AI's dirty secret: models run on A$6/day human labelers

Every production AI model you ship was trained on data that humans first had to label, test and moderate. New research shows that hidden annotation workforce is paid as little as A$6 a day, while only a tiny elite of PhD-level specialists in the Global North earn A$400–800 an hour. For ML teams, that labour reality is a data-quality, governance and supply-chain problem hiding inside every benchmark.

Amid the AI job-creation debate, new field research from The Conversation exposes a reality that rarely makes headlines: the humans who make AI appear intelligent are often marginalised, digitally literate young workers doing tedious, temporary data-labelling tasks under precarious and sometimes exploitative conditions. Researcher Elise Racine has been interviewing data workers in China and Australia, with ten interviews completed to date and fieldwork ongoing. The central finding is that the much-hyped "AI job boom" is, for many participants, a low-paid, insecure gig rather than a high-skill opportunity.

“

New research shows that hidden annotation workforce is paid as little as A$6 a day, while only a tiny elite of PhD-level specialists in the Global North earn A$400–800 an hour.

Data work is not peripheral to AI; it is foundational. Before models can "learn" anything, people must categorise, label, test and moderate enormous volumes of text, images, audio and video. This hidden global workforce prepares datasets not only for big tech but for high-stakes industries including banking, insurance, healthcare and government agencies such as defence. The outputs flow into systems that approve loans, triage patients and inform defence applications, making the conditions of data work a matter of public interest, not just labour economics. Yet this labour remains largely invisible in official employment statistics and in the marketing of "automated" AI products.

The research reveals a starkly two-tier labour market. At the top, workers with PhD-level or equivalent qualifications and STEM certifications can access more specialised tasks; if they are based in the Global North, they can reportedly earn A$400–800 per hour depending on the task. Racine notes, however, that these specialised, high-paid assignments are rare and difficult to obtain. At the bottom — where most interviewed workers sit — are 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, a sum the research states cannot cover basic living costs.

Who ends up in this work matters. The interviews indicate that precarious labour markets and marginalised social status push digitally literate young people into data labelling. Pay and task quality are shaped by qualifications and geography, producing what one worker summarises as: "We do the manual work so that they get the credit for the intelligence." That single line captures the asymmetry at the heart of the AI supply chain: the "intelligence" branded by AI firms is, in significant part, purchased manual labour rendered invisible.