Keyword-based automated screening inflates search frictions by misreading candidate skills, while semantic, vector-based matching markedly raises candidate recall and matching efficiency in simulations; adopting interoperable, verified candidate-side signals could cut matching costs but requires real-world validation.

Main Finding

The paper defines and formalizes "Artificial Frictional Unemployment" (AFU): a labor‑market inefficiency produced when deterministic, keyword‑oriented Applicant Tracking Systems (ATS) systematically reject qualified candidates because of semantic/lexical mismatch rather than true skill gaps. A controlled simulation shows that semantic, vector‑embedding based matching greatly reduces false negatives caused by vocabulary variance; a candidate‑side architecture (JobOS) is proposed to operationalize semantic competency mapping alongside existing ATS.

Key Points

• AFU: a new framing that places vocabulary/semantic misinterpretation by automated screening inside labor‑market theory and information‑asymmetry analysis (Akerlof‑style market for lemons).
• Deterministic ATS act as high‑precision, low‑recall classifiers; engineered to avoid false positives, they produce unobserved false negatives that lengthen searches and suppress effective labor supply.
• Semantic gap: equivalent competencies are expressed diversely across industries, cultures, and career paths; exact matching treats that variance as absence.
• Simulation (controlled, synthetic) isolates representation as the causal mechanism: keyword screening collapses with lexical variance while semantic matching remains robust.
• Tradeoffs: semantic retrieval raises recall substantially but lowers precision — making it better suited upstream of human review (surface more candidates for evaluation) rather than making final automated accept/reject calls.
• JobOS (design proposal): candidate‑side pipeline with ingestion/normalization, semantic embedding & vector indexing, verification/simulation (role tests), and candidate‑centric storage/governance (PII masking, revocable access). Prototype built to show feasibility (React frontend, serverless orchestration, PostgreSQL + pgvector, embeddings via Gemini/all‑MiniLM).
• Ethics & governance: semantic methods reduce vocabulary bias but do not eliminate historical/proxy bias; transparency, auditability, privacy, and candidate ownership are central design concerns.
• Scope: simulation demonstrates mechanism, not prevalence. Estimating AFU’s real‑world size requires field data, ATS logs, and trials.

Data & Methods

• Objective: isolate how lexical variance alone produces false negatives under deterministic screening.
• Synthetic dataset: N = 1,000 resume–job pairs; vocabulary of 35 skills; each job has required skill set; candidate labeled "qualified" if they cover ≥80% of required skills. After labeling, each skill is rendered with randomized surface forms (synonyms, acronyms, title variants) to inject lexical variance while preserving ground‑truth competency.
• Baseline (keyword): whole‑word matching on required terms; candidate advances if matched fraction ≥ 0.4.
• Semantic pipeline: sentence‑transformer embeddings (all‑MiniLM‑L6‑v2, 384‑d); cosine similarity with threshold τ (main result uses τ ≈ 0.49). Approximate nearest neighbor indexing used for retrieval.
• Metrics: precision, recall, F1; attention to false negatives and threshold sensitivity.
• Reproducibility: fixed random seed; code and synthetic data generator available from the author for replication.
• Prototype components: React.js frontend, serverless orchestration, PostgreSQL + pgvector for vector indexing, embedding via Gemini 2.0 Flash API. Prototype demonstrates feasibility but is not evaluated on hiring outcomes.
• Key quantitative outcomes (illustrative, from synthetic experiment):
  • As lexical variance increases: keyword F1 falls from 0.92 → 0.28; semantic F1 remains ≈0.73 → 0.69.
  • At 70% lexical variance: keyword precision 0.90, recall 0.52, F1 0.66; semantic precision 0.61, recall 0.90, F1 0.72.
  • Of 500 qualified candidates, keyword baseline rejected 239 (48% false negative rate); semantic pipeline reduced false negatives to 51.
• Limitations: synthetic data and controlled assumptions exclude recruiter behavior, downstream interviews, and real‑world noise; conclusions are about mechanism, not prevalence.

Implications for AI Economics

• Matching efficiency and measured unemployment: AFU provides a plausible algorithmic channel for an outward shift in the Beveridge Curve — vacancies and unemployment coexisting because automated upstream filters impede matches. Correcting AFU could increase effective labor supply and reduce search duration.
• Information asymmetry & institutional design: ATS were adopted to manage uncertainty (reduce costly false positives). That institutional response can deepen information asymmetry by degrading resume signals (lexical filtering reduces observable signal variety), producing negative externalities at market scale. Candidate‑side infrastructures (e.g., JobOS) reallocate where translation/verification happens, potentially reducing asymmetry.
• Precision–recall tradeoffs have macroeconomic implications:
  • Raising recall (semantic matching) surfaces more qualified applicants, increasing hiring pipelines and potential employment. But it also increases employer screening costs (more human review) or imposes additional verification expenses.
  • Firms calibrate thresholds based on private costs; absent coordination or regulation, socially optimal thresholds (minimizing aggregate search frictions) may not be chosen.
• Distributional effects: AFU disproportionately excludes "non‑traditional" or "hidden" workers (career breaks, caregiving, military backgrounds, informal skill expressions), amplifying inequality. Semantic approaches reduce vocabulary‑driven exclusion but must be paired with bias audits to avoid proxy harms.
• Labor market equilibrium and wages: by increasing the set of candidates who are seen as qualified, reduced AFU could:
  • Increase effective labor supply to particular job types, exerting downward pressure on wages in the short run, or
  • Improve match quality, raising productivity and potentially wages for better matched workers. The net wage effect is ambiguous and depends on which margin dominates.
• Employer incentives and strategic responses:
  • Lowering algorithmic rejection may induce higher applicant volumes and raise employer screening costs. Firms may respond by increasing human review capacity, investing in verification, or shifting to other hard filters (e.g., credential gating).
  • Candidates may respond strategically (resume tailoring, keyword gaming); candidate‑side verified competency signals could mitigate strategic noise but require governance and cost structures.
• Policy and regulation:
  • AFU reframes automated hiring as labor‑market infrastructure, legitimizing policy levers (coverage of hiring AIs as high‑risk, mandatory audits, transparency requirements, data access for researchers).
  • Interventions could include standards for recall/precision reporting, requirements to preserve appeal/feedback pathways, or incentives/subsidies for verification/semantic tooling in public hiring.
• Measurement needs for economics research:
  • To quantify AFU’s macroeconomic importance, researchers need ATS logs, matchedadministrative employment outcomes, randomized experiments (e.g., A/B testing different upstream representations), and sectoral studies of vocabulary heterogeneity.
• Governance and welfare:
  • Candidate‑centric control of data and verifiable competency signals could shift bargaining power and reduce asymmetry, but introduces costs (verification, credentialing) and raises privacy risks; welfare analysis should include these transaction costs.
• Research agenda:
  • Field experiments that replace/augment ATS filters with semantic pipelines and measure hires, time-to-hire, wages, and demographic effects.
  • Structural models to integrate algorithmic filtering into search/matching frameworks (e.g., extending Mortensen–Pissarides) to predict general equilibrium effects.
  • Cost‑benefit analyses comparing employer screening costs to social gains from reduced AFU.

Summary: The paper identifies an algorithmic source of matching inefficiency (AFU) driven by representation choices in automated hiring. From an AI‑economics perspective, correcting representation (semantic mapping, verification, candidate control) is plausibly welfare‑improving but involves tradeoffs in employer costs, precision vs. recall, distributional consequences, and regulatory design. Quantifying the macroeconomic significance of AFU requires field data and experimental evaluation.