Somewhere in an office park in India right now, a worker is carefully labeling an image, correcting a transcription, or grading a chatbot’s answer. The data will feed a model. In a few years, that model may be the reason the worker’s role no longer exists. And the worker knows it.
That’s the bargain at the heart of a story NPR recently reported: employees in India are being enlisted to train their own AI replacements. It sounds like a paradox, or maybe a dark joke. But it’s a fairly honest description of how automation has always worked. Somebody has to teach the machine, and the people who understand the job best are usually the people doing it.
The arrangement is more common than the headlines suggest. AI systems don’t emerge fully formed. They learn from examples, and those examples need human judgment: this document is a receipt, that sentence is sarcastic, this medical scan shows a benign lesion. In practice, much of that judgment is supplied by workers in countries where labor is cheaper, often through outsourcing firms and data-labeling shops. What’s new is the explicit framing. Instead of pretending the training is neutral background work, some employers are telling staff directly that they’re helping build the systems that will eventually do their tasks.
You can read that cynically, and plenty of people do. It’s a way to extract knowledge from workers and hand it to software. But the more interesting question is why workers agree. The answer is usually the same one that explains most labor decisions: the job pays now. A paycheck this month beats a speculative worry about next year. And there’s a subtler incentive. Workers who become good at training AI often move into higher-skilled roles, reviewing model output, catching errors, managing quality. The ladder exists, even if it’s short and not everyone gets to climb it.
That said, the optimism deserves scrutiny. The tasks that get automated first tend to be the most routine ones, which are also the easiest to hand to a model. The workers best positioned to survive are those whose jobs involve judgment, context, and accountability, the fuzzy stuff models still struggle with. The workers most exposed are the ones whose jobs can be described in a manual. Training the replacement doesn’t change that math. It just makes the transition feel collaborative rather than imposed.
There’s also a structural issue nobody likes to discuss. When a company automates a role, the savings don’t automatically flow back to the people who made the automation possible. The worker who spent months labeling data doesn’t get equity in the model. The value of that labor is captured by the employer, and the risk stays with the worker. This isn’t unique to AI. It’s the same dynamic that has played out in manufacturing, customer service, and logistics for decades. AI just compresses the timeline and makes the handoff more visible.
The global picture complicates things further. India has become a major hub for this kind of work, and governments there see it as a growth industry. Training AI is a service export, and it employs a lot of people. The concern is whether it’s a durable career or a temporary bridge. Data labeling is already being automated in part, with models helping to pre-label data that humans then verify. The human role is drifting toward supervision and edge-case handling, which is more skilled but also employs fewer people.
For workers elsewhere, the India story is a preview. Companies in the US and Europe are increasingly using similar arrangements, sometimes offshore, sometimes through contractor networks. The pattern is consistent: automate the routine, keep humans for the exceptions, and ask the humans to help build the system that handles the routine. It’s efficient. It’s also a little unsettling, because it puts workers in the position of optimizing their own obsolescence.
None of this means automation is a disaster waiting to happen. Historically, technology has destroyed some jobs and created others, often in ways that weren’t obvious at the time. The workers training AI today may end up in roles that don’t exist yet. But the transition matters, and the terms of the bargain matter. If companies want workers to help build their replacements, they owe those workers something more than a paycheck and a vague promise about reskilling. They owe transparency about what’s being automated and when, and a genuine path to the next role.
The Indian workers in the NPR story seem to understand this better than the executives do. They’re not naive. They know what they’re building. They’re doing it anyway, because the alternative is worse. That’s not a ringing endorsement of the arrangement. It’s a description of how people survive it. And it’s the most honest account of the AI transition you’re likely to get from anyone actually inside it.