For nearly a decade, Americans have been told that artificial intelligence is coming for their jobs. The warnings arrived with impressive credentials and alarming percentages. But predictions are easy to make and hard to verify, and most of them were written before the technology they feared even existed.
Now there is enough distance to check the forecasts against reality. According to Wiingy, a tutoring marketplace, the major predictions about which skills AI would erase can finally be tested against three years of real behavior, real wages, and real employment data. The findings are clarifying, and in some cases they overturn what we thought we knew.
The forecasts everyone quoted
Two studies shaped the conversation. The first, the Oxford Future of Employment report published in 2017 by researchers Carl Frey and Michael Osborne, ran 702 American occupations through a machine learning model and assigned each one a probability of automation.
The second, the World Economic Forum’s Future of Jobs report from 2025, surveyed more than 1,000 employers across 55 economies and asked which roles they expected to grow and which to shrink by 2030.
Despite the years between them, the two studies largely agreed. Physical, hands-on, interpersonal, and creative-performance work would be difficult for AI to replace. Routine text processing, data handling, and repetitive digital tasks would be easy.
The question Wiingy set out to answer was simple: were they right?
How the study tested the claims
Rather than rely on opinion, the Wiingy research team examined 29 skills across four independent datasets. ChatGPT’s launch on November 30, 2022 served as the dividing line between “before” and “after.”
The four data layers were:
- Google Trends – monthly search interest for each skill from 2020 to 2026, comparing demand before and after ChatGPT.
- Google Keyword Planner – the most current signal, measuring how many people searched to learn each skill year over year.
- Bureau of Labor Statistics wages – government median pay data, adjusted for the 8.3% cumulative inflation over the period.
- Bureau of Labor Statistics employment – the number of jobs created or destroyed in each occupation.
The team built a simple measure called the Temporal Resilience Score, dividing search demand after ChatGPT by demand before it. A score above 1.2 meant a skill was growing, while a score below 0.9 meant it was fading. The approach favored observed behavior over speculation, which is exactly what the earlier forecasts lacked.
The jobs AI is genuinely replacing
The clearest casualties confirmed the warnings. According to Wiingy, the skills predicted to be vulnerable are not just slipping in theory; they are losing search interest, wages, and headcount at the same time.
- Transcription suffered the sharpest collapse in the entire study, with course searches falling 63% in a single year. Free, near-perfect tools made human transcription training almost pointless.
- Copywriting showed the economic signature of displacement: real wages down 9.5% while more than 6,000 jobs disappeared.
- Broadcasting lost over a third of its workforce in two years, alongside the steepest real wage decline of any occupation measured.
- Proofreading has nearly vanished as a standalone career, with roughly 5,160 people left in the role nationally.
When demand, pay, and employment all fall together, the cause is rarely a coincidence. These are the fingerprints of automation, and the data shows them plainly.
The skills that refused to fade
The more encouraging story is what held steady or grew. The trades, in particular, are thriving in the AI era.
- Electrical work produced the strongest single demand signal in the study. Searches for electrician apprenticeships jumped 49% year over year, and the occupation added more than 52,000 jobs.
- Architecture posted the highest real wage growth of any skill examined, up 8.4% after inflation, as AI design tools made architects more valuable rather than less.
- Nursing remained one of the most resilient careers across every measure, with rising wages, growing employment, and one of the lowest automation risk scores on record.
- Fitness and physical instruction expanded rapidly, adding tens of thousands of positions as embodied, in-person skills proved impossible to automate.
The pattern is consistent. Work that requires a human body, human judgment, or human presence is not only surviving but, in several cases, commanding more demand precisely because machines cannot do it.
The surprises hiding in the data
Not every prediction landed cleanly, and the exceptions are some of the most interesting findings.
Plumbing was likely misclassified. Oxford assigned it a 35% automation risk, yet search demand nearly doubled. The data suggests it belongs firmly among the AI-resistant trades.
Web development revealed a “barbell” effect. Wages for web developers rose well above inflation even as the total number of jobs shrank by 11%. As AI absorbs routine work, the field is splitting: fewer people remain, but those who do handle more complex, higher-value tasks. The same shape may be forming in other digital fields.
People are still learning “doomed” skills, for rational reasons. Some workers study AI-threatened crafts not to compete with machines but to direct them. A video editor today is increasingly there to supervise and shape AI output, not to fight it. Others are visibly retraining away from collapsing fields, which is why categories like transcription showed a brief surge in interest before falling off a cliff.
What it means for American workers and leaders
The headline conclusion is that the experts were broadly correct, but the value of this study is not in vindicating a forecast. It is in replacing fear with evidence.
For workers weighing where to invest their time, the signal is unusually clear. Skills rooted in the physical world, in care, in craft, and in human relationships are gaining ground. Skills built on routine text and data processing are under real and measurable pressure.
For business and education leaders, the implications run deeper. Workforce planning, training budgets, and hiring strategies can now be grounded in three years of behavior rather than a decade-old projection.
The labor market has already begun voting with its searches, its paychecks, and its job postings, and the results are legible to anyone willing to look.
Artificial intelligence is reshaping American work, but not in the indiscriminate way the early headlines implied. According to Wiingy, the divide is sharp and the direction is set: the future belongs to skills that are difficult to digitize, and the data is finally specific enough to tell us which ones those are.
