.jpg)
Whether automation will make human workers obsolete depends on more than just how smart the AI is.
In 2016, AI pioneer Geoffrey Hinton famously said that people should stop training radiologists because deep learning would soon outperform them. He was half right. Today, the FDA has approved over 1,000 AI tools for radiology, many of which match or exceed human accuracy in analyzing medical images. Yet human radiologists are in higher demand than ever. Since 2016, their numbers have grown by 17%, vacancy rates are at record highs, and average salaries have jumped from roughly $350,000 to $570,000—making radiology one of the best-paid medical specialties in the U.S.
Fears that AI will wipe out entire professions are widespread. Anthropic CEO Dario Amodei has predicted that AI could soon eliminate half of all entry-level white-collar jobs. But the radiology example shows that predicting AI’s impact on any given career is more complex than it seems. Three key questions can help clarify how vulnerable a job really is.
1. Is your job a weak bundle or a strong bundle?
Economist Luis Garicano and co-authors of the forthcoming book *Messy Jobs* argue that most white-collar work mixes two types of tasks: “clean” and “messy.”
- **Clean tasks** are predictable, data-heavy, and objective (e.g., processing expense reports or updating spreadsheets). These are highly vulnerable to AI.
- **Messy tasks** involve ambiguity, subjective judgment, tacit knowledge, and human relationships (e.g., calming an angry client, choosing a brand strategy, or leading a team). AI still struggles here.
The critical factor is how tightly these tasks are bundled together.
In **strong bundle** jobs—like trial lawyering—the clean and messy elements are deeply interconnected. A lawyer’s trial preparation (reading cases, analyzing facts, drafting arguments) directly enables her performance in court, where she must improvise, cross-examine, and adapt in real time. Offloading too much prep to AI would undermine her ability to handle the unpredictable courtroom dynamics.
In **weak bundle** jobs—like recruiting or software development—tasks can be separated more easily. AI can now screen résumés efficiently, freeing recruiters to focus on relationship-building, interviewing, and negotiation. Similarly, developers can delegate routine coding to AI and concentrate on higher-level architecture and problem-solving.
Automation of weak-bundle jobs doesn’t automatically mean fewer workers. The outcome depends on broader economic effects.
2. If what you produce got cheaper, how much more of it would people want?
History shows that automation often triggers the **Jevons paradox**: greater efficiency lowers costs, boosts demand, and ultimately increases employment in that sector.
- Henry Ford’s assembly line slashed car prices, exploding demand, and roughly doubling U.S. auto industry employment over 35 years.
- Similar patterns occurred with power looms in textiles, ATMs in banking, and spreadsheets in accounting.
Early AI trends echo this: recruiter job openings rose 30% from 2023 to 2025, software engineer postings doubled, and call-center staffing is growing even as AI handles basic queries. Cheaper AI-driven services in finance, law, and healthcare could drive similar surges in demand and hiring.
However, demand doesn’t always rise indefinitely. Mechanized farming drastically cut food prices, but the share of Americans working in agriculture plummeted from ~40% to ~1% because people can only eat so much. In sectors with inelastic demand, automation poses a greater risk of job losses.
3. Is AI the expert, or are you?
Even when automation occurs, its effect on wages and job quality depends on whether technology enhances or commodifies human expertise.
MIT economists David Autor and Neil Thompson highlight the contrasting fates of accounting clerks and inventory clerks after computers arrived:
- **Accounting clerks**: Computers automated routine calculations, allowing survivors to focus on higher-value analysis (explaining variances, diagnosing problems). Employment fell, but remaining jobs became more professional and better paid.
- **Inventory clerks**: Computers replaced their specialized knowledge of stock, turning the role into lower-skill scanning and restocking. Employment grew, but pay declined.
Across hundreds of occupations, technology that amplifies expertise tends to create higher-quality (if fewer) jobs. Technology that deskills workers tends to expand lower-paid roles.
With AI, the picture is still evolving. Senior tech roles are currently growing while entry-level ones stagnate, but AI tools—like Schneider Electric’s “electrician’s assistant”—are already helping less-trained workers tackle complex problems. As AI improves at expert-level judgment, more professions could see their expertise “leveled.”
Applying the Framework
Radiology illustrates how the three factors interact favorably: it’s a strong bundle (image interpretation is intertwined with patient context and clinical consultation), it benefits from rising demand as scans become cheaper, and AI has so far enhanced rather than replaced radiologists’ expertise.
For my own work as a journalist, the outlook is mixed. Journalism hasn’t seen strong demand growth despite lower costs. Research can be accelerated by AI, but writing nuanced analysis still feels like a strong bundle—hard to fully separate from the messy human elements of interviewing and storytelling. For now, that interdependence offers some protection.
The Limits of Prediction
History reminds us that technology’s effects are rarely straightforward. The ATM didn’t kill bank tellers—widespread mobile banking via the iPhone did. Many of AI’s biggest impacts will be similarly unexpected.
By asking these three questions—about task bundles, demand elasticity, and expertise—we can move beyond simplistic “AI will replace everything” narratives and develop a more nuanced understanding of how this powerful technology will reshape work.