AI Can Change the Search Before It Changes the Job
A Swiss apprenticeship study found real search behavior shifted after ChatGPT. The harder question is what happens once expectations and labor-market reality start moving together.Photo by Jess Morgan on UnsplashWhen ChatGPT launched publicly in late 2022, another signal moved with it in…
A Swiss apprenticeship study found real search behavior shifted after ChatGPT. The harder question is what happens once expectations and labor-market reality start moving together.Photo by Jess Morgan on UnsplashWhen ChatGPT launched publicly in late 2022, another signal moved with it in Switzerland: where young people looked for apprenticeships.Daniel Goller, Christian Gschwendt, and Stefan Wolter analyzed search activity for Swiss apprenticeship vacancies. Their final Labour Economics study estimates that search intensity fell by about 8% after ChatGPT’s launch, with larger declines in occupations heavy in cognitive tasks and language skills.That is real behavior. It is also easy to turn into a bigger claim than the evidence permits.The researchers did not find a matching fall in signed apprenticeship contracts. They did not show that Switzerland suddenly had fewer of those jobs. And they did not record each searcher’s belief about AI and then watch that belief cause a decision.So this is not evidence that AI had already remade the Swiss labor market. It is narrower. People started looking differently.A search is not a job lossMost arguments about AI and work begin at the end of the chain.How many jobs disappeared? Which occupations shrank? Who was laid off? Did wages fall? Did employers stop hiring?Those questions matter. But a labor-market response can appear earlier.Before someone signs a contract, they search. Before they enter a field, they compare paths. Before they retrain, they decide whether the current path still looks worth the investment.The Swiss study captures one of those earlier decisions.Its design is unusually useful for that question. ChatGPT’s public launch was abrupt and difficult to anticipate. The authors exploit that timing with a difference-in-discontinuity approach, comparing changes in search behavior across occupations.The strongest declines appear in occupations with more cognitive work and greater language demands. The pattern also aligns better with later measures of large-language-model exposure than with older assumptions about which occupations were most automatable.That makes the signal difficult to wave away as a normal seasonal dip. But it leaves one important gap. The study observes the technology shock and the search behavior. It does not directly observe the expectation between them.If expectations are part of the mechanism, other evidence has to carry that step.The future people imagine can affect the choice in front of themA 2026 study of more than 7,000 Swiss adolescents by Christian Gschwendt, Martina Viarengo, and Thea Zöllner gets closer. The researchers ran a discrete-choice experiment with teenagers around age 15, when many are approaching the school-to-work transition. They evaluated jobs that differed in GenAI collaboration, automation risk, and opportunities for continuing education.The result was not a simple rejection of AI.Adolescents were generally wary of working with GenAI, but automation risk and continuing-education opportunities also changed how occupations were valued.Those are experimental choices, not completed career moves. But they show that changing the technological conditions attached to a job can change how that job is evaluated.A German randomized information experiment reaches the expectation question from another direction. Philipp Lergetporer, Katharina Wedel, and Katharina Werner gave workers information about the automatability of their occupations. The information increased concern about their professional future and, among workers in highly automatable occupations, raised willingness to pursue further training by about five percentage points.Again, willingness is not training completed. That distinction is the point. The studies measure different stages of a decision, and they should stay different.Actual training evidence is suggestive but weaker. A peer-reviewed Dutch workforce study found that workers who perceived high automation risk had a seven-percentage-point higher probability of reporting training participation. That is completed behavior, but the relationship is observational, concerns automation broadly rather than GenAI specifically, and cannot establish the same causal chain as a randomized experiment.Together, these studies support a bounded mechanism: information and perceptions about technological risk, opportunity, and the ability to adapt can change how people evaluate what to do next.This is not a flight-from-AI storyThe easiest version of this article would say young people see AI coming and run from exposed careers.The evidence refuses to stay that simple.The Swiss career-choice experiment already shows why. GenAI collaboration, automation risk, and continuing education pull choices in different directions.U.S. early-career behavior is just as mixed. Handshake’s data on Class of 2026 computer-science majors show that recent students have been sending a larger share of applications to IT, computer systems, cybersecurity, finance, marketing, and project-management roles as software engineering has become a less dominant destination.At the same time, 45% of those computer-science majors were highlighting AI skills on their résumés.And employer demand was moving too. By March 2026, more than 10% of active internships on Handshake mentioned AI, while 4.2% of full-time postings did, nearly double the share a year earlier.That is not a generation simply fleeing AI. Some people may avoid paths that look more exposed. Others may move toward AI-heavy opportunities. Others may add AI skills as a hedge or an advantage. Some may not change course at all.There is no single psychological variable measured across these studies. The safer recognition is messier: people are making present choices while weighing different signals about future risk, opportunity, and adaptability. That is a stronger story than fear.Then reality starts contaminating the experimentThere is a problem with studying expectations inside a live labor market: eventually the conditions people are anticipating may begin to change, while unrelated economic forces are changing too.By then, behavior becomes much harder to interpret.National Student Clearinghouse data show that Computer and Information Sciences enrollment fell across major institution types in spring 2026 even as undergraduate enrollment overall grew. Four-year undergraduate enrollment in the field fell 8.4%; declines reached 9.3% at primarily associate-degree-granting baccalaureate institutions and 11.2% at two-year institutions.That is observed enrollment behavior. It is not evidence that AI caused it.The technology job market was already weakening before ChatGPT appeared. In September 2022, technology-sector full-time postings on Handshake were down 40% year over year, after an earlier hiring surge and amid freezes, layoffs, slowing venture funding, and broader economic uncertainty.Later, software-engineering demand weakened further while employer demand for AI skills began to rise.A computer-science student applying outside software engineering in 2025 or 2026 might therefore be responding to expectations about AI, fewer software openings, stronger demand elsewhere, or some combination of all three.A Gallup-Lumina survey adds another piece without solving the causal problem. In an October 2025 opt-in web survey of 3,801 U.S. associate and bachelor’s students, 16% reported that they had already changed their major or field because of AI’s potential impact.That is more than an attitude: respondents reported a completed change. But both the change and its motive are self-reported. The survey cannot tell us what national enrollment would have looked like without AI, or cleanly separate expectations from labor-market signals students were already seeing. The respondents also came from an opt-in online panel.By 2026, that separation gets harder still. An August 2026 revision from the Stanford Digital Economy Lab using ADP payroll data through June found a 19% relative employment gap for workers ages 22 to 25 in highly AI-exposed occupations compared with where their employment would have been had it kept pace with less-exposed peers.The researchers also report no widespread economy-wide displacement. Their results are best treated as emerging descriptive evidence, not as proof that AI destroyed 19% of young workers’ jobs.That is enough for this story. It shows why later career behavior cannot be treated as a clean experiment in expectation. The labor market itself has started sending signals back.The labor market does not wait for the final answerNone of this establishes that AI expectations have already changed aggregate labor supply, transformed occupational composition, or created a self-fulfilling shortage in particular fields. The evidence supports a smaller claim.A measurable change followed a major GenAI shock in real occupational search. Experiments show that information about technological risk, AI collaboration, and retraining opportunities can change preferences and intentions. Later evidence shows people moving in several directions at once, just as employer demand and early-career hiring conditions make the original expectation harder to isolate.That is enough to change how we think about the labor market. It is not only a record of what technology has already done. It is also built from decisions made under uncertainty about what technology may do next.Those decisions begin before a contract is signed, before a worker is displaced, and before economists can see the final shape of the adjustment. Sometimes the first measurable trace is a search.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!AI Can Change the Search Before It Changes the Job was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI