We are accidentally automating the training out of our future

The "automation paradox" isn't just a theoretical concept for pilots or nuclear engineers anymore; it is becoming a structural crisis for the entire professional workforce. I remember designing controls for a digital system for a U.S. nuclear plant over a decade ago. We made a counter-intuitive engineering decision: we deliberately programmed manual steps into sequences that the machine was perfectly capable of doing itself. To an efficiency-obsessed manager, it looked like a mistake. To an engineer, it was a safeguard against "cold hands."

If an operator only ever supervises automation, they stop being an operator. Their mental model of the system decays. When the automation eventually fails—and it always fails when it's confused or overwhelmed—the person in the chair is left holding the reins of a machine they haven't actually "driven" in years. We built in inefficiency to preserve human competence.

Now, we are seeing this exact failure mode play out across the white-collar economy through LLM deployment and AI workflows.

The disappearing entry-level role

The data suggests we are already seeing the "junior employment gap" widen. A Harvard working paper looking at 65 million workers across 280,000 firms found that following generative AI adoption, junior employment dropped by about 9% within six quarters compared to companies that didn't adopt the tech. Meanwhile, senior employment stayed stable or grew. Stanford researchers found a similar trend in ADP payroll records: the youngest workers in AI-exposed occupations lost ground after late 2022, while experienced professionals held their positions.

There is a nuanced distinction here that people often miss:

• Automation-heavy roles: Junior employment drops as the AI takes over the "grunt work."
• Augmentation-heavy roles: Junior employment remains steady or actually rises.

The problem is that the "grunt work" is precisely where the apprenticeship happens. Whether you blame AI or the rise of remote work, the result is the same: the channel through which expertise passes from senior to junior is being severed.

Expertise cannot be downloaded

We need to stop treating expertise like a file that can be transferred via a prompt. You don't become a senior engineer, a lawyer, or a developer by simply reviewing the output of a high-performing model. Expertise is earned through the friction of real-world application: the failed builds, the grueling debugging sessions, and the "why did that actually work?" moments.

An AI is designed to spare you that friction. It provides the answer immediately, bypassing the struggle. But if you bypass the struggle, you never develop the "gut sense" required to know when a model is confidently, catastrophically wrong.

If we continue to optimize for immediate output efficiency, we are essentially cannibalizing our future talent pipeline. We are creating a generation of "supervisors" who have the theoretical knowledge to manage a model on paper, but lack the foundational experience to intervene when the machine hits a wall. In safety-critical fields, we've learned that efficiency is a trap if it costs you the ability to handle the unexpected. We are currently repeating that mistake on a global scale.