Amid all the talk about artificial intelligence (AI) both creating and destroying jobs, a troubling reality flies under the radar.
The tasks machines can’t perform well are often offloaded onto marginalised global workers who are struggling in precarious labour markets. They do ostensibly “automated” work under exploitative conditions.
Data work is an essential part of building and refining AI systems. Before AI models can “learn” anything, human data workers must categorise, label, test and moderate vast volumes of text, images, audio and video, to make the data usable for AI training.
This labour is performed by an expanding global digital workforce that prepares the datasets not only for big tech, but also high-stakes industries such as banking, insurance, healthcare and government agencies, including defence.
To understand the AI workforce, I have been interviewing workers in China and Australia who prepare datasets for AI models. The fieldwork is ongoing, but here’s what they’ve revealed so far.
Inequality is baked in
My interviews with ten people to date show that precarious labour markets and marginalised social status have pushed digitally literate young workers into the data labelling industry.
As one interviewee said:
We do the manual work so that they get the credit for the intelligence.
There’s a lot of inequality across the data labour market, shaped by people’s qualifications and geographic location.
Those with PhD-level or equivalent qualifications and STEM certifications can typically get more specialised tasks. If based in the Global North, such workers tend to be higher-paid, earning A$400–800 per hour depending on the task.
But such specialised and high-paid tasks are rare and difficult to get. Most workers I interviewed perform general tasks, such as repetitively drawing bounding boxes for images used in drones, self-driving cars and automated vending machines, or annotating audio.
These workers normally receive as little as A$6 per day or even less. The pay can’t cover daily expenses, and the long hours leave workers with chronic eye strain and back pain.
Part of the gig economy
Data work is not unlike other poorly regulated jobs in the gig economy.
Workers have no formal contracts and are not employees. They’re classified as “users”, and platforms simply call on them when there are tasks aligning with their expertise and track record.
User agreements exist primarily to protect the companies behind the outsourced work, such as requiring the workers don’t disclose any of the information they see.
This is despite the fact that datasets are already anonymised: workers often have no way of knowing which companies they conduct data labelling for. They don’t even know if humans or AI agents assess their completed work. And they have minimal rights to appeal any assessment of their performance.
All interviewees reported getting less work over time as AI advances. What’s left are more difficult and time-consuming tasks. Interviewees expressed little concern about their jobs eventually being replaced by AI, but this apparent indifference stemmed from a pessimistic outlook:
If I don’t make this money, someone else will, and I will be replaced [by AI] eventually anyway.
As one worker noted, what AI actually affects is the working class itself. This working class is expanding as more professionals are pushed into data labelling by the precarity of the current job market.
All work, little pay
How a worker gets paid is determined by the platform. US crowdsourcing platforms generally offer higher-paid tasks and pay workers when they submit the work.
Chinese platforms or companies often pay workers only after their tasks have been assessed and confirmed to meet preset standards. As a result, workers often spend hours completing tasks without receiving any payment.
In addition, workers in China can’t access US platforms; using a VPN to circumvent this risks triggering an account ban.
Companies prefer consistency in their workforce, as turnover is costly; workers require instruction and training before they can begin a task, and further time before they can complete tasks efficiently.
As workers typically get faster the longer they stay in the role, companies want to retain the experienced ones. But many workers leave because the pay is so poor.
To offset this, companies have turned to recruiting more vulnerable groups. One example is collaborating with local government initiatives supporting disabled people. These workers are less likely to quit because the job is often their last resort.
Workers reported they were unable to find other employment or were in the process of searching for full-time positions due to disability, pregnancy or being recent graduates.
The bigger picture is grim
The AI economy has created jobs. But many of these involve human workers correcting errors and handling tasks too difficult or ambiguous for machines to resolve. This work is often more cognitively and emotionally demanding than what it replaced.
And human workers don’t even know if they’re answering to human managers or AI agents. This weakens their right to bargain.
Data workers are effectively the disposable batteries of the AI economy: drained of every last charge, then discarded once they can no longer power the system that depended on them.
Australia is accelerating the pursuit of an AI-driven economy. The crucial question is not how many jobs are created, but what kind of jobs they are.
The employment gains AI promises may only exist in the short term, and come at the cost of data workers’ life quality and wellbeing. We must establish protections and a long-term plan for this workforce, so we can prevent the harm rather than merely respond to it after the fact.
- is a Research Fellow at Melbourne Law School, The University of Melbourne
- This article first appeared in The Conversation
