Pith. sign in

REVIEW 4 major objections 6 minor 52 references

Helping People Choose Careers in the Age of AI

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper argues that averaging five recent AI-exposure models reveals healthcare practice as the occupational field best combining above-median pay with below-median AI exposure.

desk verdict A genuinely useful model comparison plus a new usage-based exposure measure, but the career-advice headline rests on hand-set exposure percentages that are never sensitivity-tested. read the letter →

arxiv 2607.15506 v1 pith:CN6F2T45 submitted 2026-07-16 econ.GN q-fin.EC

Jennifer L. Steele , Isabella Cruz This is my paper · ORCID
classification econ.GNq-fin.EC
keywords AIexposureoccupationalchoicewagedifferentialscareerguidancequery-basedmeasurecross-modelaverageLLMusagedatahealthcarepractice
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

To give career guidance in an uncertain AI transition, the paper compares six existing projections of which occupations AI will touch, finds they disagree sharply, and builds a new measure from actual 2025 usage of the two dominant AI assistants. It then averages its measure with four other recent projections, using the composite to rank 863 occupations by salary and AI exposure. The central result is that healthcare practice jobs—physicians, nurses, pharmacists, therapists—are the most consistent winners: above-median pay, below-median exposure. The paper also finds that high-paying jobs tend to use AI as an assistant rather than a full substitute, and warns that strong office professions (finance, computing, law) carry above-median exposure even if demand does not collapse. A sympathetic reader would take this as evidence that a diversified, usage-informed measure can give career seekers a more stable signal than any single forecast.

What carries the argument

The load-bearing device is the paper's query-based exposure score. It is constructed by (a) assigning each of the two AI assistants' 2025 task queries to occupations, (b) grouping tasks by query volume into ventiles and deciles, (c) assigning fixed automation percentages to those buckets—80/70/60/50% for the top four ventiles, a 5% floor for never-queried tasks, and 30/25/20/15/10/5% across the deciles—with a downgrade to 40% for extremely augmentation-heavy tasks, and (d) averaging the two assistant-based job scores. This score then enters a five-model average (the paper's measure plus four recent alternative projections), which is what actually produces the salary-exposure quadrant chart.

What would settle it

Recompute the composite scoring with alternative stepped mappings — e.g., 60/50/40/30% for the top four ventiles and a 10% floor — and check whether healthcare practice still dominates the high-pay/low-exposure quadrant; or use realized 2026-2030 employment and wage changes to see whether the paper's 2025 labels track actual outcomes.

Watch

Extended reading notes

Core claim

Occupational AI exposure scores vary so much across six established projection models that no single one is trustworthy for career advice. The paper's own contribution is a usage-based measure: it takes millions of 2025 queries logged by the two leading AI assistant products, maps each query to the occupational task it serves, and translates query volume into an automation potential using a stepped, hand-specified schedule. It then averages this measure with four other recent, methodologically distinct projections to form a composite exposure score for 872 occupations. On this composite, occupations that pair above-median pay with below-median exposure concentrate in healthcare practice; ass

Load-bearing premise

The paper's own exposure measure depends on hand-set percentages (80/70/60/50/5 and 30/25/20/15/10/5) that convert query-volume ventiles and deciles into automatable task shares, and the paper does not show that other plausible percentages produce the same rankings.

Editorial extensions

If this is right

  • Averaging five heterogeneous projections yields a single usable signal for career counseling instead of six conflicting ones.
  • Under that average, healthcare practice jobs (physicians, nurses, pharmacists, therapists) most consistently pair above-median pay with below-median AI exposure; associate-degree skilled trades also do well.
  • High-paying office fields (management, finance, computing, engineering, law) mostly carry above-median exposure, so workers in them should expect task-composition changes.
  • In jobs where AI assistants are used heavily, higher pay correlates with using the tool as an assistant rather than a fully autonomous delegate, suggesting that the most-payable human work is still hard to delegate.
  • The paper's cross-model average gives a ranked list of occupations and fields that counselors can use immediately through an interactive tool.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Going beyond the paper: because the mapping from query volume to automatable share is hand-set, the tool is best treated as a ranking heuristic; a single alternative mapping could move specific jobs across the salary-exposure boundary.
  • Going beyond the paper: a natural out-of-sample test would be to check whether the high-pay/low-exposure occupations identified from 2025 data actually experienced slower task displacement or better wage growth in 2026-2030 than high-exposure fields.
  • Going beyond the paper: the finding that social and teaching tasks show high usage but mostly augmentation suggests that 'exposure' conflates assistance with substitution; a measure separating augmentable from substitutable task shares would be a sharper guide.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper compares seven occupational AI-exposure projections, introduces a new query-based exposure measure built from 2025 Anthropic Claude and OpenAI ChatGPT usage data, and then averages five models (including the new one) to produce a cross-model signal intended for career guidance. Using this average, it reports that above-median-salary, below-median-AI-exposure occupations are concentrated in Realistic, Social, and Investigative categories, associate-degree-level jobs, and healthcare practice fields (§5.7, Figure 16). The paper also examines augmentation versus automation in Claude usage and finds a modest salary premium for occupations that use Claude as a complement rather than a substitute.

Significance. If the central results are robust, the paper would provide a useful, policy-relevant synthesis of fast-moving AI-exposure research and a practical tool for counselors and career changers. Its main strengths are the careful side-by-side description of six prior projection models, the construction of a new empirically grounded measure from query-level data, the inclusion of a public data repository and interactive tool, and the explicit discussion of model heterogeneity. The cross-model averaging idea is sensible as a risk-diversification heuristic, and the healthcare-practice finding is a concrete, falsifiable claim. However, the new measure's construction rests on several hand-assigned exposure percentages that are not validated or tested for sensitivity, and the composition of the five-model average is not as clean as the text suggests. These issues affect the load-bearing career-guidance conclusion and require additional work before the paper's headline claim is fully supported.

major comments (4)
  1. [§4.3, Tables 3 and 4, §5.7] The Steele-Cruz exposure measure is built from hand-chosen exposure percentages: 80/70/60/50 for Claude ventiles 20–17, a 5% floor for the 85% of tasks with no queries (Table 3), and 30/25/20/15/10/10/5 for ChatGPT deciles 10–5 and 1 (Table 4). These constants are asserted rather than estimated, and no sensitivity analysis is provided. The paper itself notes in §4.3 that occupational exposure measures are sensitive to task aggregation, but it does not report how the results would change under alternative plausible coding schemes (e.g., lower upper-ventile percentages, a zero floor for no-query tasks, or different augmentation-downgrade thresholds). Because this measure is one of the five components of the cross-model average, and because the headline conclusion about healthcare practice (§5.7, Figure 16) is based on that average, the conclusion is not yet shown to be robust. Please repor
  2. [§4.3, Table 4] The ChatGPT decile mapping is incomplete as written. Table 4 lists deciles 10, 9, 8, 7, 6, 5, and 1, but says nothing about deciles 2–4. The accompanying text also omits them. The table's n-GWA column sums to 42 (4×6 + 18) rather than the 41 GWAs described in §4.2, so the mapping is internally inconsistent and not fully reproducible. A complete, consistent decile-to-exposure table is needed, including an explicit specification for deciles 2–4 or a statement that they are grouped with decile 1.
  3. [§4.4, §5.5] The justification for excluding Massenkoff and McCrory (2026) from the five-model average is that its correlation with the Steele-Cruz measure is 0.89, which would make the average over-reliant on Anthropic usage data. But the Steele-Cruz measure itself is 50% Anthropic-based and is also correlated 0.89 with Massenkoff-McCrory. Replacing Massenkoff-McCrory with Steele-Cruz does not remove the Anthropic signal; it retains it in a slightly different functional form. The cross-model average is therefore better described as four independent external projections plus a combined Claude/ChatGPT measure. Please show how the main conclusions change when (a) the average uses Massenkoff-McCrory in place of Steele-Cruz, or (b) both Anthropic-based measures are excluded, or (c) the average is computed with Steele-Cruz alone against the four external models without any Anthropic-based component.
  4. [§5.1, Figures 7 and 8] The claim that models published since 2020 show positive relationships between AI exposure and salaries and occupational complexity is supported only by binned scatterplots without confidence intervals or regression/rank-correlation statistics. Since the paper's discussion and career guidance rely heavily on these relationships, a quantitative summary—regression coefficients, rank correlations, or at least confidence bands for the binned means—is needed. This is particularly important because the slopes appear visually different across models, and the reader cannot assess whether the 'positive relationship' claim is statistically distinguishable from a null relationship in models such as Webb (2020) or Brynjolfsson et al. (2018).
minor comments (6)
  1. [§5.7] Typo: 'the assumption that very high-stakes takes are not suitable' should read 'tasks.'
  2. [Table 2] The first row label reads 'Steele (2026)' but should be 'Steele and Cruz (2026)' to match the author list.
  3. [Table 4] The row for decile 1 contains an extra '0' in the '% of Queries' column ('0.2 0 5%'), likely a formatting error.
  4. [§3, Table 1] The text and reference list cite 'Brynjolfsson et al. (2018),' but Table 1 labels the row 'Brynjolfsson and Mitchell (2017).' Please harmonize the citation.
  5. [§5.1] The text refers to 'the six models under consideration,' but the paper compares seven models. Clarify whether Figure 7 excludes a particular model and why.
  6. [§4.1, footnote 1] The footnote for the Anthropic data says 'data available at .' with a blank URL. Provide the full link or a DOI.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the query-based exposure measure is constructed from use data plus explicit assumptions, and the healthcare-practice conclusion is an empirical cross-tabulation with salary data, not a tautology.

full rationale

The paper's derivation chain is not circular. Its new measure (Tables 3 and 4) is built from observed Anthropic/OpenAI query volumes with explicitly assumed exposure percentages ('we must make assumptions...', §4.1; 'we set GWA-level AI exposure estimates lower...', §4.3). Those percentages are inputs, not parameters fitted to the salary/field outcomes the paper later analyzes. The central finding that healthcare practice occupations combine above-median pay with below-median exposure (§5.7, Figure 16) is an empirical cross-tab of an independently sourced salary ranking with the standardized five-model average; nothing in the definition of the exposure measure references 'healthcare' or 'median salary,' so the result is not forced by construction. The one self-reference—Table 1 listing 'Steele and Cruz (2026) (this paper)' and §4.4 choosing 'our estimates in lieu of Massenkoff's'—is a design choice about which measures to average, not an appeal to an unverified prior result; it does not carry the derivation. The paper even discloses the main non-circular weakness: the hand-set ventile/decile mappings are not sensitivity-tested, and §4.3 concedes that exposure estimates are 'somewhat sensitive to the level of task aggregation.' That is a robustness/validity concern, not evidence that the prediction is equivalent to its inputs. Therefore, under the rules requiring a concrete reduction (Eq. = Eq. by construction, or fitted parameter renamed as prediction), no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central empirical contribution, the Steele-Cruz exposure measure, relies on several hand-set coding rules that are not estimated, validated, or subjected to sensitivity analysis. The cross-model average inherits these arbitrary anchors. The rest of the analysis is descriptive, with no formal uncertainty quantification. No new physical or conceptual entities are introduced.

free parameters (4)
  • Claude ventile exposure rates = ventiles 20,19,18,17 = 80%, 70%, 60%, 50%; ventile 1 = 5%
    Hand-set in Table 3 to convert query frequency into exposure; no estimation, validation, or sensitivity analysis provided.
  • ChatGPT decile exposure rates = deciles 10,9,8,7,5-6,1 = 30%, 25%, 20%, 15%, 10%, 5%
    Hand-set in Table 4, and deliberately set lower than Claude rates because GWA-level data are coarser; no empirical justification for the specific values.
  • Augmentation-downgrade threshold and amount = 98th percentile of augmentation-automation differential; downgrade 70-80% to 40%
    Chosen to flag tasks strongly used for augmentation (teaching, counseling, scholarly writing); no independent evidence that 40% is the correct cap.
  • Model inclusion in cross-model average = exclude Frey-Osborne (2017) and Massenkoff-McCrory (2026)
    Decision in §5.4-5.5; Frey is excluded after being described as anomalous, and Massenkoff is excluded due to high correlation with the authors' own measure, which is then included.
assumptions (6)
  • domain assumption Observed query volume to Claude/ChatGPT is a valid proxy for the fraction of an occupational task that can be automated by near-future AI.
    Section 4.1 assumes high-query tasks are more automatable; the entire Steele-Cruz measure rests on this premise.
  • ad hoc to paper The ventile-to-exposure and decile-to-exposure mappings accurately reflect feasible automation.
    Tables 3 and 4; the values (80/70/60/50/5 and 30/25/20/15/10/5) are chosen by the authors, not derived from data or previous literature.
  • domain assumption Anthropic's usage-pattern labels (validation, learning, iteration = augmentation; feedback, directive = automation) capture the augmentation-automation construct.
    Section 4 uses Anthropic's taxonomy to distinguish complement vs. substitute use; this taxonomy is vendor-defined and not independently validated.
  • domain assumption Task-level Claude data can be merged with GWA-level ChatGPT data, with imputation for four missing GWAs, into a single comparable measure.
    Section 4.2-4.3; different levels of aggregation and imputation from 'similar GWAs' are assumed to be equivalent for exposure measurement.
  • ad hoc to paper Tasks with no observed Claude queries still have at least 5% AI exposure.
    Table 3, ventile 1: 'we set the exposure value at 5% rather than 0%' - this affects the 85% of tasks with zero queries and therefore almost all occupations.
  • domain assumption Standardizing each model to mean 0, SD 1 makes their ordinal rankings commensurable and averageable.
    Section 4.4/Figure 5; used for the five-model average, but the models measure different constructs (probability of full automation, share of automatable tasks, patent exposure, suitability index).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Helping People Choose Careers in the Age of AI." pith.science (2026). https://pith.science/paper/CN6F2T45

@misc{pith2026260715506,
  author       = {Pith},
  title        = {Pith review of: Helping People Choose Careers in the Age of AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CN6F2T45}},
  note         = {Machine review of arXiv:2607.15506}
}
read the original abstract

How should people choose careers when artificial intelligence (AI) is rapidly transforming the nature of work? We first compare six recent projections of occupational exposure to task automation with AI, examining their methods and assumptions. We then propose a new empirical model of occupational AI exposure based on 2025 query data from Anthropic and OpenAI. We find marked heterogeneity in model predictions, though models published since 2020 show positive relationships among AI exposure, salaries, and occupational complexity. To reduce uncertainty due to heterogeneous assumptions about task automation potential, we average the projections from five models, including our own. Using these averages, we report on likely tradeoffs between salaries and AI exposure across interest categories, O*NET Job Zones, and job fields. Jobs in healthcare practice show the strongest balance of higher pay with lower AI exposure. Among jobs making high use of Anthropic's Claude, those that use it as a complement rather than a substitute for human work are modestly higher-paying, though whether this pattern holds will depend on usage norms adopted in each field.

Figures

Figures reproduced from arXiv: 2607.15506 by the authors.

Figure 1
Figure 1. AI optimism versus senior salaries across 37 countries [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. Percent of queries by Generalized Work Activity (GWA) [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Occupational automation exposure scores from five approaches [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: Standardized exposure in seven key models [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Scatterplot matrix of standardized occupational exposure estimates [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Binned scatterplots of median salary by standardized AI exposure [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Binned scatterplots of occupational complexity by standardized AI exposure [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Number of U.S. workers by Job Zone and RIASEC Category in Millions (2022) [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 11
Figure 11. Figure 11: AI exposure estimates by RIASEC category for five most recent models [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]
Figure 13
Figure 13. Figure 13: Salary versus five-model AI exposure by RIASEC category [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]
Figure 14
Figure 14. Figure 14: Salary and AI exposure intersection by RIASEC category [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 3 canonical work pages

  1. [1]

    doi:10.2139/ssrn.6134506 , keywords =

    work page doi:10.2139/ssrn.6134506

  2. [2]

    Autor, David H. , doi =. NBER Working Paper Series , title =

  3. [3]

    Gillespie, Nicole and Lockey, Steve and Macdade, Alexandria and Ward, Tabi and Hassed, Gerard , doi =

  4. [4]

    and Levy, Frank and Murnane, Richard J

    Autor, David H. and Levy, Frank and Murnane, Richard J. , doi =. The Quarterly Journal of Economics , number =

  5. [5]

    Journal of Development Economics , keywords =

    Guzm. Journal of Development Economics , keywords =. doi:https://doi.org/10.1016/j.jdeveco.2010.08.007 , issn =

    work page doi:10.1016/j.jdeveco.2010.08.007 2010

  6. [6]

    Chatterji, Aaron and Cunningham, Thomas and Deming, David and Hitzig, Zoe and Ong, Christopher and Shan, Carl Yan and Wadman, Kevin , doi =

  7. [7]

    The Quarterly Journal of Economics , keywords =

    Emanuel, Natalia and Harrington, Emma and Pallais, Amanda , doi =. The Quarterly Journal of Economics , keywords =

  8. [8]

    AEA Papers and Proceedings , keywords =

    Brynjolfsson, Erik and Mitchell, Tom and Rock, Daniel , doi =. AEA Papers and Proceedings , keywords =

Show all 52 references
  1. [9]

    Beane, Matt , edition =

  2. [10]

    SSRN Working Papers , keywords =

    Webb, Michael , doi =. SSRN Working Papers , keywords =

  3. [11]

    Mollick, Ethan , booktitle =

  4. [12]

    and Price, Brendan , booktitle =

    Autor, David H. and Price, Brendan , booktitle =

  5. [13]

    Yang, Andrew , booktitle =

  6. [14]

    Brynjolfsson, Erik and Chandar, Bharat and Chen, Ruyu , institution =

  7. [15]

    Journal of the European Economic Association , keywords =

    Autor, David and Thompson, Neil , doi =. Journal of the European Economic Association , keywords =

  8. [16]

    My Next Move , keywords =

  9. [17]

    Nagle, Frank and Yue, Daniel , doi =

  10. [18]

    Lin, Luona , journal =

  11. [19]

    and Hazell, Jonathon and Restrepo, Pascual , doi =

    Acemoglu, Daron and Autor, David H. and Hazell, Jonathon and Restrepo, Pascual , doi =. Journal of Labor Economics, , number =

  12. [20]

    Tyrangeil, Josh , booktitle =

  13. [21]

    Oks, David , booktitle =

  14. [22]

    Strategic Management Journal , keywords =

    Felten, Edward and Raj, Manav and Seamans, Robert , doi =. Strategic Management Journal , keywords =

  15. [23]

    Noy, Shakked and Zhang, Whitney , journal =

  16. [24]

    Daedalus , keywords =

    Brynjolfsson, Erik , doi =. Daedalus , keywords =

  17. [25]

    Dreyer, Jacob , booktitle =

  18. [26]

    and Dorn, David , booktitle =

    Autor, David H. and Dorn, David , booktitle =. doi:10.3886/E112652V1 , keywords =

    doi

  19. [27]

    O*NET Online , keywords =

  20. [28]

    and Amodei, Dario and Kaplan, Jared and Clark, Jack and Ganguli, Deep , eprint =

    Handa, Kunal and Tamkin, Alex and McCain, Miles and Huang, Saffron and Durmus, Esin and Heck, Sarah and Mueller, Jared and Hong, Jerry and Ritchie, Stuart and Belonax, Tim and Troy, Kevin K. and Amodei, Dario and Kaplan, Jared and Clark, Jack and Ganguli, Deep , eprint =. arXi...

  21. [29]

    Bhuller, Manudeep and Havnes, Tarjei and McCauley, Jeremy and Mogstad, Magne , institution =

  22. [30]

    Science , month =

    Eloundou, Tyna and Manning, Sam and Mishkin, Pamela and Rock, Daniel , doi =. Science , month =

  23. [31]

    Holland, J. L. , doi =. Journal of Counseling Psychology , pages =

  24. [32]

    Economic Policy , keywords =

    Bessen, James E , doi =. Economic Policy , keywords =

  25. [33]

    Kharazian, Ara and Simon, Lisa and Stevens, Ryan , institution =

  26. [34]

    Jevons, William Stanley , publisher =

  27. [35]

    and Cruz, Isabella , institution =

    Steele, Jennifer L. and Cruz, Isabella , institution =

  28. [36]

    SSRN Working Papers , keywords =

    Lambert, Peter John and Schindler, Yannick , doi =. SSRN Working Papers , keywords =

  29. [37]

    SSRN Working Papers , keywords =

    Schubert, Gregor , doi =. SSRN Working Papers , keywords =

  30. [38]

    Economic Research by Indeed , keywords =

  31. [39]

    , booktitle =

    Acemoglu, Daron and Autor, David H. , booktitle =. doi:10.1016/S0169-7218(11)02410-5 , issn =

    doi

  32. [40]

    Science , month =

    Brynjolfsson, Erik and Mitchell, Tom , doi =. Science , month =

  33. [41]

    Acemoglu, Daron and Johnson, Simon , institution =

  34. [42]

    Technological Forecasting and Social Change , keywords =

    Frey, Carl Benedikt and Osborne, Michael A , doi =. Technological Forecasting and Social Change , keywords =

  35. [43]

    Journal of Development Economics , keywords =

    Chu, Angus C and Peretto, Pietro F and Wang, Xilin , doi =. Journal of Development Economics , keywords =

  36. [44]

    Wang, Vivian , booktitle =

  37. [45]

    Goldin, Claudia and Katz, Lawrence F , publisher =

  38. [46]

    Appel, Ruth and McCrory, Peter and Tamkin, Alex and McCain, Miles and Neylon, Tyler and Stern, Michael , keywords =

  39. [47]

    SSRN Electronic Journal , title =

    Brynjolfsson, Erik and Li, Danielle and Raymond, Lindsey , doi =. SSRN Electronic Journal , title =

  40. [48]

    Massenkoff, Maxim and McCrory, Peter , institution =

  41. [49]

    Management Science , month =

    Demirci, Ozge and Hannane, Jonas and Zhu, Xinrong , doi =. Management Science , month =

  42. [50]

    Agrawal, Ajay and Gans, Joshua and Goldfarb, Avi , journal =

  43. [51]

    Frontiers in Energy Research , keywords =

    Vivanco, David Font and Tukker, Arnold and Verbeek, Walter J V , doi =. Frontiers in Energy Research , keywords =

  44. [52]

    and Dahike, Jeffrey A

    Putka, Dan J. and Dahike, Jeffrey A. and Burke, Maura I. and Rounds, James and Lewis, Phil , institution =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.