Exposure Is Not Adoption
This series examines who actually uses generative AI at work, beginning with a deceptively simple question: if a technology can help with the tasks in a job, why does that not lead everyone in similarly exposed work to adopt it? The answer matters for any worker, manager, or policymaker trying to understand who gains early access to new capabilities. It also supports a more careful judgment than either “AI will transform every exposed job” or “people who do not use it are falling behind.”
The central distinction is between exposure and adoption. Exposure describes where a technology could matter. Adoption describes whether a person is using it. Between those two points sit skills, job design, workplace authority, digital infrastructure, training, and judgments about whether the technology is safe and legitimate. The early evidence suggests that these conditions shape the conversion of technical potential into socially distributed use—and that adoption itself should not yet be confused with a settled transformation of work.
A fast diffusion with an uneven reach
Generative AI entered working life unusually quickly. The European Working Conditions Survey conducted in 2024 asked workers in 35 European countries whether they used artificial intelligence at work that simplified complex mental tasks or made recommendations, giving examples such as ChatGPT, Llama, DALL-E, Midjourney, and Jasper. About 12% of workers answered yes. That average conceals a very wide spread: reported use was below 3% in Bosnia and Herzegovina and roughly one quarter in Luxembourg. Northern and Western European countries generally appeared ahead of Southern and Eastern European countries.
The important fact is not simply that adoption was already visible less than two years after ChatGPT’s public launch. It is that the same broad technology arrived in workplaces with very different rates of uptake. A worker in a highly exposed occupation may have a strong reason to experiment, yet still lack a computer-centered workflow, an approved tool, time to learn, confidence in evaluating outputs, or permission to put generated material into production. Conversely, a worker in a less exposed occupation may use a general-purpose assistant for occasional writing, information synthesis, or planning. Technical relevance sets the terrain; it does not determine the route taken across it.
The timing also matters. These figures describe a particular early-diffusion moment, not a permanent European adoption rate. The survey captured whether a worker used such a tool at all, not how often they used it, how deeply it was integrated, or whether the use was informal and invisible to an employer. The estimate is therefore useful for comparing where adoption was taking hold, but it should not be read as a current census of AI use or as a measure of productive mastery.
There is a further measurement qualification. The survey question is broader than the phrase “generative AI”: simplifying complex mental tasks or making recommendations could include some decision-support systems. Its named examples—ChatGPT, Llama, DALL-E, Midjourney, and Jasper—anchor respondents toward generative tools, and the phrase “at work” separates the result from purely personal experimentation. That makes the item a practical indicator of workplace engagement, not a perfect technical classification. A worker’s yes also does not tell us whether they checked the output, used an employer-approved system, or built the tool into a recurring process. Those missing dimensions matter because two people can both count as adopters while receiving very different amounts of learning, leverage, and exposure to error.
Why pay attention to an uneven pattern before large effects on jobs or earnings are established? Early use can be a channel through which people learn what the technology is good at, develop ways to check it, and gain influence over emerging workplace norms. Those possible advantages are not the same as demonstrated long-run benefits. But if some workers get repeated opportunities to experiment while others face barriers or reasonable doubts, the distribution of learning itself becomes part of the workplace question.
This is why the subject cannot be reduced to a race between adopters and non-adopters. A person who has tried a tool once may not have acquired a transferable skill. A person who has not used it may be making a considered choice under constraints that the adoption statistic cannot see. At the same time, a workplace that reports high use has not necessarily redesigned its jobs or improved their quality. The relevant questions are more demanding: who can act on the opportunity, under what conditions, with what authority to check and adapt the tool, and with what consequences for the work that remains? Starting with those questions keeps the focus on workplaces rather than on an abstract technology leaderboard.
Four different things that are easy to confuse
The first confusion is to treat exposure as a forecast. The European study measures occupational susceptibility with an index built from tasks: it asks whether current large language models could plausibly reduce the time required for certain activities, directly or through software that incorporates them. The index is assigned at the level of broad occupations. Information and communications technology professionals, for example, score highly because their work contains many tasks involving language, code, analysis, or synthesis. That score says something about the potential fit between technology and task content. It does not say that every worker in the occupation has access to the same system, wants to use it, or can safely rely on its output.
The second confusion is to treat a worker’s report of use as a measure of intensity. The European survey records the extensive margin: whether the worker uses AI at work. It does not distinguish a one-off experiment from daily use embedded in a workflow. Nor does it see a backend system that affects a process without being salient to the person doing the job. A firm’s statement that it uses AI is different again. Enterprise reporting may include automated systems workers never see, while workers may be using publicly available tools without formal organisational approval. These measures can move together while still describing different realities.
The third confusion is to treat a high adoption rate as proof that work has already been redesigned. Workers who use AI report more technology-related task displacement and task creation than non-users, and occupations with more AI users also report more task restructuring. That pattern is important as a description. It is not yet proof that AI caused the change. People in jobs already being reorganised may be more likely both to adopt AI and to report new or removed tasks. An association between exposure and use cannot, by itself, settle what happened to the task bundle.
The distinction is easy to miss because the survey also reports large overall rates for technology-related change: 24% of workers said technology had removed some tasks and 34% said it had created new ones. Those figures describe technology in general, not a generative-AI treatment effect. Even the stronger comparison between adopters and non-adopters is vulnerable to selection: a worker who is already involved in a changing process may be both more open to a new tool and more likely to notice that their tasks are shifting. To learn whether AI itself is responsible, researchers need a design that separates the places where adoption is high from the occupational and national tendencies that would have produced change anyway.
The fourth confusion is to treat unequal use as a single individual preference. Adoption is a decision made inside a work system. A worker’s education and ability to judge an output matter. So do the abstract demands of the job, the worker’s control over how tasks are performed, the organisation’s training and procurement practices, the country’s digital readiness, and the perceived social risks of the technology. These factors can coexist. A person may understand generative AI well and still decide not to use it because the privacy, accuracy, labour, environmental, or mental-health implications remain unacceptable to them.
The levels should not be read as a checklist in which one missing item explains every case. Education can make experimentation easier but cannot create a usable workflow. A computer may be present while an employer restricts data entry. A worker may have discretion over the order of tasks but little say in whether a new system is introduced. And concern about harm may be a response to the technology’s actual uncertainties rather than a lack of knowledge. The point is to keep these possibilities visible until the evidence can tell us which relationships are strongest.
The real question is conversion
The most useful way to read the evidence is to ask not only where generative AI could apply, but how effectively a setting converts exposure into adoption. Across European workers, use rose from 1.5% in the least exposed occupational quintile to nearly one quarter in the most exposed. This is strong evidence that technical fit organises early uptake. Yet use was not zero in the least exposed group, and the relationship varied markedly by country. Exposure is neither irrelevant nor destiny.
The European estimates make the distinction concrete. Before additional adjustments, a one-standard-deviation increase in occupational exposure was associated with a 10.1 percentage-point increase in the probability of reported workplace use. Adding education, seniority, and computer use reduced the estimate to 6.6 points. Adding employment relations, workplace characteristics, and industry controls reduced it again to 4.9 points. The pattern says that exposed occupations contain more workers and jobs already positioned for digital adoption, but also that a substantial conditional relationship remains.
It does not say that exposure itself causes adoption. The study is observational, and the index is an occupation-level proxy. The remaining relationship may reflect unmeasured skills, employer decisions, access, job autonomy, or other conditions that travel with occupational exposure. The point of the gradient is therefore diagnostic. It tells us where to look for the missing mechanisms.
Those mechanisms operate at several connected levels. Within an occupation, workers differ in formal education, the kind of judgement their tasks require, and the freedom or influence they have to alter a workflow. Within a country, workplaces differ in digital readiness and training opportunities. Across countries, the stock of jobs and institutions differs as well. Gender can shape the translation of the same exposure into use, while perceived social risk can make hesitation more or less reasonable. Finally, even if adoption occurs, the evidence must still answer a separate question: has the task bundle changed because of AI?
That is the problem this series will keep in view. Generative-AI exposure identifies a potential relationship between tools and work. Adoption is the outcome of a conversion process involving capability, context, authority, infrastructure, and legitimacy. The next step is to examine how exposure is measured and how much of that potential reaches workers in practice—without mistaking a statistical gradient for a list of occupations already transformed.