What the 2025 Swiss Apprenticeship Study Adds to AI Exposure Evidence

What did the final 2025 article report—and where is the 8% stated?

The final 2025 Labour Economics article, “This time it’s different – Generative artificial intelligence and occupational choice,” reports that young people’s searches for Swiss apprenticeship vacancies declined after ChatGPT’s release. The publisher’s article highlights state an average decline of 8%. That figure belongs to the highlights: the accessible abstract describes a strong short- and medium-term decline but does not give a percentage. The abstract also says the search response was larger for occupations with more cognitive tasks and higher language requirements, while signed apprenticeship contracts did not decline and applicant quality fell, particularly in commercial apprenticeships. The wording matters because “average decline” characterizes the study’s search-intensity outcome, not an identical response by every young person or occupation.

This corrects a consequential version distinction. The 2023 IZA Discussion Paper, “This Time It’s Different” – Generative Artificial Intelligence and Occupational Choice, is an earlier working-paper version, not a substitute for the final journal article. A figure found in that PDF must be attributed to the 2023 version unless the final publisher record independently confirms it. Here, the 8% is confirmed by the 2025 publisher highlights, but it should not be described as a number stated in the accessible abstract. The abstract’s omission is a reporting difference within the final publication, not evidence that the percentage appears only in the earlier paper.

The population and outcome also set a firm boundary on what the result means: it concerns young Swiss apprenticeship searchers and their search behavior. It is not an estimate of job losses, a probability that an individual will be displaced, or a finding about experienced workers already employed in occupations exposed to generative AI. The study adds evidence that anticipated AI change can coincide with a shift in early-career occupational search; its reported percentage does not convert that search response into a forecast of employment outcomes.

Sources: This time it’s different – Generative artificial intelligence and occupational choice

What did a vacancy search measure, and what did it leave out?

A vacancy search is an observable act of looking for an apprenticeship, not a direct reading of what a young person thinks or what they ultimately do. In the final 2025 article, the outcome is search behavior; the abstract discusses it separately from signed contracts. The 2023 IZA Discussion Paper describes the underlying platform evidence as vacancy-query activity. That operational detail should remain attached to the working-paper description rather than silently being presented as a newly verified method in the final article’s accessible abstract.

This measure sits near the beginning of a longer sequence: someone can search, decide whether to apply, submit an application, receive or decline an offer, sign a training contract, begin and complete the apprenticeship, and later enter employment. These are distinct events. A search signal indicates that a vacancy is being sought; it does not establish that an application followed, that an employer made an offer, or that a contract was signed. A signed contract is further along the pathway, while completion and later employment are later outcomes that the abstract does not report as consequences of the search measure.

The sequence matters because counts can change at one stage while later stages remain stable or move differently. A young person might stop searching after accepting an offer, switch to a different training route, pause a search, or alter how they use the platform. These are possible explanations for an individual search ending, not reasons identified by the aggregate study. Likewise, fewer observed queries could reflect less searching through that channel without showing that the person abandoned apprenticeship, left education, or became unemployed. The recorded behavior narrows the claim to search activity; it does not disclose the unobserved next step for each searcher.

Search behavior is more concrete than asking respondents to predict what they might do, because the measure concerns activity rather than a hypothetical answer. Yet aggregate query counts still cannot reveal individual motives or link each searcher to an application, contract, completion, or job. Without person-level trajectories and reported reasons, the measure cannot tell whether a change reflects fear of automation, a different occupational preference, a completed search, or another circumstance. Those distinctions are essential when interpreting a change in searching: it describes one stage in a choice process, not the choice process in full. Nor does a query count reveal whether the same person searched repeatedly, how many vacancies they considered, or whether platform activity captured searching elsewhere. These limits concern the interpretation of the measure, not a reason to dismiss it: the observed behavior is informative about active search on the recorded channel, while remaining silent about unrecorded decisions and outcomes.

Sources: This time it’s different – Generative artificial intelligence and occupational choice; “This Time It’s Different” – Generative Artificial Intelligence and Occupational Choice (IZA Discussion Paper 16638)

How does the 2023 launch-based design support—and bound—the causal claim?

The 2023 IZA Discussion Paper, “This Time It’s Different” – Generative Artificial Intelligence and Occupational Choice, treats ChatGPT’s public launch on November 30, 2022 as a sudden, salient event that could alter how young people viewed future occupations. Its design compares apprenticeship-vacancy query activity before and after that date, asking whether search changed when generative AI became widely visible. The paper therefore gives a launch-related interpretation to a change in observed behavior: the timing is consistent with a response to the release, rather than a direct observation of any searcher’s thought process.

The IZA paper describes platform query observations aggregated across occupation, canton and time. Those comparisons let the authors examine whether changes differed across occupational groups and locations around the event, rather than relying only on one overall before-and-after count. This is useful because a broad time comparison can be sensitive to seasonality or other common changes; differences across groups and places can provide additional context for whether the pattern tracks occupational characteristics. These are features of the 2023 working-paper design as described in that version. The accessible 2025 journal abstract does not supply these detailed data or specification descriptions, so they should not be silently carried over as verified final-article methods.

A before-and-after design becomes more persuasive when the outcome changes near the event and when the result remains under reasonable alternative comparisons. The 2023 IZA Discussion Paper presents its evidence as a causal effect of the launch on search behavior, and its reported robustness analyses strengthen that interpretation. But the causal object remains limited: it concerns aggregate search activity in the platform data around a public release. The underlying platform data are confidential, which means readers cannot independently reconstruct the query panel or rerun the paper’s specifications from the PDF alone. Robustness checks can show that a finding is not dependent on a particular modeling choice; they cannot make confidential observations directly inspectable.

The distinction is important for causal language. A comparison around a dated event can support a claim that the launch influenced a population-level outcome when the design’s assumptions hold; it does not make the launch equivalent to a randomized assignment of beliefs or exposure. The unit summarized in the data is platform activity grouped across occupations, places and time, not a traced sequence of one person’s expectations and decisions. The design is thus informative about timing and distribution of the response while leaving its individual pathway unresolved.

The launch also bundled more than a technical capability demonstration. Public attention, news coverage and discussion about future work could have changed at the same time as access to ChatGPT. Even if the launch timing is the relevant shock in the authors’ design, the study does not isolate whether an individual searched less because of fear, changed expectations, conversations with family or teachers, or some other response to the event. Nor does a search change establish that the person’s occupation was technically automatable, that employers adopted AI, or that jobs later disappeared. The defensible reading is causal influence on the measured search outcome under the 2023 paper’s design, with the psychological channel and downstream labor-market consequences left open.

Sources: “This Time It’s Different” – Generative Artificial Intelligence and Occupational Choice (IZA Discussion Paper 16638); This time it’s different – Generative artificial intelligence and occupational choice

Which occupational task dimensions distinguished the search response?

The final 2025 Labour Economics article, “This time it’s different – Generative artificial intelligence and occupational choice,” reports larger search declines for occupations with more cognitive tasks and higher language requirements. That pattern distinguishes occupational groups by the kinds of work associated with them; it does not show that an AI system performed those tasks well, that employers deployed it, or that any particular job was eliminated. The result is evidence about how search behavior varied across occupational categories after the launch.

These dimensions describe different features of work. Cognitive versus manual refers broadly to whether activities center on mental processing or physical action. Language demand concerns how much work depends on understanding or producing language. Routine versus non-routine, by contrast, describes how structured and repeatable tasks are, rather than whether they are mental or physical. A task can be cognitive and routine, cognitive and non-routine, or language-intensive without fitting neatly into a single routine category. Treating these labels as interchangeable would obscure what each comparison can test.

The 2023 IZA Discussion Paper, “This Time It’s Different” – Generative Artificial Intelligence and Occupational Choice (IZA Discussion Paper 16638), reports occupation-group comparisons that include cognitive tasks, language demands and routine-task share. In that earlier working-paper version, the routine versus non-routine split did not show a meaningful difference in the search response, while the cognitive and language patterns were more pronounced. The final 2025 abstract retains the cognitive- and language-demand finding; it is the IZA version that supplies the specific routine-category comparison. Keeping that version boundary matters because a working-paper subgroup result should not be presented as if the journal abstract independently reported it.

This is not a contradiction: routine structure and language or cognitive content are separate axes, so a null split on one need not erase a pattern on the others. It also does not mean routine tasks are immune or that non-routine work is protected. The comparison tells us which occupational classifications showed different search responses in this study, not which task types a system can complete reliably under workplace conditions.

Together, the dimensions suggest that young searchers’ responses aligned more clearly with occupations associated with cognitive and language work than with the older routine/non-routine distinction. One possible interpretation is that public expectations about generative AI were especially salient for work perceived as involving text, information or analysis. That is an interpretation of the pattern, not a measured explanation: the study does not establish what searchers knew, which capabilities they expected, or whether their expectations matched actual task performance. Occupational averages also conceal variation among individual roles and workers. A person in a language-intensive occupation may spend substantial time on coordination, client trust, physical procedures or accountable decisions that a broad category cannot represent. The finding therefore describes choices across occupation groups, not a capability ranking of tasks or a reliable forecast of which individual roles will change.

Sources: This time it’s different – Generative artificial intelligence and occupational choice; “This Time It’s Different” – Generative Artificial Intelligence and Occupational Choice (IZA Discussion Paper 16638)

What does a task-based exposure index measure beside searches?

The International Labour Organization’s 2025 report, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” asks a different question from the Swiss apprenticeship study. It estimates how far current generative AI could apply to work tasks; it does not record whether young people search for apprenticeships, whether a workplace uses AI, or whether a job disappears. The Swiss article measures a change in observed search behavior. The ILO index models potential task applicability. Reading the two side by side can clarify the difference between a response to perceived change and an estimate of technical overlap, but it cannot establish that the searchers acted on the index or accurately foresaw AI capability.

The index begins with task-level judgments rather than assigning one exposure label to an entire job. In its methods summary, the ILO reports worker assessments for 2,861 tasks drawn from a sample of 29,753 tasks in Poland’s occupational classification. The assessments came from 1,640 workers and were compared with international expert input; AI-assisted scoring then helped extend task estimates across occupations. Each component contributes a different kind of information: workers assess task applicability from work knowledge, experts provide comparative judgments, and model-assisted scoring broadens coverage beyond the directly assessed tasks. The resulting occupational estimates are therefore a synthesis and extension of judgments, not a census of AI use in workplaces.

The ILO’s companion explainer, “How might generative AI impact different occupations?”, describes four exposure gradients that use both an occupation’s average task score and variation among its tasks. That second dimension matters: two occupations can have similar average exposure while one contains many tasks with comparable scores and another combines highly exposed tasks with less exposed ones. The gradient can therefore distinguish broad occupational profiles that a single average or job title would flatten. It still summarizes task estimates at occupational level; it does not show how a particular worker divides their time, how tasks are organized in one employer’s workflow, or which tasks carry accountability, trust or verification requirements.

The ILO’s global employment shares should be read in that same modeled frame. They describe how many jobs fall within estimated exposure gradients under the index, not how many workers already use generative AI or will lose work. A global share can show the scale of potential applicability across a broad labor market, making it useful as a comparison to a study of one country’s apprenticeship searches. It cannot be used as a forecast for Swiss apprenticeships, a local employer, or an individual career decision. Differences in occupational mix, task organization, infrastructure, cost and rules intervene between a capability estimate and actual implementation.

This task-based index is more informative about variation inside occupational categories than a single broad job label, but it remains a model-assisted potential assessment. Its task scores may help identify where to investigate work in detail; they cannot explain why Swiss youth changed their searches or predict what a specific employer will automate, augment or retain. The comparison is useful precisely because the measures answer different questions: one observes behavior around occupational choice, while the other estimates potential overlap between AI capability and tasks. Neither supplies the missing link from perceived or modeled exposure to an employer’s decision or a worker’s eventual outcome.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; How might generative AI impact different occupations?

Why can an exposure signal not explain the search response by itself?

The measures form a chronology, not a single prediction test. Older expert assessments of automation risk classified occupations using capabilities and task assumptions available before the current generative-AI wave. The 2025 Swiss apprenticeship article reports later search behavior after ChatGPT’s release, while the ILO’s 2025 index estimates potential generative-AI applicability task by task. The study says the search pattern contrasted with prior expert automation assessments and aligned with newer exposure assessments. That correspondence is informative about which occupational patterns resemble each other across measurement approaches; it does not mean young searchers consulted later indexes, which were unavailable when the choices occurred.

The distinction between an older automation-risk assessment and a newer exposure index is substantive. A broad automation-risk estimate asks whether jobs may be automatable under a particular set of assumptions. A task exposure measure asks whether AI capabilities could apply to some work activities. The ILO’s Generative AI and Jobs index makes the latter more granular by modeling tasks and variation within occupations. The Swiss study, by contrast, observes a behavioral outcome: searches for apprenticeship vacancies shifted across occupational groups. Similarities between occupational rankings or group patterns do not show that people understood the underlying capabilities, nor that the exposure classifications caused the search change. They are a suggestive correspondence between unlike measures.

The OECD’s 2023 chapter, “Artificial intelligence and jobs: No signs of slowing labour demand (yet),” explains why task overlap does not settle what happens to work. AI may substitute for some tasks, reducing the labor required for them, while complementing workers on others or raising productivity in ways that change demand. New tasks can also emerge as production and services reorganize. These channels can operate together, and their net effect depends on adoption, task redesign, demand and workplace conditions. The chapter’s framework largely predates recent generative-AI use, so it clarifies the mechanisms that make exposure ambiguous; it does not establish which channel prevailed in the Swiss apprenticeship setting.

This mechanism boundary prevents a category error. A task can be technically exposed without being automated in practice; a tool may assist a worker while leaving review, context, customer interaction or accountability with that worker. Even when a task is substituted, the effect on a role depends on whether remaining tasks expand, shrink or change and whether demand for the output responds. Accordingly, an exposure estimate cannot convert the apprenticeship search result into a forecast of contracts, employment or wages. And a search response cannot prove that an occupation was technically exposed: it records behavior, which may reflect expectations or other influences rather than verified capability.

The study’s reported alignment with newer exposure assessments is therefore best treated as a clue about how perceived occupational change and modeled capability overlap, not evidence that the young searchers anticipated AI accurately. It does not establish a causal path from the index to search decisions, and neither measure predicts the employment result for a particular occupation or worker. A stronger conclusion would require evidence connecting task capability to actual employer adoption, changed workflows and labor demand over time. Until then, the comparison helps organize questions about change while leaving the causal and employment links unresolved.

Sources: This time it’s different – Generative artificial intelligence and occupational choice; Artificial intelligence and jobs: No signs of slowing labour demand (yet)

Why do contracts and applicant quality complicate the story?

The final 2025 Labour Economics article, “This time it’s different – Generative artificial intelligence and occupational choice,” reports a decline in apprenticeship-search intensity alongside no decline in signed contracts. These measures sit at different points in a selection process. Search queries capture activity among people looking; contracts count matches completed between applicants and firms. A drop in one therefore need not produce an immediate drop in the other. The article also reports declining applicant quality, especially for commercial apprenticeships, adding a dimension that a contract total alone would miss: the composition of the pool can change even when the number of agreements does not.

Several mechanisms could produce that combination, but the reported pattern does not identify which one operated. Employers might select among fewer applicants and still fill a similar number of places; this is a plausible selection mechanism, not the authors’ established explanation. Applicants might redirect their search among occupations or channels, or vacancies and applicant numbers might adjust at different speeds. These possibilities have different implications, and the aggregate mismatch cannot distinguish them. It would be unwarranted to turn any one into a factual account of employer behavior without data linking the relevant firms, candidates and decisions.

The quality result deserves similarly careful interpretation. The article’s wording points to a change in the assessed quality of applicants, particularly in the commercial-apprenticeship group; it does not, by itself, reveal the assessment’s practical consequences for each employer or candidate. If a smaller or differently composed pool changes who enters training, that could matter for the pipeline. But the study’s reported result does not establish that employers later faced shortages, that training completion fell, or that graduates had worse employment outcomes. Pipeline composition is a reason to ask what happened next, not proof that a downstream problem occurred.

To establish a longer chain, evidence would need to connect searchers to applications, selection and signed contracts, then follow apprentices through training completion and into employment. Applications would show whether reduced searching translated into fewer candidates at the next stage; completion data would show whether intake changes persisted through training; later employment would show whether credentials led to work. The abstracted outcomes do not supply those person-level links. Stable contract totals thus complicate a simple story of immediate displacement, while leaving open whether composition or later outcomes changed. They cannot demonstrate that future adjustment was absent, just as lower applicant quality cannot establish subsequent harm.

Sources: This time it’s different – Generative artificial intelligence and occupational choice

Why does the Swiss apprenticeship setting limit transfer to current workers?

A Swiss apprenticeship is a structured route into an occupation, not simply a short course chosen by someone already established in a job. The Swiss career-information page “Doing an apprenticeship in Switzerland” describes training that combines work at a host company with vocational schooling and leads to a recognized vocational qualification. Programs commonly last two, three or four years, with pathways onward to employment or further education. The 2025 Labour Economics article, “This time it’s different – Generative artificial intelligence and occupational choice,” studies search behavior in this early occupational-choice setting. The decision is about which occupation and training route to enter while a person is still building their work history.

An incumbent worker faces a different decision set. Their choice is shaped by skills and expertise already accumulated, current income, existing qualifications and the jobs available within a realistic travel or relocation range. A move may also need to fit health, caregiving and family obligations. These constraints are relevant to applying the finding to experienced workers, but the apprenticeship study did not measure them or estimate how they affect career transitions. Treating them as part of transfer is an editorial inference about the difference between entering a pathway and adjusting an established working life, not a result attributed to the Swiss researchers.

What can reasonably transfer is narrower: expectations about technology may matter to occupational choices before later outcomes such as signed contracts visibly change. That makes early search behavior worth studying, but it does not tell a current worker to abandon a field, repeat vocational training or retrain for an entirely new occupation. For an experienced person, the relevant alternatives may include changing a task mix, learning a tool within the current specialty, moving to an adjacent role or making a larger transition. Which route is practical depends on the worker’s own prerequisites and constraints; the study does not compare these paths or identify a generally superior one.

The distinction also changes what evidence would help. A prospective apprentice choosing a multi-year entry route needs information about training and the occupations open to them. A current worker deciding whether to adapt needs evidence about the actual tasks in their role, how their employer is changing the workflow, and what nearby roles require. The Swiss study can inform a discussion of how perceived AI change enters early occupational choice; it cannot estimate an experienced worker’s personal exposure, the value of their existing qualifications, or the costs of changing jobs. Its setting supports attention to expectations while leaving incumbent decisions to a separate, task- and circumstance-specific assessment.

Sources: Doing an apprenticeship in Switzerland; This time it’s different – Generative artificial intelligence and occupational choice

What additional workplace conditions matter beyond technical exposure?

Distinguish potential capability from change in your role by following one recurring task from input to accountable output. The OECD’s 2025 report, “Emerging divides in the transition to artificial intelligence,” says occupational exposure alone cannot predict employment or income outcomes; liability, task criticality, complementarity and skills also matter. The OECD’s 2023 chapter, “Artificial intelligence and jobs: No signs of slowing labour demand (yet),” explains why: AI can substitute for some tasks, complement workers on others, raise productivity, or contribute to new tasks. Those are possible channels, not a forecast for a particular workplace. The practical question is what your employer permits, adopts and changes, and what happens to the surrounding work.

Start with a task you perform repeatedly, described specifically enough to observe: for example, turning a service request into a customer response, not “communications.” Note which inputs are digital and repeatable, which parts involve drafting or language, and where context, trust, physical presence or accountable judgment enter. These features help identify where a tool might assist or where a person must remain responsible. They do not by themselves establish that a tool performs the task reliably or that an employer will use it.

Next, look for local adoption evidence. Is an approved tool available for this workflow? Are workers actually using it, and under what rules for confidential information? If an output needs review, how much checking does it require, who corrects errors, and who owns the final decision? Has the workflow changed in a way you can verify, such as a new review step, a different handoff, or a changed responsibility? These observations separate a capability demonstration from an implemented process. A public announcement or a successful one-off trial may be worth investigating, but neither alone establishes how routine work or staffing will change.

Consider an illustrative case: a support specialist drafts replies to routine requests. A language model may produce a first draft, while the specialist still checks account context, handles unusual cases, follows policy and takes responsibility for what is sent. If the organization has not approved a tool or altered this process, capability alone has not changed that role. If it has, the specialist can observe which steps moved, what review remains, and whether expectations for response volume or judgment changed. This example illustrates questions to ask; it asserts no measured result about support jobs.

Then compare responses in proportion to what you observed. An in-role upgrade might mean learning the approved tool and strengthening evaluation or workflow skills. An adjacent move might use the same domain knowledge in a role with different task responsibilities. A larger transition may require new credentials or substantial retraining. Compare each against prerequisites, likely salary floor, available roles in your location, training cost and time, and effects on health or family responsibilities. Also check labor-demand signals for the relevant local roles; a task’s technical exposure says little about whether nearby work is hiring. A bounded workflow review can inform the next question, but it cannot predict future adoption, staffing plans or the market, and employer evidence may be incomplete.

Sources: Emerging divides in the transition to artificial intelligence; Artificial intelligence and jobs: No signs of slowing labour demand (yet)

What is the next proportionate step?

Before paying for retraining or making a larger move, choose one recurring task and record what you do, what a tool could plausibly assist, and what still requires your context or sign-off. Then verify a local signal: an approved-use policy, changed handoff, new review requirement, or role posting that names a changed skill. If signals are absent or unclear, gather better information before treating exposure as a reason to leave. Reassess if credible evidence shows workflow, staffing, demand or role requirements changing. Keep a factual note of the task, the policy or process source checked, and what remains uncertain. A manager’s description, team procedure and vacancy notice answer different questions; read each for what it directly shows, not as proof of a market-wide trend. If the workflow is stable but you want to improve your position, a short, low-cost learning step tied to that task may be easier to evaluate than an immediate credential or career pivot.

The free task checker at /ai-job-risk-checker organizes task-level change-pressure signals; its result is not a validated probability of displacement or an employer forecast. The paid roadmap at /career-roadmap compares scenarios against your constraints but does not guarantee employment or income. Both are optional aids, not substitutes for employer policy, local labor information or your full circumstances. Choose one only if it answers the next question you have.

Questions readers ask