AI systems may be approaching a critical bottleneck — not from a lack of computing power or data, but from the disappearance of the very human experts they depend on to improve. According to a new analysis, the automation of entry-level knowledge jobs is quietly eroding the pool of skilled evaluators needed to train advanced models, raising the risk that future AI progress could stall.
“We are investing billions in making models smarter, but almost nothing in preserving the human expertise that allows them to learn,” said Dr. Elena Vasquez, a computational sociologist at the Institute for AI Workforce Dynamics. “If this trend continues, we could see entire fields of knowledge atrophy from within.”
Background: The Expertise Pipeline Under Threat
The issue stems from a simple dependency: to keep improving in knowledge work, AI systems need either a reliable mechanism for autonomous self-improvement or human evaluators capable of catching errors and generating high-quality feedback. Industry efforts have overwhelmingly focused on the first — pouring resources into reinforcement learning and self-play — while neglecting the second.
Major tech companies have cut new graduate hiring by half since 2019. Tasks once handled by junior professionals — document review, first-pass research, data cleaning, code review — are now performed by AI models. Economists call this displacement; companies call it efficiency. But neither side is paying attention to the long-term consequence: the next generation of experts is not being trained.
“Every time we automate a junior role, we lose an opportunity for someone to develop the judgment that makes a senior expert valuable,” Vasquez explained. “We are eating our own seed corn.”
Why Self-Improvement Has Limits in Knowledge Work
Proponents of autonomous AI often point to successes like AlphaZero, which mastered chess, Go, and Shogi through self-play without human data. In the 2016 match against Lee Sedol, the famous “Move 37” — a move no human professional would have played — emerged purely from AI self-discovery.
But such breakthroughs depend on stable environments with clear rules and immediate feedback. The game of Go has a fixed state space, unambiguous rules, and a perfect reward signal: win or lose. Knowledge work has none of these properties.
In legal, medical, or financial domains, rules are constantly rewritten. A legal strategy that worked in one jurisdiction may fail in another. A medical diagnosis may take years to confirm. “Without a stable environment and an unambiguous reward signal, you cannot close the loop,” Vasquez said. “You need humans in the evaluation chain to continue teaching the model.”
The Formation Problem: A New Kind of Knowledge Loss
History offers examples of knowledge dying — Roman concrete, Gothic construction techniques, lost mathematical traditions. But those collapses were caused by external forces: plague, conquest, institutional breakdown. What is happening now is different.
“We are seeing fields atrophy not from catastrophe, but from a thousand individually rational economic decisions,” Vasquez noted. “Each layoff, each automation of a junior role is sensible on its own, but collectively it is hollowing out expertise.”
The AI systems being built today were trained on the expertise of people who went through a rigorous formation process. That pipeline is now being dismantled. The result could be a future where AI continues to generate plausible-sounding answers, but lacks the grounding that only deep human judgment can provide.
What This Means: A Call for a New Priority
The analysis argues that the human evaluation problem deserves the same investment and rigor as building model capabilities. Without deliberate intervention — such as preserving entry-level pathways, investing in human-in-the-loop systems, or creating synthetic training environments that mirror real-world complexity — the AI industry could hit a wall.
“We are not advocating for stopping automation,” Vasquez said. “But we need to recognize that knowledge work is not like a board game. If we remove the ladder by which experts are formed, we will all be stuck on the ground.”
For now, the warning is clear: the efficiency gains of today may come at the cost of tomorrow’s expertise. And unlike a chess match, there may be no second chance to recover what is lost.