How western AIs are exploiting millions of Indians to teach them how to think

Raw Large Language Models (LLMs) are chaotic. To make them helpful and safe, they require Reinforcement Learning from Human Feedback (RLHF). This means humans must read the AI's outputs, rank them for quality, and manually correct its mistakes. Every time a new model is released, its polish is directly attributable to millions of human hours spent tweaking its behavior, one prompt at a time.

This work isn't done in glossy California tech parks. It is outsourced through platforms like Scale AI, Remotasks (now Outlier), and Appen. In India, a sprawling network of contractors and gig workers log into these portals from cyber cafes, small outsourcing shops in Tier-2 cities, or their own phones. They draw bounding boxes around pedestrians for self-driving cars, categorize thousands of images, and grade AI essays for grammar.

To teach an AI what not to do, the AI must first generate horrific content — and a human must read it. Indian data workers are frequently tasked with “red teaming” or safety filtering. For 8 to 12 hours a day, workers are exposed to graphic violence, hate speech, child exploitation material, and self-harm instructions. They must meticulously label and categorize this trauma to ensure the AI knows to block it in the final consumer product, often with zero mental health support.

This is modern piecework. Workers are not salaried employees of the big tech firms; they are independent contractors paid by the task. Labeling an image or grading a paragraph might pay a few cents. Even working at breakneck speeds, many Indian annotators report making between $1 to $2 an hour (approx. ₹80 to ₹160). When tech giants boast about slashing AI training costs, it is achieved through this ruthless geographical labor arbitrage.

The workers are managed not by humans, but by algorithms. Platforms track “dwell time” — exactly how many seconds a worker spends on a single task. If a worker is deemed too slow, or if their quality score drops below a hidden threshold (often judged against other underpaid workers), the platform simply suspends their account. There is no HR department, no severance, and no appeals process.

This labor model creates a massive blind spot. Western AI models are heavily shaped by the values, biases, and cultural contexts of the underpaid Indian workers grading them. If a Silicon Valley prompt requires nuance about American political humor, an annotator in rural Uttar Pradesh is forced to guess, leading to the bizarre hallucinations and cultural disconnects frequently seen in global AI outputs.

The ultimate paradox of the RLHF economy is that these workers are building the very automation designed to make human labor obsolete. The more accurately an Indian annotator corrects an AI's code or categorizes a dataset, the less necessary that annotator becomes for future iterations. They are in a race to perfectly train the algorithm before the algorithm no longer needs them.