AI Is Stealing Your Tasks, Not Your Job — Yet | Politics

Employment rates are near all-time highs despite AI. But the structure of work is shifting fast. Here's what three new job archetypes tell us about surviving the transition.

Geoffrey Hinton once predicted AI would make radiologists obsolete. Radiologists are now in higher demand than ever. That's not a fluke — it might be the most important data point in the entire debate about AI and jobs.

The Numbers Don't Lie (Yet)

If you've been waiting for the labor market apocalypse, you're still waiting. Prime-age employment rates in the US are hovering near all-time highs. A recent survey of corporate CFOs found "little evidence of near-term aggregate employment declines due to AI." A parallel survey of European firms reached the same conclusion: productivity is up, headcounts are not down.

The most rigorous data comes from economists Humlum and Vestergaard, whose 2026 study tracked workers in AI-exposed occupations across Denmark. Their findings are striking in their precision: despite widespread chatbot adoption, measurable productivity gains, and significant task reorganization, the effect on wages and recorded hours was statistically indistinguishable from zero — ruling out changes larger than 2% over two years post-ChatGPT. AI is reshaping what people do at work. It is not, at least not yet, eliminating the people doing it.

The pattern emerging from the data is consistent: AI replaces tasks, not jobs. Software engineers who spent most of 2024 writing code now spend most of 2026 reviewing and maintaining code written by AI. The job title survived. The job description did not.

Why Some Jobs Are Stickier Than Others

This is where the economics gets genuinely interesting. A new theoretical framework from Luis Garicano, Jin Li, and Yanhui Wu offers a clean explanation for why some occupations resist automation far longer than others. Their key concept: the difference between strongly bundled and weakly bundled jobs.

In a weakly bundled job, the tasks that make up the role are separable. You can hand some to an AI and keep others for a human without the whole thing falling apart. These jobs are vulnerable — not immediately, but structurally. In a strongly bundled job, the tasks are so interdependent that the same person needs to do all of them for the work to function properly. Automating one piece doesn't help much if the rest still requires a human.

The radiologist case fits perfectly. Yes, AI can read scans with impressive accuracy. But scan-reading is only one piece of a radiologist's actual job. Patient consultation, clinical judgment calls, coordination with surgeons, liability — these tasks are tightly woven together. Strip out one and the whole bundle unravels. Garicano et al. predict that strongly bundled jobs will resist automation until AI capabilities become extraordinarily comprehensive, not just narrowly excellent.

Advertise with Us

[email protected]

These are the Specialists — and they're the easiest archetype to understand. A blogger, a therapist, a trial lawyer, a jazz musician. Jobs where the human perspective, the relational context, or the holistic judgment is inseparable from the output itself. AI can assist, but it can't substitute, because the bundle won't hold without the human at the center.

The Return of the Generalist

But what about weakly bundled jobs — the ones that are vulnerable? This is where the most counterintuitive insight of the current moment emerges.

AI's capabilities are famously "jagged." It's extraordinarily good at some things and surprisingly bad at others, and crucially, which things shifts constantly. Claude Code didn't just improve code-writing gradually — it effectively replaced the task almost overnight. Companies that hired engineers for their coding ability in 2025 suddenly needed to evaluate engineers for their ability to check and maintain AI-generated code. Those are different skills. They don't reliably co-occur.

The rational corporate response to this kind of unpredictable, shifting capability landscape isn't to hire narrower specialists. It's to hire people who can move fluidly between tasks — people whose core skill is figuring out what the AI is getting wrong today, and filling that gap. Cedric Savarese describes the arc: the initial euphoria of AI-assisted productivity gives way to a subtler, harder skill — developing a "mental model of the AI mind." Learning to recognize confidently incorrect outputs. Knowing when to push back and when to trust. That capacity for critical, curious, adaptive engagement with AI tools is not a technical skill. It's a disposition.

This is the Salaryman archetype — and the historical precedent is instructive. Japan's traditional corporate system rotated employees through HR, accounting, operations, product design, and back again. It was widely mocked for producing generalists with shallow expertise and contributing to notoriously low white-collar productivity. But in a world where AI handles the deep expertise and humans handle the gaps, the salaryman model suddenly makes structural sense. You don't need someone who is a great accountant. You need someone who can do a passable job at accounting when the AI accountant breaks down — and who will notice when it does.

There's a second-order implication worth noting: salaryman roles are sticky. A specialist's human capital is portable — take your skills to a competitor. A generalist's value is embedded in institutional knowledge, internal relationships, and familiarity with how a specific company's AI stack behaves. That makes generalists harder to poach and more likely to stay. Long job tenure, which has been in structural decline for decades in Western labor markets, could make a quiet comeback — not for sentimental reasons, but for economic ones.

What This Means for the Career Decisions Being Made Right Now

The hardest part of this analysis isn't theoretical. It's practical. Parents, students, and mid-career professionals are making decisions today — about majors, certifications, job offers — without knowing which tasks will still require humans in five years.

The honest answer from researchers who have spent years on this question is: nobody knows. Predictions about AI's impact on specific occupations have been falsified almost as quickly as they've been made. The "learn to code" consensus of the early 2010s is now a cautionary tale. Whatever today's equivalent is, it may age just as poorly.

What the evidence does suggest is a rough sorting mechanism. If your job involves tasks that are tightly interdependent with human judgment, relationships, or context — you're likely in a strongly bundled role, and relatively protected for now. If your job is a collection of separable tasks that could in principle be distributed across different agents, the question isn't whether AI will restructure it, but when and how fast.

For policymakers, the implication is that the coming disruption may look less like mass unemployment and more like mass dislocation — workers whose job titles persist but whose actual work transforms faster than retraining programs can track. The Denmark data suggests the transition can happen without visible wage effects for a while. That's reassuring and alarming in equal measure: reassuring because it suggests some cushion, alarming because it means the structural shift may be well underway before the aggregate statistics catch it.