Interviews
Lucy Mugo & Michael Geoffrey Abuyabo Asia The human workforce behind Artificial Intelligence
Artificial intelligence is often presented as something automatic, seamless and almost magical. But behind many AI systems there is a vast, largely invisible workforce that collects, labels, moderates, cleans and classifies the data that allows these systems to function. In this conversation, Lucy and Michael from the Data Labelers Association speak about the realities of data work in Kenya and across Africa: low pay, lack of transparency, mental health harms, legal gaps, and the urgent need to recognise data workers as essential actors in the AI supply chain.
To begin, how would you describe data work to someone who has never heard of it?
Lucy: Data work is the hidden labour force behind AI systems. In simple terms, I would say it contributes maybe 70% or more.
As data workers, we are given raw data: images, videos, text or other materials. We are then asked to label, tag or annotate them. Different companies use different names for the work, but basically we are labelling the data so that, later, it can be used to train an AI system.
For example, if you are training a self-driving car, you have to label things like: this is a human, these are road markings, this is a tree, this is a traffic light. So when that self-driving car reaches a zebra crossing, it can identify that people are crossing and stop. That identification happens because of the labour we have done.
Michael: Data work revolves around training these systems. Normally, data workers are given raw data and a set of instructions, or a manual, and they are supposed to label the data according to that manual.
It begins with what is being built. Once the developers have a clear idea of what they want to develop, they create a set of rules. Then workers are trained, often by teaching themselves through training tasks. For instance, if a company wants facial recognition data, it may take random pictures of people and ask workers to draw boxes around faces and identify them.
You need thousands and thousands of pictures, because people are different: tall, short, children, people in wheelchairs, people on bicycles. All of that has to be labelled so that the system becomes precise
The same applies to self-driving cars. You have to help the system understand what can be found on the road: a human being crossing, an animal, a fire hydrant, another car, road markings, traffic lights. And this cannot be done with just a few images. You need thousands and thousands of pictures, because people are different: tall, short, children, people in wheelchairs, people on bicycles. All of that has to be labelled so that the system becomes precise.
Even something like the sky has to be labelled correctly. If you do not label the sky, the system might later confuse a blue sky with a body of water because both are blue. So data work is about helping machines identify objects of interest in a very precise way.
When people hear terms like “training data” or “datasets”, they often imagine something purely technical. What human realities are hidden behind those words?
Michael: The first thing to understand is that this is not really “artificial” intelligence. It is human intelligence. There is nothing artificial about the intelligence behind these systems. There is pure human intelligence behind the screens.
What looks like magic is actually human labour. I would say we are the human magic behind the machines. That human aspect cannot be removed from this conversation, even if many big tech companies try to hide it.
This workforce exists in many different categories. One person may be working on captions for videos, another on drone footage, another on security data, another on medical data. Every system is developed for a specific need. If a country wants to develop a security system, it will need one kind of data. If someone is developing a medical tool, they will need another. The work is very diverse, but in every case workers have responsibility for how the world is classified and interpreted by these systems.
What are the most urgent issues that the public and policymakers should understand about data work, especially across Africa?
Lucy: The first thing is that people should stop believing the narrative that AI is fully automated. Artificial intelligence cannot reason on its own. It cannot function without humans in the loop.
There are many examples. One day you label a person standing upright as a human. Another day, the system sees a person in a wheelchair, or a person crawling because they do not have access to a wheelchair. The system may mistake that person for something else if it has not been trained properly. We have corrected these kinds of cases many times.
So if AI cannot function without human beings, and if these systems are making billions, then the workers should be taken care of. But companies outsource this work to countries where there are fewer protections, where no one checks whether minimum wage is being paid, and where workers have very little power.
Many workers do not have contracts. They do not have social benefits. The working conditions are very poor
Sometimes we are paid so little that we cannot even afford to go to work. We cannot afford transport or food. Many workers do not have contracts. They do not have social benefits. The working conditions are very poor.
I used to be a trainer, and sometimes we did not even have proper workstations. I had to stand for 14 hours straight, from 7 a.m. to 9 p.m., because as a trainer I was expected to move around and train hundreds of people. These are the kinds of conditions that need to be made visible.
It is not that we do not want the jobs. We want the jobs. But if these companies bring jobs here, they should bring decent jobs. They should pay at least minimum wage, or above it.
Michael: For lack of a better word, the conditions are pathetic.
If you are developing a system that is going to be used in Europe or the United States, why is it so hard to pay the people doing the work properly? As far as we are concerned, many of these technologies are not even developed for us. We are building systems for others, yet we are paid as if the work has no value.
Companies also set conditions for workers without taking responsibility for them. They tell you that you need a laptop with specific specifications, maybe 16 GB of RAM, SSD storage and a certain internet speed. But they do not provide the laptop or the internet. So to work, you have to invest your own money.
Because the pay is so low, people end up taking multiple jobs. Sometimes workers stay online for 20 hours a day just to earn something reasonable. We have seen cases where someone made only $0.10 in a whole week. We have also seen dirty or graphic content paid at $0.30 for 1,000 tasks. How long does it take to complete 1,000 tasks? How do you call that anything other than modern slavery?
And then there is discrimination. Sometimes workers in Kenya are paid $0.30 for 1,000 tasks while someone doing similar work in the United States may be earning $50 per hour. The targeting is strategic.
A lot of the most disturbing work is done here. Workers view graphic content, sometimes for long periods, with no mental health support. I know of cases where people became deeply affected by what they saw. But there are no protections, no statutory provisions, and companies use gaps in the legal framework against workers.
How did you begin organising and building advocacy around this work?
Lucy: We started locally in Kenya. It would not have been possible to begin by organising across every country at once. We first needed to grow locally, and now, with the momentum we have, we are beginning to engage others. We already have members from other countries, even if they are not yet actively involved.
We also connect with other platform workers. For example, I had the opportunity to go to Geneva for the International Labour Conference, where I met workers from other sectors: ride-hailing workers, nanny associations and others. We realised that we cannot fight this alone.
The companies may not always be exactly the same, but the model is similar. We need global campaigns. We need unity among platform workers, whether they are in data work, care work, ride-hailing or other sectors.
With the passing of Convention 193, we need a global campaign to push governments to implement these rules and improve working conditions.
Many labour abuses remain invisible because they are difficult to document. How do you collect evidence and testimonies from workers?
Michael: We have faced many challenges with documentation. Not everyone wants to come out publicly because they fear victimisation. So we try to gather stories carefully and quietly, making sure we understand the space before making things public.
There is also retaliation. Sometimes, after violations are raised, companies change jurisdictions. That makes it difficult to pursue legal action because they are no longer operating under the same jurisdiction. We saw this with the Meta case in Kenya, when more than 1,100 workers were laid off. After reports came out about the work, the company changed jurisdiction.
Some platforms monitor workers through cameras during the entire shift. If someone passes behind you, that may be enough for you to lose your job
We also have to consider NDAs. Workers are told not to talk about their work, sometimes not even with their partners. Imagine working from home, but your partner is not supposed to know what you are doing. Some platforms monitor workers through cameras during the entire shift. If someone passes behind you, that may be enough for you to lose your job.
This isolates workers from society. It makes them easier to control.
As an association, we know these issues are sensitive. We do not want to move recklessly, because people can lose their jobs. When more than 1,100 people lost their jobs, we were not in a position to offer an immediate solution. But the fact that workers spoke about their conditions and the company responded by leaving shows that these practices are deliberate.
That is why we did not restrict the association to one country. Our membership platform is open to workers across the world. But as a young organisation, we also need to build strong structures before expanding more widely.
Who becomes a data worker? Is this mainly young people, or does it cut across generations?
Lucy: It cuts across the whole demographic. During my time as a tasker, trainer and team lead, I trained all kinds of people. Sometimes whole families came: a mother, a daughter, a cousin.
Unemployment in Kenya is very high, and I think that is one of the reasons big tech companies outsource cheap labour here. Anyone who does not have a job may try to join these platforms.
The platforms present the entry process as easy. You just sign up and get enrolled. But once you are inside, you realise that the work has many requirements. You have to understand long training materials, often written in complex English, and translate that into practical tasks.
Many people drop off because training is unpaid. If training is on-site, workers have to travel every day, feed themselves and still support their families. Older workers often cannot stay on the platforms for long because they have more bills and responsibilities. Sometimes they leave to do cleaning jobs that pay better.
But many people, especially mothers and fathers, are very persistent. I trained one project for three straight months. Workers were promised live tasks, which are the tasks you complete and get paid for. But management kept saying they were not ready because the project was “sensitive”. Workers waited for months, hoping they would finally be paid. Eventually many left frustrated.
So you might see many young people in the sector, but it is not only young people. It affects many kinds of workers.
Are data workers uniquely positioned to identify risks, harms or blind spots in AI systems?
Michael: Yes, but many workers are vulnerable and may not know when their rights are being violated. Sometimes they know, but they ignore it because they need the money.
For example, there are projects where workers are asked to take pictures of people between 60 and 70 years old, or babies between nine months and two years old. A worker may do that without realising it is a violation. They may only understand later, when they enter advocacy spaces, that what they were asked to do was wrong.
These companies often target vulnerable groups. In some training schools, workers were asked whether they came from a slum. If you did not come from a slum, you might not be selected. That tells you something. They target people who are desperate for work and less likely to challenge violations.
Do workers know how the data they label is eventually used?
Lucy: No. There is almost no transparency.
I cannot tell you how any data I worked on has been used. Companies are not transparent about what they do with our productivity.
For example, there was a project where we had to take selfies of ourselves. The selfies had many specifications: in a dark room, in a well-lit room, wearing a helmet, without a helmet, showing your hands, from different angles. We had to submit batches of around 15 photos.
I cannot tell you how any data I worked on has been used. Companies are not transparent about what they do with our productivity
At the time, we were told we would be paid around $10, which sounded like a lot of money just for taking selfies. So people did it. Some of us recommended it to friends. But after submitting, the company would say that one image did not meet the requirements, so the whole task was rejected. They already had the photos, but nobody was paid.
Where did those photos go? What were they used for? We were never told.
Other projects ask people to take pictures of babies. That is a violation, but many workers do not know their rights. Companies take advantage of that lack of knowledge.
Even the names of projects are hidden. We would work on projects with names like Austria, Porcupine, Zebra or Giraffe. These names tell you nothing. You do not know whether your work is being used for Tesla, smart vacuum cleaners, medical robots, drones or something else.
If our labour helps create products that make billions, do we not have some form of intellectual property relation to that? At minimum, we need transparency.
Michael: People are still in the dark because AI is presented as magic. But AI is never magic. There is no machine that has ever built itself. It was built by human beings. There is no machine that got information by itself. It was fed by human beings.
I also worked on a project called chat moderation. At first, I thought I would be responding to tickets or helping users with issues. But over time, the work changed.
We were told to chat with lonely people, people looking for love online. But we had to impersonate different identities. I am a man, but sometimes I had to respond as a woman. Sometimes I had to act as a gay person, or a lesbian, or an older woman who had lost her husband. I had to provide emotional attention, but under strict rules. I could not share my personal information. If I did, the system could detect it and suspend my account.
The person on the other side might be told they were talking to AI, but in reality they were talking to human beings. We were expected to type at least 50 words per minute and maintain the flow of conversations that were passed from one worker to another. If one worker stopped, the conversation moved to another available worker, who had to quickly read the previous messages and continue as if nothing had changed.
That is another form of disinformation. People think they are talking to AI, but they are talking to human workers who are paid very little, sometimes $0.05 per message.
Could AI be developed differently, in ways that benefit local communities?
Lucy: Yes, and we would really love that. Right now, we do not benefit locally from much of this work.
For example, with the selfies, we later realised they may have been used to develop facial recognition software. But we did not know that at the time. Someone else somewhere else may have been given our images to label. We never know.
In Kenya, I have never seen a self-driving Tesla. Maybe there is one somewhere, but I have never seen that technology being used here. I once ordered a smart vacuum cleaner online because I was excited. I had worked on similar things and thought it might help around the house. But what arrived was like a toy. It could barely fit in my palm and just kept rotating in the same spot while the cats ran around it.
We do not experience the benefits of these technologies. If companies are making billions from our labour, they should give back to the communities that help develop their products.
They could support AI systems that help with traffic, infrastructure, schools or public services. Workers could label images from our own roads and streets. That could help identify potholes, improve infrastructure and support local needs. But right now, most of the value leaves the country.
What kind of advocacy and training does the Data Labelers Association provide?
Lucy: We organise worker gatherings. We had one in February and another on the 13th of this month. Each time we bring workers together, we focus on different subjects.
We invite professionals such as lawyers to teach workers about policies and rights. We also invite therapists, because mental health is one of the biggest issues for data workers.
These companies isolate us. Workers sign NDAs, and many people around them cannot understand what they are going through. People break down. Some commit suicide. So we try to create safe spaces where workers can talk, share experiences, meet others and receive support.
Michael: When we did our mapping, we wanted to understand what the most urgent issues were. Mental health came out very strongly. That is why we created DLA Health, which focuses on mental health. We also have DLA Voice, which focuses on advocacy, and DLA Community, which focuses on engagement.
We bring workers together for sessions on mental health, civic education, rights and policy. We also engage partners, civil society organisations and government actors, because we need changes in policies and regulations.
Sometimes we face retaliation. We are told that we are pushing investors away. But our point is simple: let us sit down and address these issues
It has been a difficult journey. We are a young organisation, and funding is a challenge. Sometimes we face retaliation. We are told that we are pushing investors away. But our point is simple: let us sit down and address these issues.
Many people did not even know what data labelling was. They confused data labelling with content moderation or content creation, which are different things. So part of our work is simply making people understand what data labelling is, and why it matters.
How can people outside the sector support your work?
Lucy: People can help by naming and recognising the Data Labelers Association. Awareness is important, especially now, with Convention 193 being passed. We need people to contribute to the global campaign and push governments to implement protections for platform workers.
At the moment, the most important thing is to join us in the global campaign.
Michael: Lucy has said almost everything. This work should go a long way, and we cannot do it alone.
