From Unequal Uptake to Equitable Work
The answer to this series’ opening question is now more precise than “some workers are ready for AI and others are not.” Generative-AI exposure identifies where a job could be changed, but adoption depends on whether workers can evaluate and integrate the tool, whether workplaces give them access and influence, whether national systems support digital work and learning, and whether the technology is judged trustworthy enough to use. This final synthesis connects those conditions to the separate question of task transformation, so the reader can distinguish what the evidence supports from what remains unknown and make a more defensible judgment about equitable adoption.
Use is not the same as change
The most important boundary concerns what a reported adoption rate can tell us about work. The European survey records whether a worker uses AI at work. It does not, by itself, reveal whether the worker’s tasks have been removed, newly created, accelerated, degraded, or quietly reorganised. That distinction matters because a tool can be useful inside an existing task without changing the task category a worker recognises. A person may use AI to draft, search, or organise information and still perform the same broad responsibility, with new checking and editing work folded into it.
The data show why a simple comparison can mislead. European adopters report 13 percentage points more technology-related task displacement and 10 percentage points more task creation than non-adopters after extensive controls. Across occupations, adoption and reported task restructuring are strongly associated, with a Spearman correlation of 0.92. Those patterns are consistent with real changes around adoption. They are not yet evidence that generative AI caused them.
Selection is the central problem. Workers in jobs that are already being reorganised may be more likely to try a new tool, and they may also be more likely to notice that their tasks have changed. A firm investing in new processes may encourage AI use while independently changing the way work is allocated. An occupation undergoing broader digitalisation may show high exposure, high adoption, and high restructuring even if the AI tool itself is not the cause of the change. Comparing adopters with non-adopters without accounting for those paths risks attributing a changing workplace to the newest technology present in it.
The study therefore uses a more demanding design. It restricts the analysis to workers who use a computer at work, the population with the basic digital infrastructure to use AI software, and collapses the data into 344 one-digit occupation-by-country cells. It then uses occupation and country fixed effects to absorb tendencies for some occupations to restructure faster and for some countries to experience common changes. A shift-share instrument combines occupational exposure with leave-one-out country adoption, creating variation in cell-level adoption that should be stronger for exposed occupations in countries where uptake is generally higher.
Under the design’s assumptions, the striking descriptive relationships do not survive as detectable average AI effects. The instrumental-variable estimate cannot reject the null for either task displacement or task creation. The main first stage is strong, with an F-statistic of 43.2. Adding the share of workers who sometimes work from home weakens the first stage to 18.7, suggesting that part of the variation may be connected to broader digitalisation. Using an independent enterprise-adoption change as the shift produces substantively similar results, with a first-stage F-statistic of 32.5 in the smaller sample.