The Last Thing You'll Ever Teach

There's a particular kind of workplace request that sounds like opportunity. A manager pulls someone aside, explains that the company is investing in new technology, and asks if they'd be willing to help document their workflow. Maybe record a few screen sessions. Walk through the decision tree they use when a customer calls with a billing dispute. It's framed as a chance to shape the future. To be part of something. The employee says yes, because of course they do.

That scene is playing out right now across business process outsourcing centers in Hyderabad and Chennai and Manila. Workers are sitting in front of cameras and annotation tools, carefully explaining the logic behind tasks they've performed for years. They are, in the most literal sense, writing their own replacements. Al Jazeera documented this pattern in June 2025, and the phrase that stuck with me was almost bureaucratically neutral: workers "training AI robots to take their jobs." Not a metaphor. A job description.

The scale of what's actually happening is worth sitting with for a moment. Spain's government projects a net loss of 400,000 jobs between 2025 and 2035 from AI adoption, a number drawn from structural economic modeling, not think-tank speculation. India's IT services sector employs roughly five million people in roles that involve exactly the kind of structured, documented, repeatable knowledge work that current AI systems are best at automating. The Philippines built an entire middle class on BPO contracts. These aren't abstract futures. The displacement is running now, and the mechanism is specific: companies are using existing workers to capture institutional knowledge before those workers exit. It minimizes transition cost. It's cheaper than hiring AI specialists to reverse-engineer the workflows from scratch. And it lets the company describe the process as "upskilling" in the press release.

This is not corporate malice. That's the part that makes it hard to argue against. It's the mathematically rational move for any firm with margin pressure and access to AI tooling. The worker who refuses to participate doesn't save their job. They just become less useful during the transition and get cut first. So they participate. And their participation accelerates the outcome for everyone else in the same role. Game theorists have a name for this structure, and it's one of the most durable concepts in social science precisely because it describes situations where individual rationality produces collective catastrophe.

The prisoner's dilemma was formalized in 1950 by researchers at the RAND Corporation, which is its own kind of irony given RAND's current role in AI policy research. The original framing involved two suspects who couldn't communicate, each deciding whether to cooperate with the other or defect to the authorities. If both stay quiet, both get a light sentence. If one talks, the talker goes free and the other gets the maximum. If both talk, both get hammered. The dominant strategy, for each individual, is to talk. The collective outcome is the worst possible one. No coordination mechanism exists to prevent it.

The workers in Hyderabad aren't talking to each other either. They're in separate contracts, separate companies, separate cities. There's no union meeting where someone stands up and says, "What if we all just stopped documenting our workflows?" Even if there were, the person who stopped first would just be the one who got cut first. The structure of the situation makes cooperation against the outcome essentially impossible, which is why naming the structure matters more than assigning blame. You can't shame a company out of a dominant strategy. You can only change the payoff matrix.

History has a few examples of labor transitions where the mechanism was similarly invisible at the individual level. The handloom weavers of early industrial England didn't lose their livelihoods because mill owners were uniquely cruel. They lost them because each individual weaver who took a job at the mill, even at lower wages, made the economics of handloom weaving slightly worse for everyone who stayed. The Luddites understood this, which is why they attacked the machines rather than the mill owners. They were wrong about the tactic and right about the diagnosis: the machine was the coordination point, and destroying it was the only way to change the math. They failed, obviously. But they failed because they were operating against a system where the incentives were already locked in, not because they were irrational.

The framework I keep returning to is this: AI deployment at scale isn't a technology story. It's an incentive architecture story. The technology is the mechanism. The architecture is what determines who bears the cost of the transition and who captures the benefit. Right now, the architecture is almost perfectly designed to transfer cost onto workers and benefit onto capital. Companies that report earnings with explicit AI-driven headcount reductions are being rewarded by investors. The displacement funds the AI development, which justifies further displacement. The feedback loop is self-reinforcing, and the workers participating in their own replacement are not victims of deception so much as participants in a system where every available option leads to the same place.

What the Spain number and the India pattern together reveal is that we've crossed a threshold. The conversation about AI and jobs has been dominated for three years by the question of whether displacement would happen at scale. That question is answered. The new question is about the speed of policy response relative to the speed of deployment, and if you look at how quickly Spain's projection became a news item versus how quickly the EU will actually legislate a coherent response, you can form your own view on that race.

The worker in Chennai who spent last Tuesday recording herself triaging customer complaints made a reasonable choice with the information she had. So did her manager. So did the executive who approved the AI deployment budget. Every individual decision in the chain was defensible. The aggregate outcome was not something anyone chose, exactly. It was something the structure of the situation produced.

I keep thinking about that manager pulling someone aside to say this is an opportunity. Maybe they believed it.

Sources: [Al Jazeera, "India's workers are training AI robots to take their jobs" (June 2026)](https://www.aljazeera.com/gallery/2026/6/11/photos-indias-workers-are-training-ai-robots-to-take-their-jobs); [The Corner, "AI could result in net loss of 400,000 jobs in Spain between 2025 and 2035"](https://thecorner.eu/news-spain/ai-could-result-in-net-loss-of-400000-jobs-in-spain-between-2025-and-2035/126183/); [Platformer, "How to help knowledge workers who lose their jobs to AI"](https://www.platformer.news/how-to-help-knowledge-workers-who-lose-their-jobs-to-ai/); [The Conversation, "Canada's 'AI for All' strategy has ambitious growth targets, but it falls short on workers and the environment"](https://theconversation.com/canadas-ai-for-all-strategy-has-ambitious-growth-targets-but-it-falls-short-on-workers-and-the-environment-284648)