The projections were wrong in both directions. The jobs that researchers most confidently predicted would be disrupted by AI have, in many cases, held up remarkably well through 2026. And several occupations that appeared safely removed from automation — jobs requiring judgment, credentials, and institutional trust — are being hollowed out in ways that the models missed entirely.
Understanding which predictions failed, and why, is more useful than simply reading the 2026 employment data at face value. The failures reveal something about the assumptions built into the forecasting models, and those assumptions will keep producing wrong predictions until someone changes them.
Start with radiology, the canonical example. AI-assisted diagnostic imaging has been cited as an imminent displacement since at least 2016. Andrew Ng’s famous statement — that AI will replace radiologists the way X-rays replaced diagnosticians who had to look at bones through the skin — became the ur-text of AI displacement commentary. In 2026, radiology employment in the US is flat. In the UK, it’s up slightly. The technologists were right that AI can read imaging studies accurately. They were wrong about everything else: the liability architecture that requires physician sign-off on every read, the expanded demand for imaging created by AI-assisted disease detection in primary care, the new subspecialty work of reviewing and auditing AI outputs, and the simple institutional fact that hospitals do not retrain their staffing models at the pace that software improves.
The radiologist story is not a story about AI failing. It’s a story about how task competence and job displacement are related much more loosely than the forecasting literature assumed.
Now consider two occupations that surprised in the opposite direction.
Compliance officers at mid-size financial institutions were considered safe by most 2022 models. The work requires regulatory judgment, relationship management with examiners, institutional memory, and the kind of contextual interpretation that doesn’t compress into a rule set. What the models underweighted was the way AI changes the economics of compliance functions from the top down. Large institutions can now run AI-driven regulatory monitoring that previously required teams of 15-20 junior compliance analysts. They can generate BSA/AML reports, flag suspicious transaction patterns, and draft regulatory response letters with minimal human input. This doesn’t eliminate the senior compliance officer role. It eliminates the organizational justification for 12 of the 15 people on her team, and — this is the part that takes two years to show up in employment data — it eliminates the career pipeline that would have produced the next generation of senior compliance officers.
Technical writers experienced a sharper and faster shift than almost anyone in the field expected. The common view in 2022 was that technical documentation required enough domain knowledge, audience modeling, and structural judgment to resist AI substitution. By 2025, that view was empirically untenable. The combination of large language models with decent context windows and company-specific documentation corpora produced outputs that, for standard API documentation, user guides, and product release notes, were good enough to eliminate the first-draft human role. What remained — and this is the pattern that appears across multiple knowledge work occupations — was the editorial, QA, and strategic layer. Not “write the docs,” but “decide what docs are needed, review AI drafts for technical accuracy, and maintain the style and knowledge standards that make AI output usable.”
The headcount implications are severe. A technical writing team of eight is now, in many software companies, a team of two plus AI tooling. The two remaining humans do more cognitively demanding work. They are paid more. The six people who had the other jobs are not.
More surprising than any individual occupation is the pattern of what I’d call liability-anchored stability. Occupations where errors produce legally or professionally consequential outcomes — physicians, licensed engineers, certified public accountants — have shown substantially more employment stability than their task-exposure profiles would predict. This is not because AI can’t perform their core cognitive tasks with reasonable accuracy (in many narrow domains it can). It’s because the institutional and legal architecture of these professions requires a credentialed human to take professional responsibility for the output. The AI does the work; the human signs the document; the human carries the liability; the human keeps the job.
This is a genuinely precarious stability. It depends on regulations and professional licensing structures that were written for a world without AI, and those structures are now under active political and regulatory pressure. The debate within the American Bar Association about AI-supervised legal practice — whether an AI system supervised by a licensed attorney constitutes unauthorized practice of law — is not an abstract professional turf dispute. It’s a negotiation about whether the liability anchor holds.
An unexpected winner: urban logistics coordinators. Not the warehouse workers (who have experienced their own disruption story with robotics, a separate and earlier wave), but the humans who coordinate last-mile delivery in dense urban environments — the dispatch managers, route optimization specialists, and exception handlers who deal with the cases that automated systems can’t handle. This category has seen wage growth of roughly 8% year-over-year in 2026. The explanation is counterintuitive: the more AI is deployed in logistics, the more valuable the human who can handle the exceptions becomes, because the system increasingly routes all exceptions to that human while handling all routine cases itself. The human’s work becomes purely non-routine, which means it becomes harder to automate, which means the human becomes scarcer and more valuable.
This is a general pattern that the forecasting models largely missed because they modeled occupation-level displacement rather than within-occupation task dynamics. When AI absorbs the routine tasks in a role, the human worker is left with only the non-routine tasks — which sounds like a reduction in scope, but is often actually an increase in the cognitive intensity and thus the value of what remains.
The occupation I would watch most carefully for 2027 is mid-career manager in knowledge work organizations. This category has been, so far, relatively insulated from direct AI displacement. But the structural logic for its disruption is building. Managers in knowledge work organizations exist primarily to coordinate information flows, allocate resources, monitor performance, and make prioritization decisions. AI systems are becoming competent at all four of these functions in the specific context of software development, content creation, and analytical services. The remaining function — motivating and developing people — is real and hard to automate, but it’s not obviously sufficient to sustain current management headcount ratios if the coordination and monitoring functions are absorbed by software.
The history of technological disruption in white-collar work is full of roles that seemed stable until they weren’t — not because the technology changed overnight, but because the organizational logic for maintaining those roles quietly dissolved over several years before anyone noticed. Middle management in manufacturing went through exactly this between 1980 and 2000, driven by enterprise software and flat organizational ideologies. The middle knowledge worker may be on a similar trajectory.
What the occupational surprise list tells us, in aggregate, is that the employment effects of AI are mediated by three factors that the simple task-exposure models ignored: institutional architecture, liability structures, and the value premium placed on exception-handling. The occupations that change most are the ones where all three of these mediating factors are weak. The occupations that hold are the ones where at least one of them is strong.
The forecasters who got it wrong were not naive. They were applying technically rigorous models to a problem where the critical variables weren’t captured in the data they were modeling. That’s a different kind of wrong than being uninformed — it’s the kind of wrong that takes longer to correct.
One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.