Not-so-artificial intelligence: the human workers who power AI

After graduating from university in California, Dylan Baker was hired by Google. It was the classic career path of a young software engineer looking to pay off his student loans. He joined the tech powerhouse in 2017 to work on machine learning (ML), a technology by which artificial intelligence (AI) devices can “learn” from data without being given precise instructions.

To do this, ML uses what is known as “labelled data”: information to which explanations are attached as to their meaning or content. For example, an image of a cat may be labelled with the location of its ears and snout, or a video of a person with a transcript of what they said or a description of their emotion.

Dylan and his colleagues use these bite-sized chunks of labelled data to feed their AI systems. “At that point in my career, I didn’t even know that labelling data was a job in its own right”, recalls Baker. “We were getting so-called labelled data, but who was labelling it, and how? We didn’t ask questions.” It was during his research that the young engineer discovered the reality of the working conditions of the people who labelled his data.

AI trainers without rights

Some of these people, known as AI trainers, are employed in large centres located in countries where labour is cheap. But many work for platforms such as Amazon Mechanical Turk or Clickworker.

Spread across the four corners of the world, these workers do what is known as “click work”. They perform small, standardised, low-skill tasks on demand. Companies and other organisations send these assignments to the platforms, paying them just a few cents per unit.

By 2022, Baker’s “cognitive dissonance” (as he puts it) between his values and his work was getting to be too much. He had concerns about the biases of AI and about the working conditions of the people producing the data, but these were ignored by his superiors. So he left Google to join the Distributed AI Research Institute (DAIR), founded by Timnit Gebru, an engineer and researcher in AI ethics (who was sacked by Google). Now 28, Baker is researching ethical AI and campaigning for better working conditions for the people who train it.

It was in this capacity that Dylan Baker was invited to take part in a discussion at the European Parliament on 21 November 2024, organised by French MEP Leïla Chaibi (GUE/NGL, left). “I’m here to give an engineer’s point of view, but above all to say ‘listen to the workers and workers’”, he said.

Pennies for jobs, from Venezuela and Syria to the USA and Spain

Sitting next to him, her face hidden by lavender-tipped brown hair, Oskarina Fuentes is one of those workers. Speaking in Spanish, the 34-year-old Venezuelan explains her occupation for the last decade. She works with a number of platforms, but can only name one – Appen – because she is bound by confidentiality agreements.

These are the only things she has signed with any of the companies: Fuentes works without a contract and is paid by the job. She got into the business when she was studying at university, to earn a bit of income. At the time, she was aiming to join Venezuela’s national oil company, but inflation had already made the local currency worthless. The platforms paid her in US dollars. It was only a few cents per job, but this “was still better to live on than the Venezuelan minimum wage”, she explains. She ended up making “100%” of her living from this activity. Using a basic laptop, a model which the government gives to schoolchildren and that she picked up on the black market, she spends her days switching between five platforms.

In 2019, “life in Venezuela had become impossible”, with inflation, power cuts and internet outages. Oskarina took a bus to Colombia. Just a few months after her arrival there, she fell ill. She was diagnosed with type 1 diabetes, which crippled her to the point where she couldn’t work a normal day. She had no choice but to continue living off the work of the platforms.

In all the time she has been training AIs, Oskarina has learned to do everything. Typically, she is checking and scoring the outputs of algorithms (such as a Google search result), updating data on a company or a person, or determining which age bracket a video corresponds to. Over the last few years, jobs have become increasingly rare.

‘Without constant human input, AI models will eventually self-destruct’ – Dylan Baker, ethical AI researcher

To keep earning a decent income, she has to keep signing up to different platforms: “I have all the windows open at the same time on the small screen of my computer. It’s a bit hard on the eyes, but I have no choice, I have to make enough money to pay rent and bills.” In general, a unit of work is paid between $0.01 and $0.05. Assignments are few and far between and the labour pool large, so this young woman is never really disconnected: “Sometimes I get up at three in the morning just to make a few cents.”

Big Tech has effectively made these microworkers invisible: they toil alone, without income security, and compete with each other for simple tasks. Without a contract, they have no job security. A company can even refuse a piece of work if it considers that it is not sufficiently well done. In this case, the worker will not be paid, even if the customer gets to keep the labelled data. The result is lost revenue and precious time wasted. Often, the workers affected don’t even know why their work has been refused.

Yasser Al Rayes attends the event at the European Parliament remotely from Syria. Behind him, a large window gives a glimpse of the buildings of Damascus. He is a young graduate in computer science and AI. “We don’t have a stable internet connection, the electricity cuts [are recurrent], and it is expensive to work in a place with good bandwidth”, he explains. “However, the customers of the platforms sometimes set really high standards for how a task should be carried out.” If workers are disconnected in the middle of a task, they may be refused payment. “Or even get kicked off the platform” if it happens too often, complains this young Syrian.

In a documentary about his daily life, made for the Data Worker Inquiry project, Yasser recounts the hours spent trying to understand the instructions for a task. “I’ve finished all my tasks for the day, and they’ve all been validated by my supervisors. And then you see that the customer has refused them all. I have to start all over again.” Ten hours of work for nothing.

To fight back against such abuses and to share advice about vague working instructions, the microworkers felt they had no choice but to get organised.

Krystal Kauffmann lives in the USA. Like Oskarina Fuentes, she began working for platforms when a chronic illness forced her out of conventional employment. “This was in 2015, before the pandemic, and remote working didn’t really exist in my area”, says the Michigan-born woman. “So I Googled work-from-home opportunities and came across Amazon Mechanical Turk”, she recalls.

After years of working alone behind her screen, she joined and then took the helm of Turkopticon, an organisation created by and for workers on microworking platforms. There she discovered the glaring inequality between herself and her international colleagues: “People in Latin America or India were paid much less than me for exactly the same work.”

Turkopticon started out as a simple forum for reviewing jobs and clients. Today, it brings together people from all over the world in different discussion channels, and advocates for their rights. “In an ideal world, data workers would be recognised for being the experts that they are. They would have access to an equal amount of work, equal pay, psychological support, and so on”, argues Kauffmann.

“Generative AI will always need humans”

“We’re at a tipping point: the European Union is looking at how to regulate AI and AI work”, explains Leïla Chaibi. “These workers upstream of the algorithm”, she says, should be an absolute priority. Invisible behind their telephones and laptops, the microworkers have been overlooked in the European debate on regulating AI.

Nacho Barros, a Spaniard, recalls his first steps on the platforms during the lockdown of 2020: “At first, I found it rather fascinating. I liked some of the jobs. But I soon realised that all the time I spent choosing my tasks, signing up to platforms, qualifying for different assignments, was unpaid.” As a career it was simply too insecure, so Barros went back to working in the hotel industry. But he is still involved in campaigning for the regulation of click work. Because if there were a protective framework – “and a decent wage”, he stresses – then Nacho could easily see himself returning to the job full-time.

“Generative AI will always need humans – language is constantly changing”, says Krystal Kauffman. Dylan Baker agrees: “It’s a well-crafted marketing strategy on the part of the platforms to claim that one day AI will no longer need humans. But that is absolutely not viable. Without constant human input, AI models will eventually self-destruct.”

👉 Original article on Basta!

🤝  This article it’s published within the Come Together collaborative project