Workers powering the AI industry face terrible conditions, but they shouldn’t have to – interview 

By the time consumers get their hands on a brand-new product, the natural resources and human labour that go into its production are long forgotten.

Distance, time and marketing abstract the long supply chains that enable the flows of goods and services. However, even what feels like the simple act of engaging with an AI chatbot is only made possible by a global system that encompasses everything from cobalt mines in Africa, to data labellers in the Philippines, to the sunny California headquarters of leading technology companies.

In the AI industry, the biggest concentration of human labour is in data work, where workers, mostly based in low-income countries, label, annotate and refine the data foundational for AI models under generally very poor working conditions involving low pay, exposure to health risks and exploitative practices. Action-research project Fairwork, founded by University of Oxford professor Mark Graham, argues this does not have to be the case.

The project, started in 2018, has developed a series of principles – fair pay, fair conditions, fair contracts, fair management and fair representation – to work with organisations to improve the conditions of various digital work modes: location-based platform work, cloud work, sex work and AI. It audits companies’ supply chains to provide them with a score based on these principles, which then works as a benchmark of fair work for industry, consumers, workers and policymakers.

Recently, Graham spoke with Investment Monitor about the nature of data work, why it is sensible for companies to audit their AI supply chains ahead of regulatory crackdowns, the importance of consumer awareness and the feasibility of reform.

Getting ahead of regulation

What do other digital platforms such as Uber and Deliveroo have in common with the AI industry? When both of these sectors came into existence, Graham notes, they did so in lightly regulated business environments. This enabled them to grow and innovate, but it also “opened the door for some of the harms and risks, especially for workers, that we have seen as a commonality in both domains”, he says. Once these harms become common knowledge, a regulatory backlash eventually follows.

“One of the things we have really tried to do is help companies get ahead of the curve, because at the end of the day, these sorts of risks and harms for workers, they will never remain totally swept under the carpet […] It is really a question of, to what degree do individual firms or the sector as a whole want to get ahead of that and want to try and protect their businesses from some of the risks, for instance, of overzealous regulation that might then swing the pendulum back?”

Indeed, reports about the poor labour conditions faced by data workers employed by companies such as Meta and OpenAI (mainly through outsourcing centres) are becoming increasingly prevalent. In 2022, Time reported that Samasource, a California-based outsourcing company employed by Meta for content moderation work, was subjecting workers to mental trauma, union suppression, intimidation and low pay in Kenya. The company also provides data labelling services for AI and machine learning models.

In 2023, Sama invited Fairwork to audit its working conditions in Kenya and Uganda. Fairwork’s initial findings found that the company did not meet any of its principles, after which the team collaborated with Sama to enact changes. While the company received a five out ten score following some improvements, the most recent evaluation for 2024–25 saw this score drop to three.

In the EU, there has certainly been a regulatory backlash against tech companies. In the past few years, the bloc has handed out billions of euros in fines for alleged breaches of antitrust laws, data privacy rules and legislation such as the Digital Services Act and the Digital Markets Act. GlobalData’s recent Global AI Regulatory Landscape report reinforces Graham’s argument, finding that companies that are already aligned with international AI standards will have an advantage over their competitors when regulations become fully applicable.

Awareness of data labour work 

The type of awareness needed to create societal and political pressure for regulation targeting the AI supply chain is improving, but “it is not fully there yet”, Graham says. Still, reports such as the one in Time have continued and gained more notoriety. Once the idea the idea that exploitation is embedded in the AI supply chain sets in at a consumer level, he warns, it will be very difficult to shake off.

“A lot of people don’t fully realise just the quantity and quality of human labour that is involved in, not just producing, but maintaining that [AI] service they are using,” Graham outlines. “These are people in predominantly low-income countries, with absolutely terrible labour protections and mostly with terrible jobs.”

He draws a parallel with the fashion industry, in which companies did not foresee the onslaught of reporting about sweatshops in the late 1990s and early 2000s. By the time the reality that certain practices in a product’s supply chain are exploitative reaches the mainstream, the money and time that is spent on public relations and due diligence is a lot more than it would have been to “just address the issue at the root from the beginning”, he notes.

While ‘fast fashion’ and its associated problems still exist, awareness around poor business practices have made it harder for companies to look the other way. Shein, the Singaporean ultra-fast fashion online retailer, has had its IPO in London repeatedly stalled and faced major pushback in France over alleged human rights abuses.

Is low-data content work being phased out?  

In the past few months, there have been reports, both by media outlets and major AI companies, that suggest this type of low-level data work is being phased out by the industry’s top companies, as the focus shifts to building ‘smarter’ models. This might be the case for individual companies that have already completed low-level training for their products, “but, in the industry as a whole, absolutely not”, Graham says. As the sector continues to grow, so does the pipeline of new products that require the building of proprietary databases.

The managers of data work centres in Asia and Africa that Graham speaks to are aware that, in a sense, they are putting themselves out of a job by training machines to do work formerly performed by humans. However, demand is strong, and companies return to these centres to refine their models, try new ideas, fix issues in past data sets, and more.

“There is just a constant demand for this sort of work that I don’t see going away any time soon,” Graham says.

Race to the bottom 

Herein comes the dilemma for low-income countries that want to attract investment. There is high demand for low-level data work but also a lot of competition. Raising regulatory standards risks spooking investment away altogether.

“In the Philippines, I think there is very much an awareness that they are competing in a global market for these jobs, and so there is extreme downward pressure on wages and working conditions,” Graham outlines. Maintaining the regulatory balance to attract investment and have fair conditions is “very difficult”.

However, Graham stresses that there is a misconception that any reform will be expensive and complicated. Some of Fairwork’s suggested policy changes are straightforward, such as establishing transparent lines of communication with management, limiting shift lengths and clarifying the terms set out in initial contracts.

“It is not just ‘pay the workers more’, which is, of course, absolutely crucial for decent work […] There are dozens and dozens of things, based on years of research, that can make meaningful improvement to workers’ lives that don’t actually cost anything,” he says. “There doesn’t always have to be this concern on the part of either industry or government that empowering workers or providing rights to workers will necessarily mean that we as a firm have more cost to bear.”

Anti-regulatory sentiment  

Countries in the Global North have the capacity to lead this space. Talking to regulators in Germany, Graham says, there is an awareness that the country is “the apex of many supply chains”, and therefore has more influence in establishing better standards, which can then spread downstream. This is in the interest of regulators, he argues, as it can help prevent a “global race to the bottom” and local workers being undercut by markets with lower labour costs in other parts of the world.

This point is particularly important given the anti-regulatory sentiment growing in the EU, as evidenced by the recent omnibus, which weakened some of the bloc’s digital regulations. However, Graham still considers that the broad trend in the EU is towards establishing more transparency and accountability in supply chains.

“We are not going to go back to a world where, in a market like the EU, leading firms are given free rein to just say, well, this is none of our business,” he says.

This, along with a growing awareness of data work and the wide array of reforms available, are drivers of change in the industry.

“These are terrible jobs, these are miserable jobs, and these are extractive and exploitative jobs – and at the end of the day, there will be a sufficient number of people anywhere who will find them distasteful and want nothing to do with that,” he says. “These [reforms] are just small factors in the overall pie of production, but they can make a big difference to a lot of people.”

Unlock up to 35% savings on GlobalData reports

Use the code at checkout in the report store

• 20% OFF

  Buy 2 reports

  Use code:

  Bundle20

• 25% OFF

  Buy 3 reports

  Use code:

  Bundle25

• 30% OFF

  Buy 4 reports

  Use code:

  Bundle30

• 35% OFF

  Buy 5+ reports

  Use code:

  Bundle35

Valid on all reports priced $995 and above. Cannot be combined with other offers.

Still deciding what will work best for your business?

Ask our experts for help.

Enquire before buying