So Will the Job Change?

The subject of this series is the path from AI capability to occupational exposure, workplace adoption, and task-level change. The final question is whether an exposure signal can tell us that a job will actually change. The most defensible answer is no—not by itself. Exposure identifies where AI may be close to some of the capabilities a role requires. Adoption shows whether workers and organisations act on that potential. Evidence of task restructuring is a separate test, and the European study did not detect an average change in worker-reported task bundles attributable to early adoption.

That answer is not a dismissal of exposure measures. It is a statement about what each link in the chain can establish. A capability map can be useful without being predictive in the everyday sense. It can show where to investigate, which bottleneck matters, and what kind of evidence a workplace should gather next. The mistake is to ask one score to stand in for a process that includes technology, work design, incentives, authority, and time.

Use is not the same as redesign

Henseke’s European study makes the final distinction concrete. In the 2024 European Working Conditions Survey, workers who reported using AI were also more likely to report that technology had removed tasks or created new ones. After accounting for digital tool use and a broad set of characteristics and fixed effects, adopters reported about 13 percentage points more task displacement and 10 points more task creation than non-adopters. At first glance, that looks like evidence that adoption changes jobs.

But the comparison has a plausible alternative explanation. Workers in jobs that were already being reorganised may have been more likely both to adopt AI and to report task changes. A firm experimenting with new processes might attract early adopters, or a worker with more discretion might seek out a tool because their job was changing for other reasons. In that case, adoption and task change would be correlated even if adoption itself caused little of the observed restructuring.

The study therefore moves to a more demanding analysis. It restricts attention to people who use a computer at work, the population with the basic infrastructure to use AI software. It collapses observations into one-digit occupation-by-country cells, adds occupation and country fixed effects, and instruments cell-level adoption with occupational exposure interacted with leave-one-out country adoption. In plain language, the design uses the prediction that highly exposed occupations should show more adoption in countries where adoption is generally more advanced, while absorbing broad differences between occupations and countries.

The descriptive association weakens as those comparisons become more demanding. Task displacement loses its association with adoption after occupation fixed effects. Task creation remains associated after that step but weakens further after country effects. The instrumental-variable estimates cannot reject zero for either outcome. Robustness checks that account for working from home or use an independent change in enterprise-level AI deployment reach the same substantive conclusion, although one check suggests that part of the instrument also captures broader digitalisation.

The conclusion is deliberately narrower than “AI has not changed any jobs.” The evidence supports no detectable average effect of early generative-AI adoption on worker-reported task displacement or creation in this cross-section. A null estimate can be consistent with several realities: workers may select into tools after their jobs begin changing; AI may be fitted into existing processes without producing a task workers recognise as new or removed; effects may be concentrated in particular occupations and cancel out in an average; or the survey may not measure the relevant change well.

Timing matters. The survey captures adoption roughly eighteen months after ChatGPT’s release, and its adoption measure is binary. Someone who used a tool once and someone who relies on it daily can both be counted as adopters. The task questions are also technology-generic: they ask whether technology has removed or created tasks, not whether generative AI was the specific cause. Backend systems invisible to workers are outside the worker-reported measure. These limitations make the null less sweeping, not less useful. They define what “no detectable effect” means.

The result also explains why visible use can precede visible redesign. A worker may ask a system to draft a response, then spend the saved time checking the draft, correcting errors, or completing the same downstream process. The tool has changed the worker’s method without necessarily changing the task categories recognised by the survey or the organisation. A manager may call this productivity improvement; a worker may call it extra quality control.