pith. sign in

arxiv: 2606.07939 · v1 · pith:GE25S4KBnew · submitted 2026-06-06 · 💻 cs.CY

Stable Geometry, Reversing Poles: The Bipolar Structure of AI Occupational Substitutability and Its Decade-Scale Inversion

Shuyao Gao , Minghao Huang (aSSIST University , Seoul , South Korea) This is my paper

Pith reviewed 2026-06-27 19:25 UTC · model grok-4.3

classification 💻 cs.CY
keywords AI substitutabilityoccupational automationbipolar structuremicro-actionsO*NETpole inversionlabor marketsemantic typology
0
0 comments X

The pith

AI occupational substitutability has a stable bipolar structure whose high-risk pole has inverted over a decade.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests the assumption of a continuous gradient in AI exposure scores by breaking occupations into micro-actions and mapping them onto a 7-category semantic typology. It finds two poles: Tool-Mediated Physical activities with very low automation index and Planning & Design with much higher, separated by a large statistical gap that holds under multiple checks. The middle categories sit in a narrow band rather than spreading evenly. The structure itself stays consistent, but which activities count as high-risk has flipped from assessments made ten years ago. This matters because it shows that the jobs most at risk change as AI capabilities advance, rather than following a fixed ordering.

Core claim

Decomposing 1,961 O*NET Detailed Work Activities into 15,817 micro-actions and projecting the prior Occupational Automation Index onto a 7-macro typology produces a bipolar geometry. Tool-Mediated Physical (mean OAI 0.054) and Planning & Design (mean OAI 0.499) sit at the extremes with Cohen's d of 2.41. The geometry remains stable when resolution increases, the encoder changes, or external ratings are used, yet the order of the poles reverses relative to Frey-Osborne rankings, shown by a macro-level Spearman correlation of -0.750.

What carries the argument

The 7-macro semantic typology applied to micro-actions from O*NET Detailed Work Activities, serving as the projection space for the Occupational Automation Index to reveal polarity.

If this is right

  • The six middle macro categories form a low-contrast band with few equivalent pairs under equivalence testing.
  • The bipolar gap widens when the typology is refined from 7 to 15 categories.
  • Alternative encoders and external task ratings replicate the same lead for the LLM-based OAI at the poles.
  • The inversion means that earlier high computerisation risk for physical tasks now corresponds to low exposure in current projections.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the bipolar pattern persists at finer scales, labor analyses could shift from ranking all jobs on one axis to tracking two distinct frontiers.
  • Policy responses might need to address the moving target of which pole faces higher substitution as AI improves in different domains.
  • Replicating the decomposition on updated O*NET data could test whether the geometry continues to hold or begins to erode.

Load-bearing premise

The 7-macro typology derived from the LLM pipeline with expert calibration accurately reflects distinct substitutability dimensions at the level of individual micro-actions.

What would settle it

Human-coded automation exposure scores on the full set of micro-actions that do not produce a statistically significant separation between the Tool-Mediated Physical and Planning & Design groups.

Figures

Figures reproduced from arXiv: 2606.07939 by Minghao Huang (aSSIST University, Seoul, Shuyao Gao, South Korea).

Figure 1
Figure 1. Figure 1: UMAP-2d projection of the 15,817 micro-actions, coloured by K=7 macro assignment. [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Macro-cluster profiles across seven dimensions: stage shares (IC, NA, PD, ME, FV), [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: OAI distribution by analysis group, ordered by group median ascending. M2 (left) [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Resolution sweep K = 7 to K = 15. Left axis: the bipolar polar gap (high-pole mean OAI minus low-pole mean OAI) grows monotonically as the cut becomes finer. Right axis: the middle-pair non-significance rate falls from 100% at K = 7 to 48.9% at K = 12, then rises slightly to 56.4% at K = 15. The bipolar shape is structurally robust to resolution; the middle’s low-contrast band gradually resolves as finer c… view at source ↗
Figure 5
Figure 5. Figure 5: The M4 chimera. K=7’s M4 (Diagnostic Analysis) splits at [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: M2 intra-cluster heterogeneity. Left: overall OAI density for M2 with Hartigan dip statistic. Right: per-micro OAI distributions sorted by mean [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Bipolar reproduction under four independent exposure indicators. Each panel plots the [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: DWA-level scatter between our OAI and Eloundou’s GPT-4 task-rating exposure [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Intelligence-type composition by macro (BGE labels, row-normalised %, with M4 [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: DWA-level OAI distribution by dominant intelligence type ( [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Polar reversal at the macro level. Left: mean FO (2013, Oxford Martin original) on the x-axis vs mean OAI (2026) on the y-axis; the OLS regression slopes downward, with Spearman ρ = −0.650. Right: same plot against Eloundou’s 2023 task labels; the inversion is stronger, with Spearman ρ = −0.750. M2 and M4-HVAC (the high-FO, low-LLM-era macros) and M7 and M4-data (the low-FO, high-LLM-era macros) anchor th… view at source ↗
Figure 12
Figure 12. Figure 12: MPNet versus BGE intelligence-type labelling. Left: per-class share under each [PITH_FULL_IMAGE:figures/full_fig_p055_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Confusion matrices on the 150-row human audit. Left: human label vs MPNet [PITH_FULL_IMAGE:figures/full_fig_p056_13.png] view at source ↗
read the original abstract

Empirical research on the labor-market impact of artificial intelligence has converged, since Frey and Osborne (2017), on a continuous-gradient representation in which each occupation is assigned a real-valued exposure score on [0,1] obtained by linear aggregation across capability dimensions. This continuity is rarely articulated as an assumption and has not been tested at the micro-action level where substitution actually occurs. We decompose 1,961 O*NET Detailed Work Activities into 15,817 micro-actions using a multi-agent LLM pipeline with 31-expert HITL calibration, then project the DWA-level Occupational Automation Index from our prior work onto a 7-macro semantic typology. The result is a bipolar structure. Tool-Mediated Physical (M2, mean OAI = 0.054) and Planning & Design (M7, mean OAI = 0.499) form two extremes separated by Cohen's d = 2.41 (H = 172.88, p = 6.21e-34). The geometry is robust under three independent stress tests: resolution (K=7 to K=15, polar gap widens from 0.45 to 0.57), encoder swap to BGE (LLM-class OAI lead replicates at 3.37x), and Eloundou's GPT-4 task ratings (DWA-level rho = 0.635). The six middle macros form a low-contrast band between the poles (TOST at d=0.2 admits only 1/15 pairs as equivalent), not a flat plain. The geometry's stability does not, however, extend to its content. Across a decade, the polarity has inverted. Frey-Osborne (2013) placed Tool-Mediated Physical near the highest computerisation risk and Planning & Design near the lowest; our LLM-era OAI reverses that order, with macro-level FO-Eloundou Spearman rho = -0.750, p = 0.020, against the original Oxford Martin appendix. Which pole is high is therefore contingent on the era's dominant capability frontier, while the stable geometry itself is the structurally robust object.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper decomposes 1,961 O*NET Detailed Work Activities into 15,817 micro-actions via a multi-agent LLM pipeline with 31-expert HITL calibration, projects the authors' prior Occupational Automation Index (OAI) onto a custom 7-macro semantic typology, and reports a stable bipolar geometry: Tool-Mediated Physical (M2, mean OAI=0.054) and Planning & Design (M7, mean OAI=0.499) are separated by Cohen's d=2.41 (H=172.88, p=6.21e-34). The geometry is robust to three stress tests, the six middle macros form a low-contrast band, and the polarity inverts relative to Frey-Osborne (2013) with macro-level Spearman rho=-0.750 (p=0.020).

Significance. If validated, the result would demonstrate that occupational substitutability is not a flat continuous gradient but exhibits stable bipolar structure whose content (which pole is high-risk) is era-dependent, offering a falsifiable alternative to linear aggregation models and a framework for tracking capability-frontier shifts.

major comments (3)
  1. [Methods (LLM pipeline and typology construction)] Methods (LLM pipeline and typology construction): The reported means, Cohen's d=2.41, and bipolar claim rest on assignment of 15,817 micro-actions to the 7 macros; no inter-annotator agreement, error rates, confusion matrix, or released labeled dataset from the 31-expert HITL calibration is supplied, so systematic bias in micro-action bucketing cannot be ruled out and directly affects the polarity and inversion results.
  2. [Results (inversion and FO comparison)] Results (inversion and FO comparison): The decade-scale inversion is quantified by macro-level FO-Eloundou Spearman rho=-0.750 (p=0.020) on 7 categories; with such small N the result is sensitive to the custom typology choice and to any projection artifacts from the prior OAI, yet no bootstrap, leave-one-macro-out, or alternative grouping robustness check is reported.
  3. [Stress tests section] Stress tests section: The three stress tests (K=7 o15 resolution, BGE encoder swap, Eloundou GPT-4 ratings) are invoked to support robustness, but quantitative outcomes (e.g., exact polar-gap values, replication of d=2.41, or classification agreement metrics) are only partially stated, leaving open whether they address LLM-induced classification bias.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'three independent stress tests' is used without naming them or giving the key numeric outcomes (polar gap widening, 3.37x lead); a one-sentence enumeration would aid readability.
  2. [Notation] Notation: OAI is referenced before its first full expansion in some passages; ensure consistent first-use definition throughout.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful comments on our manuscript. We provide point-by-point responses below and commit to revisions where appropriate to address the concerns raised.

read point-by-point responses

  1. Referee: Methods (LLM pipeline and typology construction): The reported means, Cohen's d=2.41, and bipolar claim rest on assignment of 15,817 micro-actions to the 7 macros; no inter-annotator agreement, error rates, confusion matrix, or released labeled dataset from the 31-expert HITL calibration is supplied, so systematic bias in micro-action bucketing cannot be ruled out and directly affects the polarity and inversion results.

    Authors: We agree that additional details on the HITL calibration would strengthen the manuscript. In the revised version, we will include inter-annotator agreement statistics, a confusion matrix, and error rates from the 31-expert process. The labeled dataset will also be released publicly to allow independent verification. These additions will help rule out systematic bias in the micro-action assignments. revision: yes

  2. Referee: Results (inversion and FO comparison): The decade-scale inversion is quantified by macro-level FO-Eloundou Spearman rho=-0.750 (p=0.020) on 7 categories; with such small N the result is sensitive to the custom typology choice and to any projection artifacts from the prior OAI, yet no bootstrap, leave-one-macro-out, or alternative grouping robustness check is reported.

    Authors: The small sample size for the correlation is a valid concern. We will add bootstrap resampling to provide confidence intervals for the Spearman rho and perform a leave-one-macro-out sensitivity analysis in the revised results section. This will quantify the robustness of the inversion finding to the specific typology and any projection effects. revision: yes

  3. Referee: Stress tests section: The three stress tests (K=7 to 15 resolution, BGE encoder swap, Eloundou GPT-4 ratings) are invoked to support robustness, but quantitative outcomes (e.g., exact polar-gap values, replication of d=2.41, or classification agreement metrics) are only partially stated, leaving open whether they address LLM-induced classification bias.

    Authors: We will expand the stress tests section to include full quantitative details, such as the exact polar-gap values under each test, confirmation of d=2.41 replication where applicable, and classification agreement metrics between the LLM pipeline and expert ratings. This will more explicitly demonstrate that the tests mitigate concerns about LLM-induced bias. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The derivation projects a DWA-level OAI taken from prior work onto a 7-macro semantic typology obtained by LLM decomposition of 15,817 micro-actions. The reported means (M2 = 0.054, M7 = 0.499), Cohen's d = 2.41, and macro-level Spearman rho = -0.750 with FO are computed outputs from this projection and external comparison; they are not equivalent to the inputs by construction. The typology is defined semantically rather than fitted to maximize separation on OAI, the stress tests (resolution, encoder swap, Eloundou ratings) are independent robustness checks, and the polarity inversion is measured against the external Frey-Osborne benchmark. No self-definitional equations, fitted-input predictions, or load-bearing self-citations that collapse the central claim appear in the provided text. The chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim depends on a custom 7-macro semantic typology and LLM-based micro-action decomposition whose validity is asserted through robustness checks but not derived from first principles or external benchmarks; OAI values are taken from prior author work.

free parameters (1)
  • Number of macro categories (K)
    Set to 7 for projection of DWA-level OAI; stress test shows polar gap changes with K but choice is not derived from data.
axioms (2)
  • standard math Cohen's d, Kruskal-Wallis H, and TOST equivalence tests are appropriate for comparing OAI distributions across the 7 macros
    Invoked to establish statistical separation and non-equivalence of middle macros.
  • domain assumption The 7-macro semantic typology partitions substitutability dimensions without systematic bias from LLM classification
    Central to identifying the bipolar structure and its stability.

pith-pipeline@v0.9.1-grok · 5948 in / 1599 out tokens · 37688 ms · 2026-06-27T19:25:13.871231+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 8 canonical work pages · 3 internal anchors

  1. [1]

    Tasks, Automation, and the Rise in U.S

    Daron Acemoglu and Pascual Restrepo. Tasks, Automation, and the Rise in U.S. Wage Inequality. Econometrica, 90(5):1973–2016,

    1973

  2. [2]

    Deming and Kadeem Noray

    David J. Deming and Kadeem Noray. Earnings Dynamics, Changing Job Skills, and STEM Careers.The Quarterly Journal of Economics, 135(4):1965–2005,

    1965

  3. [3]

    Monroe and J

    DOI 10.1126/sci- ence.adj0998. Data files used in this paper are taken from the 2023 working-paper release (arXiv:2303.10130); references to “Eloundou et al. 2023” in the body text refer to the vintage of the underlying GPT-4 task ratings rather than to the publication year. Martha S. Feldman and Brian T. Pentland. Reconceptualizing organizational routine...

    work page doi:10.1126/sci- 2023

  4. [4]

    Working-paper antecedent of Frey and Osborne (2017). Cited here for the 702-row appendix table parsed programmatically in this paper’s external-indicator alignment; the published 2017 version contains the identical probability values (verified at Spearmanρ= 1.000across 653 matched SOCs). Carl Benedikt Frey and Michael A. Osborne. The Future of Employment:...

    2017

  5. [5]

    Bounded by Risk, Not Capability: Quantifying AI Occupational Substitution Rates via a Tech-Risk Dual-Factor Model

    arXiv:2604.04464. Paweł Gmyrek, Janine Berg, and David Bescond. Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality. ILO Working Paper 96, International Labour Organization,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    arXiv:2604.00186. J. A. Hartigan and P. M. Hartigan. The Dip Test of Unimodality.The Annals of Statistics, 13 (1):70–84,

    work page arXiv

  7. [7]

    How Exposed Are UK Jobs to Generative AI? Developing and Applying a Novel Task-Based Index

    arXiv:2507.22748. Lucija Ivančić, Dalia Suša Vugec, and Vesna Bosilj Vukšić. Robotic Process Automation: Systematic Literature Review. In Claudio Di Ciccio et al., editors,Business Process Man- agement: Blockchain and Central and Eastern Europe Forum (BPM 2019), volume 361 of Lecture Notes in Business Information Processing, pages 280–295. Springer, Cham,

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  8. [8]

    Leland McInnes, John Healy, and James Melville

    DOI 10.1007/978-3-030-30429-4_19. Leland McInnes, John Healy, and James Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction.arXiv preprint,

    work page doi:10.1007/978-3-030-30429-4_19

  9. [9]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    arXiv:1802.03426. Rachel Metz. OpenAI Sets Levels to Track Progress Toward Superintelligent AI. Bloomberg News, jul

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Re- ports OpenAI’s five-stage internal classification: Chatbots, Reasoners, Agents, Inno- vators, Organizations

    Industry framework; non-peer-reviewed. Re- ports OpenAI’s five-stage internal classification: Chatbots, Reasoners, Agents, Inno- vators, Organizations. URL: https://www.bloomberg.com/news/articles/2024-07-11/ openai-sets-levels-to-track-progress-toward-superintelligent-ai. Henry Mintzberg.The Nature of Managerial Work. Harper & Row, New York,

    2024

  11. [11]

    arXiv:2311.02462 (preprint Nov. 2023). Fionn Murtagh and Pierre Legendre. Ward’s Hierarchical Agglomerative Clustering Method: Which Algorithms Implement Ward’s Criterion?Journal of Classification, 31(3):274–295,

    work page arXiv 2023

  12. [12]

    URL: https://www.oecd.org/en/publications/2025/06/ introducing-the-oecd-ai-capability- indicators_7c0731f0.html

    Nine ability domains ×five capability levels. URL: https://www.oecd.org/en/publications/2025/06/ introducing-the-oecd-ai-capability- indicators_7c0731f0.html. Edith T. Penrose.The Theory of the Growth of the Firm. Basil Blackwell, Oxford,

    2025

  13. [13]

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Lan- guage Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China,

    2019

  14. [14]

    Iterative Repetition

    DOI 10.1145/3626772.3657878. Working-paper antecedent: arXiv:2309.07597 (2023). The BGE family (BAAI/bge-large-en-v1.5) is released as part of this package. AK= 5Robustness Cut The headlineK = 7partition (§3.4) was selected on dendrogram inspection of natural breakpoints. As a robustness backup, we report the raw Ward output atK = 5(the next informative c...

    work page doi:10.1145/3626772.3657878 2023

  15. [15]

    review the operational manual

    BGE matches the human ground truth at κ= 0.893(92 .0%accuracy); MPNet matches atκ= 0.769(82 .7%). The 9.3-point accuracy gap concentrates entirely on the 30 disagreement rows: on the 120 encoder-agreement rows, both encoders match the human at97.5%(only the three deliberate reversals are misses); on the 30 disagreement rows, BGE matches the human at70%ver...

    2018