Exposure Is Not the Same as Change
The subject of this series is the relationship between AI capability, occupational exposure, workplace adoption, and changes in the tasks people perform. The question is narrower—and more useful—than “Which jobs will AI replace?” It is: when a measure says that an occupation is exposed to AI, what has actually been established, and what still depends on workers, organisations, costs, institutions, and time? The answer gives readers a way to interpret exposure claims without turning a technical indicator into a prophecy about their livelihood.
That distinction matters because the word exposure carries two meanings in ordinary conversation. It can suggest vulnerability, as if a worker has been placed in the path of an unavoidable force. In the OECD’s framework, however, exposure means that the capabilities required by an occupation are relatively close to capabilities that current AI systems can perform. It is a measure of potential contact between a technology and a job’s requirements. It is not a measure of whether an employer has bought the technology, whether a worker uses it, or whether the job has already been redesigned.
The difference is visible in the two sources considered here. The OECD’s AI Capability Gap Index maps occupational requirements against a current AI frontier. Its results place routine, codifiable information work close to that frontier: Office and Administrative Support has the lowest major-group gap, at 0.8. Groups built around contextual judgement, social interaction, responsibility, or complex physical performance sit farther away; Community and Social Service has a gap of 6.4. Those numbers describe different capability distances. They do not say that an administrative worker has lost a job, or that a social-service worker is insulated from AI assistance.
The European evidence shows why another layer is necessary. In a 2024 survey of approximately 36,600 workers in 35 countries, reported generative-AI use at work averaged 12%, ranging from under 3% in Bosnia and Herzegovina to about a quarter in Luxembourg. Adoption rose from 1.5% in the least exposed occupational quintile to nearly a quarter in the most exposed. Exposure therefore helps explain who is more likely to use AI. But it does not explain everything: the association became smaller after accounting for skills, computer use, workplace characteristics, and industry, and it varied substantially across countries.
Nor does reported use automatically mean that the job’s task bundle has changed. Workers who adopted AI were more likely to report that technology had removed or created tasks. Yet when the study used a more demanding design to separate adoption from the tendency of already-changing jobs to attract adopters, it found no detectable average effect of early adoption on worker-reported task displacement or creation. That is not evidence that no task changed. It is evidence that exposure, adoption, and visible restructuring are separate empirical questions.
The shortcut from exposure to replacement
Imagine two claims about a medical transcriptionist and a judge. The first occupation is close to current AI capabilities in structured language work: transcribing, recording, classifying, and handling information. The second requires language too, but also interpretation in context, responsibility for consequential decisions, and forms of reasoning that are not captured by routine text processing. A language-centred exposure measure may place both occupations in a similar analytical neighbourhood. A broader capability map can place them farther apart because it asks what the whole occupational profile demands.
Neither result tells us what a particular workplace will do. A transcription office may lack the funds, workflow integration, or permission to deploy a capable system. A court may use language tools for research or drafting even if the occupation’s full capability profile remains distant from current AI. The first role can experience assistance without immediate elimination; the second can experience substantial task-level support without becoming an automatable role. “Exposed” and “unchanged” are not opposites, and “assisted” and “replaced” are not synonyms.
The OECD report makes this boundary explicit. It is designed to be forward-looking because it measures potential exposure before diffusion is complete. It is transparent because it keeps the assessment of AI capabilities separate from the assessment of occupational demands. It is updateable because either side can be revised as evidence changes. And it is linked to human abilities because it does not reduce work to one technology or one benchmark; it compares occupations across cognitive, social, sensory, and physical domains.
Those design choices answer a specific question: where is the current AI frontier close to the capabilities that occupations require? They do not answer the economic and organisational questions that follow. Is the system cheaper than a human operator? Can it be integrated with existing data and workflows? Will regulation permit its use? Do managers want to reorganise the process? Do workers trust the outputs and have enough influence to adapt the work? Are there reasons to retain human responsibility even when a system can perform part of a task? Capability proximity makes these questions more urgent; it does not settle them.
The European study supplies a useful reality check. A worker must first have a setting in which software can be used and a reason to use it. The survey’s adoption item asked whether the worker used AI at work that simplified complex mental tasks or made recommendations, with examples such as ChatGPT and DALL-E. It measured the extensive margin—whether the person used such a tool—not frequency, quality, or depth of integration. It also missed backend systems that workers may not notice. A capability map can therefore identify opportunity or pressure even when adoption is absent or invisible.
The arrows in this diagram are conditional, not inevitable. The comparison produces an exposure signal. Conditions shape whether that signal becomes use. Only then does it make sense to ask whether workers perceive tasks being removed, created, or rearranged—and even that observation needs careful identification, because jobs may already have been changing for other reasons.
Four different questions hidden in one score
A headline exposure number often compresses at least four questions.
First: what can the AI system do? Capability is domain-specific. A system may be strong at language processing or codified knowledge while remaining less capable in social understanding, contextual judgement, fine manipulation, or action in an unpredictable environment. “AI” is not one capability that advances everywhere at the same speed.
Second: what does the occupation require? An occupation is a bundle of tasks and responsibilities, not a single activity. A role can contain routine components that are highly exposed and other components that depend on capabilities still beyond the frontier. This is why an occupation-level score is a useful map but an incomplete description of a person’s work. The OECD identifies moving from occupations to tasks as an important next step for precisely this reason.
Third: will someone adopt the technology? The European results show a strong exposure-to-adoption gradient, but also show that the slope depends on enabling conditions. Workers with more formal education, more abstract task content, and greater influence over organisational decisions were more likely to convert exposure into uptake within the same occupational exposure. At the country level, digital intensity was a prerequisite, while workplace training and the abstract content of the national job stock helped distinguish countries that translated exposure into adoption more efficiently.
Fourth: what happens after adoption? Use may save time without changing the recognised task list. It may shift effort from drafting to checking, or from searching to deciding. It may create new activities that are too small or too recent for workers to report as distinct tasks. It may also leave the task bundle intact because the organisation has not redesigned roles around the tool. The European study’s null result on average early task restructuring is therefore compatible with real use and with local experimentation; it simply does not support a broad claim that early adoption had already caused measurable task displacement or creation across the studied setting.
This separation also changes how “high” and “low” exposure should be discussed. A low remaining capability gap indicates that current assessed AI is close to the requirements measured for that occupation. It does not mean the whole job will be automated. A high remaining gap indicates that important capabilities remain unmatched. It does not mean the job is safe from assistance, gradual redesign, or pressure to change. The gap is about the breadth and composition of the role as currently defined, not a permanent property of the people who hold it.
The practical implication is not to ignore exposure measures, but to ask better follow-up questions. Which capability domain is driving the score? Which task uses that capability, and how central is the task? What is the remaining bottleneck—social judgement, responsibility, dexterity, environmental adaptation, or something else? Who has access to the technology and the skills to evaluate it? Who can influence how work is reorganised? What evidence shows actual change rather than potential or self-selection?
That is why the central subject is not a ranking of occupations. It is the relationship between a changing capability frontier and a changing organisation of work. Exposure tells us where to look. Adoption tells us whether people and institutions are acting on that possibility. Task-level evidence tells us whether the work itself is being rearranged. The next step is to examine how the OECD turns that first question—how close is AI to an occupation’s requirements?—into a measurable capability gap, and why the details of that construction determine what the score can legitimately mean.