A shared AI professional vocabulary crystallized quickly in early 2024, but practitioners never coalesced into a distinct occupation; instead, AI skills diffused into existing jobs rather than creating a new 'AI Engineer' class.

Main Finding

The paper introduces a zero-assumption NLP method that defines an occupation as a bipartite "co-attractor" of people and vocabulary, and operationalizes tests for its emergence in resume data. Applied to 8.2 million US resumes (Aug 2022–Jan 2026), the method finds that AI exhibits a striking asymmetry: a tightly cohesive AI tooling vocabulary crystallized rapidly in early 2024, but the practitioner population never formed a mutually sustaining occupational cohort. In short, AI behaves like a diffusing technology absorbed across existing careers, not as a newly emergent occupation — though creating an “AI Engineer” occupational category could, in principle, catalyze population cohesion around the already-formed vocabulary.

Key Points

• Co-attractor concept: an occupation is a mutually sustaining pair — a cohesive vocabulary and a cohesive population — each necessary for the other.
• Push vs. Pull: compression/algorithmic artifacts (“push”, e.g., frequent tokens like “ai”) can produce apparent topics; genuine semantic co-occurrence among terms and genuine population similarity (“pull”) must be separated.
• Dual, symmetric tests:
  • Vocabulary cohesion: permutation-tested co-occurrence on XTX (term-term) to detect pull beyond frequency expectations.
  • Population cohesion: permutation-tested co-occurrence on XXT (document-document) to detect groups of practitioners who are similar beyond focal terms.
• Ablation: removing the candidate vocabulary from resumes and retesting population cohesion determines whether vocabulary is necessary for the observed population cluster (tests mutual dependence).
• Trifactor NMF (X ≈ F S Gᵀ): used to independently identify document groups F (populations), term groups G (vocabularies), and the coupling S; large diagonal S[k,k] indicates strong within-topic population–vocabulary coupling.
• Empirical AI result: three independent timelines show (1) rapid vocabulary lock-in (early 2024), (2) dissolution/absence of a cohesive AI practitioner population, and (3) population signals driven mainly by generic signaling terms rather than tooling vocabulary.
• Validation: method recovers established occupations in held-out tests (paper reports validating on two known occupations).
• Interpretation: AI is observed as a diffusing general-purpose technology rather than a self-contained emerging occupation; prior AI practitioner communities fragmented as AI tools became mainstream.

Data & Methods

• Data
  • Source: BOLD platform ecosystem resumes (users of resume-builder tools).
  • Sample: 8.2 million US resumes, August 2022 – January 2026.
  • Temporal resolution: monthly windows (42 months).
• Preprocessing / representations
  • Document-term matrix X (rows = resumes, columns = terms/skills/vocabulary tokens).
  • Cosine similarity used for pairwise co-occurrence measures.
• Push vs. pull filtration (compressionless co-occurrence test)
  • Null construction: independently permute each column of X 200 times to preserve term frequencies but destroy co-occurrence structure.
  • For each term pair (or document pair), compare observed similarity to the empirical permutation distribution; mark edges significant after Benjamini–Hochberg correction.
  • Result: Boolean masks Mvoc and Mpop that keep only edges driven by pull (co-occurrence beyond frequency).
• Validation of groups (hypergeometric density test)
  • Given the pull-filtered graph, compute whether a candidate vocabulary or population group has more significant internal edges than expected by chance using a hypergeometric/urn model; report density ratio (>1 indicates over-connection).
• Topic/co-cluster discovery
  • Trifactor NMF (X ≈ F S Gᵀ) to identify document clusters (F), vocabulary clusters (G), and coupling strengths (S).
  • Use S to measure within-topic coupling; use ablation (zero-out vocabulary terms in X and retest population cohesion on XXT) to test necessity of vocabulary for population coherence.
• Multiple complementary diagnostics: concentration of F[:,k], cohesion on raw XXT, significant vocabulary density on XTX, and ablation-driven changes.
• Statistical thresholds and corrections reported (permutation repeats = 200; multiple testing control via Benjamini–Hochberg).

Implications for AI Economics

• Measurement and monitoring
  • Existing taxonomies (SOC, O*NET) and five-year surveys lag true occupational dynamics; the paper demonstrates a scalable, high-frequency resume-based approach to detect nascent occupations or diffusion in near–real time.
  • The co-attractor tests separate buzzword-driven signals from genuine occupational formation — important for accurate labor market measurement and forecasting.
• Labor market interpretation
  • If AI is primarily diffusing across occupations rather than forming its own occupation, policies focusing on retraining should emphasize upskilling within existing career ladders (role-specific tool adoption, integration into domain tasks) rather than creating separate AI-only career tracks.
  • Employer classification and hiring taxonomies that treat AI expertise as a cross-cutting skill (rather than a discrete occupational label) may better reflect current labor-market structure — unless active steps are taken to institutionalize a new occupation.
• Policy and credentialing
  • Introducing a formal occupational category (e.g., “AI Engineer”) — via standard occupational classification, credentialing programs, or industry hiring taxonomies — could act as a coordination mechanism and potentially catalyze population cohesion around the already-formed vocabulary (i.e., complete the co-attractor). Whether this is desirable depends on trade-offs (e.g., labor market segmentation vs. clearer training pipelines).
• Workforce forecasting and education
  • Forecasts that assume rapid emergence of a unified AI occupation may overstate labor reallocation; instead expect AI skill diffusion to alter task content and productivity within many occupations.
  • Educational and training programs should prioritize embedding AI tool fluency into domain curricula and continuing education for incumbent workers rather than primarily funding standalone “AI degrees.”
• Research and policy caveats
  • The method relies on resume text (supply-side signal); employer-side uptake and job-posting dynamics matter for wages and demand. A combined analysis (resumes + job postings + employer surveys) would give a fuller picture.
  • Data-source considerations: BOLD’s resume sample is described as broader than LinkedIn’s, but all resume platforms have selection biases that should be acknowledged when generalizing to the whole labor force.
  • Ablation establishes necessity of vocabulary for population cohesion in the data, not causal mechanisms in the broader labor market.
• Practical recommendation
  • Statistical occupational monitoring systems (national statistical agencies, workforce boards, and research centers) should adopt dual-side co-occurrence tests (vocabulary and population) and ablation diagnostics to distinguish occupation formation from technology diffusion. For AI-specific policy, prioritize cross-occupation upskilling pathways while tracking whether institutional actions (classification, credentials) are producing emergent co-attractor dynamics.