How workers on platforms like Amazon's Mechanical Turk earn pennies to train AI
Each morning when she wakes up, Kristy Milland powers up her home computer in Toronto, logs into Amazon Mechanical Turk, and waits for her computer to ding. — Amazon Mechanical Turk (AMT) … Tweets: @wadhwa , @paulsingh , @alexrubalcava , and @brianroemmele . Thanks: @om Tweets: Vivek Wadhwa / @wadhwa : How half a million people are being paid pennies to train AI http://secure.techrepublic.com/ ... >Really sad. And they're training their replacements Paul Singh / @paulsingh : “an invisible, online workforce—one that is in demand for their vital role in helping train intelligent machines” http://www.techrepublic.com/ ... Alex Rubalcava / @alexrubalcava : How Amazon's Mechanical Turk works ... I always thought most of this was done overseas, not in the US. http://www.techrepublic.com/ ... Brian Roemmele / @brianroemmele : In many ways we are all Mechanical Turks. As time moves on this job will grow to every electronic device you own. Humans train AI. http://twitter.com/... Thanks: @om
Context & Ripple Effects
This 2016 profile of Kristy Milland's daily grind on Amazon Mechanical Turk captures crowdwork before it had an industry around it: half a million people waiting for a ding, paid per micro-task, much of it labeling data for machine learning. Vivek Wadhwa's widely shared framing — workers 'training their replacements' — made the human cost of AI's data pipeline a public talking point years before ChatGPT.
The related coverage traces what happened next. By 2018, Samasource was running managed annotation operations in Nairobi for clients including Google and Microsoft, while psychology researchers flagged bot contamination degrading data quality on Mechanical Turk itself. By 2023 the labor had professionalized into a hidden 'tasker' underclass working through firms like Scale AI, and TIME's investigation found OpenAI paying Sama workers in Kenya $2/hour to label toxic content — a wage floor barely above Milland's pennies-per-task era.
First-order effects
- Requesters on Mechanical Turk acquire labeled training data at near-zero marginal cost, while crowdworkers like Milland absorb all the income volatility — unpaid waiting time, rejected tasks, no benefits.
- Wadhwa's 'half a million people' framing puts Amazon's marketplace, not a lab, at the center of the debate over who bears the human cost of building AI systems.
Second-order effects
- Managed intermediaries — Samasource, then Sama and Scale AI — capture enterprise demand for annotation, but the corpus shows formalization changed working conditions far more than pay, with OpenAI's Sama contract still landing at $2/hour.
- Bot infiltration of MTurk surveys forces data buyers to add verification layers, raising the real cost per verified task and pushing quality-sensitive clients toward managed vendors who can vouch for their workforce.
Third-order effects
- If the pattern holds, AI rests on a two-tier structure: a visible model-building industry atop a hidden annotation labor market stretching from Toronto to Nairobi, whose workers are simultaneously the population most exposed to being automated out — a dynamic echoed in later accounts of tech workers fired as managers cited AI.
- Persistent subsistence wages across a decade of coverage point toward eventual pressure on enterprise buyers and regulators over labor standards in AI supply chains, though no enforcement framework appears anywhere in this corpus.
The trend: The humans who label AI's training data form a persistent, low-paid hidden layer of the industry, migrating from ad-hoc crowdsourcing on Mechanical Turk toward managed annotation firms in Kenya and elsewhere without escaping subsistence pay.