What is an hour of AI training labor actually worth? The honest answer, per Melbourne Law School researcher Fan Yang’s fieldwork reported by Mirage News on August 23, 2026, is: it depends entirely on your credentials and your zip code. Workers with PhD-level qualifications in the Global North can pull A$400-800 per hour on specialized tasks. Everyone else — the people drawing bounding boxes for drones and self-driving cars, or annotating audio — reported earning as little as A$6 a day, an amount Yang notes doesn’t cover basic living expenses.
That two-tier structure isn’t new to anyone who has followed this beat, but Yang’s interviews add texture that aggregate industry data can’t: workers classified contractually as “users” rather than employees, no formal right to appeal performance assessments, and — strikingly — no way to know whether a human or an AI agent is grading their work. That last detail matters more than it sounds. If a worker can’t identify who or what evaluates them, there’s no one to bargain with, no grievance process that means anything, and no leverage as the easy tasks dry up and only the hardest, most ambiguous ones remain — which Yang’s interviewees report is already happening.
A pay range running from A$6 a day to A$800 an hour isn’t a labor market — it’s a caste system dressed up as a gig platform.
The China-Australia comparison sharpens a payment-mechanics problem that’s been reported piecemeal elsewhere: US crowdsourcing platforms generally pay on submission, while Chinese platforms or intermediary firms often withhold payment until work clears a quality assessment, leaving workers laboring for hours with no guaranteed compensation. SOMO’s research, cited independently, found Amazon, Google, Meta, Microsoft, and Nvidia collectively route work through at least 30 intermediary companies — a fragmentation that lets Big Tech set pricing pressure and deadlines without directly employing anyone subject to those terms. Yang’s finding that companies increasingly recruit disabled workers and new graduates through local government programs because they’re less likely to quit adds a harder edge to that dynamic: retention isn’t solved by better pay, it’s solved by narrowing who else has options.
None of this is disclosed voluntarily. SOMO notes that Amazon, Google, and Meta all declined in 2025 to identify which annotation vendors they use, which is precisely why single-researcher fieldwork like Yang’s — however small a sample, ten interviews so far — carries weight: it’s one of few ways to get inside a supply chain that a market intelligence estimate cited by SOMO expects to reach $10.2 billion by 2034. Watch whether Yang’s completed study, or a comparable audit from Fairwork, produces employer-identified, quantified pay data at scale; until then, the $10.2 billion valuation and the A$6-a-day wage are two facts about the same industry that no one has yet been forced to reconcile in public.
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.
