners who see low earners as their comrades.

In this sense, we urge researchers and journalists to also focus on the role that precarious workers have in the functioning of “AI” and stop deceiving the public with narratives of totally autonomous machines with the capacity action similar to that of a human. These machines are built by armies of low-paid employees and precarious employees around the world. With a clear understanding of the labor exploitation behind the current proliferation of AI systems, people can fight for better working conditions and real consequences for companies and corporations that fail to comply.