Pith

open record

sign in
Browse

arxiv: 2605.27399 · v1 · pith:HLS6NPC5 · submitted 2026-04-23 · cs.CY · cs.AI

Short-Term Gain, Long-Term Fragility: AI Labor Substitution and the Erosion of Sustainable Capability

Wolfgang Rohde This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 reserved 2026-07-05 01:27 UTCglm-5.2pith:HLS6NPC5record.jsonopen to challenge →

Figure 1
Figure 1. Figure 1: Mechanism of capability masking and capability erosion under AI labor substitution. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] reproduced from arXiv: 2605.27399
classification cs.CY cs.AI
keywords AI labor substitutioncapability erosioncapability maskingapprenticeship pipelineinstitutional short-termismtechnical debtcapability debtpolitical economy of AI
0
0 comments X

The pith

AI hiring freezes mask slow erosion of organizational capability

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

The paper argues that when organizations replace human workers with AI systems, they often mistake visible AI-generated output for genuine capability substitution. This creates a 'capability masking' effect: managers see plausible code, documents, or analyses and conclude that skilled human labor is no longer needed, leading to hiring freezes, staff reductions, and deferred investment in training. Meanwhile, the human systems that actually sustain long-term capability — tacit knowledge transfer, apprenticeship pipelines, institutional memory, peer review — quietly weaken. The paper calls this mechanism 'capability masking and capability erosion,' and frames it as a form of debt accumulation: technical debt in artifacts, capability debt in the human layer, and institutional debt in the broader structures that reproduce skill and resilience. The author synthesizes evidence from AI-assisted coding studies (where generated code remains uneven in correctness, security, and maintainability and still requires substantial human verification), labor-market research (showing reduced demand for AI-exposed tasks and softening junior developer hiring), and political-economy analysis (showing how managerial cost incentives, geopolitical competition, and platform concentration amplify substitution pressure). The central claim is that current AI adoption patterns systematically transfer burden from the present to the future in ways that standard productivity metrics cannot detect, producing systems that appear more efficient short-term while becoming more fragile over time.

Core claim

The paper identifies and formalizes a five-step mechanism: (1) AI produces visible, plausible output that appears to substitute for human work; (2) managers and policymakers infer from that output that underlying capability has been replaced; (3) organizations respond with hiring restraint, leaner staffing, and deeper dependence on external platforms; (4) hidden costs accumulate through verification burdens, loss of tacit knowledge, and contraction of the apprenticeship pipeline; (5) these firm-level decisions scale into broader fragility — weaker organizational resilience, narrower entry paths into professions, and more concentrated power. The key conceptual move is distinguishing between '

What carries the argument

The central mechanism is 'capability masking and capability erosion.' Masking is the perceptual and accounting-level error: visible AI output creates the appearance that organizational capability has been replaced. Erosion is the slower structural consequence: the human systems that actually sustain capability — tacit knowledge, apprenticeship pipelines, institutional memory, peer review — weaken over time. The paper also frames this as a layered debt accumulation: technical debt (short-term expedients in artifacts), capability debt (weakening human skill and review capacity), and institutional debt (deferred investment in governance and social systems).

Load-bearing premise

The paper's load-bearing premise is that the five-step mechanism is actually occurring at meaningful scale in real organizations — specifically, that firms are reducing human staffing based on the appearance of AI capability substitution, and that this is causing measurable erosion of tacit knowledge and apprenticeship pipelines. The paper is a conceptual synthesis rather than an empirical study, and the causal chain from 'AI output looks plausible' to 'organizations reduce'

What would settle it

The mechanism would be substantially weakened if firms using AI for labor substitution were simultaneously maintaining or expanding their apprenticeship and training pipelines, or if follow-up empirical studies found that organizations adopting AI aggressively showed no degradation in institutional memory, tacit knowledge transfer, or recovery capacity over multi-year horizons. It would also be challenged if AI-generated code quality improved to the point where verification burdens dropped to negligible levels and junior developer hiring rebounded.

Figures

Figures reproduced from arXiv: 2605.27399 by Wolfgang Rohde.

Figure 2
Figure 2. Figure 2: Related work map: four literatures illuminating different dimensions of the mechanism of [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A locally valid change can violate non-local constraints in a large system. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Surgical edit versus hammer rewrite in AI-assisted coding. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Apprenticeship pipeline contraction and delayed expertise loss. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Short-term visible gain versus long-term capability and resilience loss. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Masking versus solving: visible output above unresolved structural dependence. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
read the original abstract

What looks like acceleration can be a quiet transfer of burden from the present to the future. Attempts to replace human labor with AI systems are often presented as rational responses to technological progress, but that view is often structurally short-sighted. Across software development and adjacent knowledge industries, AI is increasingly attractive because it appears to reduce labor costs, speed output, and improve short-term metrics. Yet those gains may be achieved by drawing down human capabilities that are slow to build and difficult to restore. This paper develops a mechanism of capability masking and capability erosion under AI labor substitution. AI-generated output can create the appearance that organizational capability has been replaced, even when dependence on skilled human labor remains. That appearance can support hiring restraint while slower costs accumulate in the background. Evidence from AI-assisted coding shows that generated output still requires substantial human verification and remains uneven in correctness, maintainability, and security. Repository-level studies also suggest limits in handling broader codebase context. More broadly, labor-market, political-economy, and industrial-strategy evidence suggests that substitution pressures are being driven by managerial cost incentives and national competition while increasing risks of concentration and platform control. The result is a system that may look more efficient in the short term while becoming more fragile over time.

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

2 major / 8 minor

Summary. This manuscript proposes a conceptual mechanism —

Significance. The paper identifies a cross-domain mechanism — capability masking and capability erosion — that connects AI-assisted coding evidence, labor-market research, political economy, and industrial strategy into a single causal account. The contribution is integrative rather than empirical, and the paper is commendably honest about this scope (§2). It draws on a diverse, external evidence base (peer-reviewed studies, the METR RCT, BLS statistics, Senate testimony, OECD analysis) without circular self-citation. The five-step mechanism in §4 is clearly stated and falsifiable in principle: if firms that reduce staffing based on AI adoption do not subsequently experience capability erosion, the mechanism would be weakened. The debt framing (technical, capability, institutional) is a useful organizing device. The policy recommendations in §10, while somewhat generic, are proportionate to the argument.

major comments (2)
  1. §4, five-step mechanism, steps 1–2: The central load-bearing claim is that AI output creates a 'persuasive appearance' of capability substitution (step 1), which managers then act on (step 2). However, the paper's own evidence base argues against this persuasiveness for direct users. The METR study (ref 5, discussed in §5) shows experienced developers were 19% slower with AI and accepted fewer than 44% of generations. Refs 1–2 show code quality is uneven in correctness, maintainability, and security. The paper itself states 'human verification remains essential' (§4). This creates an internal tension: if AI output is visibly flawed to practitioners, who is actually being 'masked'? The paper gestures at this in §4 ('optimistic interpretation of AI output, and the political and financial convenience of accepting visible productivity as proof of real substitution'), but this passage实际上支持 an
  2. §4, five-step mechanism, steps 1–2: The central load-bearing claim is that AI output creates a 'persuasive appearance' of capability substitution (step 1), which managers then act on (step 2). However, the paper's own evidence base argues against this persuasiveness for direct users. The METR study (ref 5, discussed in §5) shows experienced developers were 19% slower with AI and accepted fewer than 44% of generations. Refs 1–2 show code quality is uneven in correctness, maintainability, and security. The paper itself states 'human verification remains essential' (§4). This creates an internal tension: if AI output is visibly flawed to practitioners, who is actually being 'masked'? The paper gestures at this in §4 ('optimistic interpretation of AI output, and the political and financial convenience of accepting visible productivity as proof of real substitution'), but this passage实际上支持 an
minor comments (8)
  1. §1: The paper cites corporate statements about layoffs and AI authorship rates (refs 16–19) as evidence that 'AI is already being used as part of the managerial justification for workforce reduction.' The causal link between AI adoption and specific layoffs is not established by these statements, and the paper acknowledges this ('These examples do not prove a uniform one-to-one replacement'). This is fine, but the framing could be tightened to avoid implying stronger causation than the evidence supports.
  2. §7: The evidence for apprenticeship pipeline contraction relies on a single CIO article (ref 7) and labor-market research on AI-exposed tasks (refs 8–9) that does not isolate junior software roles. The paper acknowledges this ('These findings do not isolate junior software roles on their own'), but the claim that 'this risk is no longer hypothetical' (§7) overstates what the cited evidence shows. Consider softening to 'this risk is becoming visible in early indicators.'
  3. §6: The GitHub Copilot study (ref 31) is described as 'vendor-produced' and the paper appropriately cautions against treating it as neutral. The DORA report (ref 29) is also vendor-produced (Google). The paper treats DORA more favorably without the same caveat. Consider applying the same disclosure consistently.
  4. Figures 1–7 are referenced but not visible in the reviewed manuscript text. Ensure that figures are legible, properly captioned, and that Figure 1 (the causal chain) clearly labels all five steps of the mechanism.
  5. §8: The discussion of data limitations (refs 33–35, Epoch AI and model collapse) is interesting but tangential to the core mechanism. Consider whether this material strengthens the argument or distracts from it. If retained, connect it more explicitly to the mechanism — e.g., how uncertainty about model maturity affects step 2 (managerial inference).
  6. §9: The discussion of platform concentration (refs 36–40) is wide-ranging and somewhat disconnected from the core capability-erosion argument. The link between concentration and capability erosion could be made more explicit.
  7. The paper uses epigraph-like sentences at the start of each section (e.g., 'When institutions mistake visible output for durable capability, speed becomes a solvent of memory.'). These are stylistically distinctive but may read as editorializing for a research audience. Consider whether they match the target venue's conventions.
  8. References: Several sources are from 2026 (refs 15, 19, 20–22, 26, 30, 32). Ensure all citations are publicly accessible and properly archived, as some appear to be very recent news or government sources that may not be stable.

Simulated Author's Rebuttal

1 responses · 0 unresolved

The referee raises a sharp internal-tension objection: if the paper's own evidence shows AI output is visibly flawed to practitioners, then who is actually being 'masked'? We agree the manuscript needs to make the masking target more precise and will revise accordingly. The referee's comment appears truncated, but the core challenge is clear and substantive.

read point-by-point responses

  1. Referee: §4, steps 1–2: The central load-bearing claim is that AI output creates a 'persuasive appearance' of capability substitution (step 1), which managers then act on (step 2). However, the paper's own evidence base argues against this persuasiveness for direct users. The METR study shows experienced developers were 19% slower with AI and accepted fewer than 44% of generations. Refs 1–2 show code quality is uneven. The paper itself states 'human verification remains essential.' This creates an internal tension: if AI output is visibly flawed to practitioners, who is actually being 'masked'?

    Authors: The referee identifies a genuine gap in the manuscript's argumentation, and we agree it must be addressed. The core issue is that the paper does not sufficiently distinguish between two audiences for AI output: practitioners who directly use the tools and decision-makers who are one or more organizational layers removed from the code. The masking mechanism does not require that experienced developers are fooled. It requires that managers, executives, and policymakers — who see aggregate metrics, cost savings, dashboards, and narratives rather than individual code generations — interpret visible output as evidence that capability has been replaced. The METR study actually strengthens this argument rather than weakening it: it shows that the people closest to the work are not persuaded, while the people making staffing decisions (e.g., the corporate examples in §1 — Microsoft, Amazon, Salesforce, Anthropic) are nonetheless acting as though substitution has occurred. This is precisely the asymmetry the mechanism describes. However, the referee is correct that the manuscript does not make this distinction explicit enough. The passage in §4 about 'optimistic interpretation of AI output, and the political and financial convenience of accepting visible productivity as proof of real substitution' gestures at the answer but does not develop it. We will revise §4 (steps 1–2) to clarify that the masking target is primarily organizational decision-makers rather than direct practitioners, and to make explicit that the paper's own evidence about practitioner skepticism is not a counterargument to the mechanism but a component of it: the very gap between practitioner awareness and managerial action is what makes the masking possible. We will also add a brief discussion noting that the revision: no

Circularity Check

0 steps flagged

No circularity: conceptual synthesis built entirely from external evidence with no self-citation or fitted-parameter predictions

full rationale

The paper is explicitly a conceptual synthesis (§2: 'This paper is a conceptual synthesis rather than a new empirical study') that integrates external evidence from four literatures. None of the 40 references are self-citations by the author (Wolfgang Rohde / AiSuNe Foundation). The five-step mechanism in §4 is a qualitative causal chain, not a mathematical derivation, so there are no equations that could reduce to inputs by construction. No parameters are fitted to data and then presented as predictions. No uniqueness theorem or prior ansatz from the same authors is invoked. The paper's definitions of 'masking' and 'erosion' are standard conceptual definitions, not circular (masking is defined as the perceptual phenomenon of overreading AI output; erosion is defined as the structural consequence of acting on that misreading — they are distinct concepts, not defined in terms of each other). The debt-framing (technical/capability/institutional debt) is explicitly presented as analogy, not as a new derivation. The argument's load-bearing evidence comes from independent external sources: METR's RCT (ref 5), NBER/Fed working papers (refs 8-9), BLS statistics (ref 22), GAO reports (ref 32), Nature (ref 34), and peer-reviewed software-engineering studies (refs 1-4). The paper is self-contained against external benchmarks and exhibits no circular reasoning.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 3 invented entities

The paper introduces no free parameters (it is a conceptual synthesis, not a quantitative model). Its axioms are domain assumptions about organizational behavior and AI capability trajectories, most of which are plausible but not directly tested within the paper. The three invented entities — capability masking, capability debt, and institutional debt — are conceptual labels for phenomena the paper argues are real but does not operationalize. None has independent falsifiable evidence outside the paper's own framing. The paper is transparent about this: it states it is a 'conceptual synthesis rather than a new empirical study' (§2).

axioms (5)
  • domain assumption AI-generated output creates a 'persuasive appearance' that organizational capability has been replaced, sufficient to influence managerial decisions about hiring and staffing.
    Stated in §4 as the first two steps of the mechanism. The paper provides illustrative evidence (corporate layoff announcements, Anthropic's AI authorship claims) but does not directly measure whether managers are actually making staffing decisions based on the appearance of AI capability.
  • domain assumption The human systems that sustain capability — tacit knowledge transfer, apprenticeship pipelines, institutional memory — are slow to build and difficult to restore once eroded.
    Invoked throughout §5 and §7. The paper draws on the COBOL/legacy systems analogy (ref 32, GAO) and labor-market research (refs 8-9) as supporting evidence, but the claim that these systems are 'difficult to restore' is treated as an assumption rather than directly tested.
  • domain assumption The mechanism extends by analogy from software development to other apprenticeship-based domains.
    Stated in §1 and §5: 'The discussion begins from software development, where the evidence is richest, and then extends by analogy to other apprenticeship-based domains.' The paper acknowledges software is used as a 'leading-indicator sector' but the analogical extension is assumed rather than demonstrated with domain-specific evidence.
  • domain assumption Current AI development paths may not mature enough to deliver the reliability assumed by full substitution narratives.
    Invoked in §8 to argue that firms are behaving as if substitution quality will improve predictably. The paper cites data pipeline constraints (ref 33, Epoch AI) and model collapse (ref 34, Nature) as supporting evidence, but the claim about future AI capability trajectories is inherently uncertain.
  • standard math Short-termism in corporate governance causes underinvestment in workforce training and innovation.
    Referenced in §1 and §4 via refs 23-24 (Terry 2023, Econometrica; Green et al. 2016). This is an established result in the corporate governance literature that the paper builds upon rather than derives.
invented entities (3)
  • Capability masking no independent evidence
    purpose: Names the perceptual/accounting phenomenon where visible AI output creates the appearance that capability has been replaced.
    Introduced in §4 as a conceptual label. The paper provides illustrative evidence (corporate behavior, AI authorship rates) but does not provide a falsifiable test that distinguishes 'masking' from actual capability substitution. No operationalization or measurement protocol is specified.
  • Capability debt no independent evidence
    purpose: Extends the technical debt metaphor to the human skill layer: short-term output preserved by borrowing against future human understanding.
    Introduced in §1 and §4. The debt framing is analytically useful but is a metaphorical extension, not a measured quantity. No accounting method or metric for capability debt is proposed.
  • Institutional debt no independent evidence
    purpose: Extends the debt metaphor to governance and social systems: deferred costs from underinvestment in institutions that reproduce skill and resilience.
    Introduced in §1. Same status as capability debt — a conceptual label without operationalization or falsifiable prediction.

pith-pipeline@v1.1.0-glm · 16772 in / 3923 out tokens · 396868 ms · 2026-07-05T01:27:41.281244+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

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    (2023).Evaluating the code quality of AI-assisted code generation tools

    Yetiştiren, B., et al. (2023).Evaluating the code quality of AI-assisted code generation tools. arXiv.https://arxiv.org/abs/2304.10778

    work page arXiv 2023

  2. [2]

    (2023).Security weaknesses of Copilot-generated code in GitHub projects: An empirical study

    Fu, Y., et al. (2023).Security weaknesses of Copilot-generated code in GitHub projects: An empirical study. arXiv.https://arxiv.org/abs/2310.02059

    work page arXiv 2023

  3. [3]

    (2023).RepoCoder: Repository-level code completion through iterative retrieval and generation

    RepoCoder. (2023).RepoCoder: Repository-level code completion through iterative retrieval and generation. arXiv.https://arxiv.org/abs/2303.12570 16

    work page arXiv 2023

  4. [4]

    RepoBench: Benchmarking Repository-Level Code Auto-Completion Systems

    RepoBench. (2023).RepoBench: Benchmarking repository-level code auto-completion systems. arXiv.https://arxiv.org/abs/2306.03091

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    (2025, July 10).Measuring the impact of early-2025 AI on experienced open-source developer productivity.https://metr.org/blog/2025-07-10-early-2025-ai-experienced -os-dev-study/

    METR. (2025, July 10).Measuring the impact of early-2025 AI on experienced open-source developer productivity.https://metr.org/blog/2025-07-10-early-2025-ai-experienced -os-dev-study/

    work page 2025

  6. [6]

    (2024, October 3).Introducing canvas.https://openai.com/index/introducing-c anvas/

    OpenAI. (2024, October 3).Introducing canvas.https://openai.com/index/introducing-c anvas/

    work page 2024

  7. [7]

    (2025, September 23)

    CIO. (2025, September 23). Demand for junior developers softens as AI takes over.https: //www.cio.com/article/4062024/demand-for-junior-developers-softens-as-ai-takes -over.html

    work page arXiv 2025

  8. [8]

    Hampole, M., Papanikolaou, D., Schmidt, L. D. W., & Seegmiller, B. (2025).Artificial intelli- gence and the labor market(NBER Working Paper 33509).https://www.nber.org/papers/w3 3509

    work page 2025

  9. [9]

    (2025).How retrainable are AI-exposed workers? Federal Reserve Bank of New York Staff Reports, No

    Hyman, B., Lahey, B., Ni, K., & Pilossoph, L. (2025).How retrainable are AI-exposed workers? Federal Reserve Bank of New York Staff Reports, No. 1165.https://www.newyorkfed.org/r esearch/staff_reports/sr1165

    work page 2025

  10. [10]

    (2025).The ethics of AI or techno-solutionism?.Journal of Education Policy.https: //www.tandfonline.com/doi/abs/10.1080/01425692.2025.2502808

    Selwyn, N. (2025).The ethics of AI or techno-solutionism?.Journal of Education Policy.https: //www.tandfonline.com/doi/abs/10.1080/01425692.2025.2502808

    work page doi:10.1080/01425692.2025.2502808 2025

  11. [11]

    (2024, March 26).AI washing and SEC enforcement.https://www.thomso nreuters.com/en-us/posts/investigation-fraud-and-risk/ai-washing-enforcement/

    Thomson Reuters. (2024, March 26).AI washing and SEC enforcement.https://www.thomso nreuters.com/en-us/posts/investigation-fraud-and-risk/ai-washing-enforcement/

    work page 2024

  12. [12]

    (2024).Global investor survey 2024.https://www.pwc.com/th/en/press-room/pres s-release/2024/press-release-26-12-24-en.html

    PwC. (2024).Global investor survey 2024.https://www.pwc.com/th/en/press-room/pres s-release/2024/press-release-26-12-24-en.html

    work page 2024

  13. [13]

    (2025).Competition in artificial intelligence infrastructure.https://www.oecd.org/e n/publications/competition-in-artificial-intelligence-infrastructure_623d1874-e n.html

    OECD. (2025).Competition in artificial intelligence infrastructure.https://www.oecd.org/e n/publications/competition-in-artificial-intelligence-infrastructure_623d1874-e n.html

    work page 2025

  14. [14]

    Ishkhanyan, A. (2025). The sovereignty-internationalism paradox in AI governance: Digital federalism and global algorithmic control.Discover Artificial Intelligence, 5, Article 123.https: //link.springer.com/article/10.1007/s44163-025-00374-x

    work page doi:10.1007/s44163-025-00374-x 2025

  15. [15]

    Papyshev, G., & Chan, K. J. D. (2026). AI regulatory strategies for digital sovereignty: The role of geopolitics and technological disparities.Electronic Markets, 36, Article 8.h tt p s: //doi.org/10.1007/s12525-025-00870-z

    work page doi:10.1007/s12525-025-00870-z 2026

  16. [16]

    CNBC. (2025a, May 13).Microsoft laying off about 6,000 people, or 3% of its workforce.https: //www.cnbc.com/2025/05/13/microsoft-is-cutting-3percent-of-workers-across-the-s oftware-company.html

    work page 2025

  17. [17]

    (2025, June 17).Message from CEO Andy Jassy: Some thoughts on Generative AI

    Amazon. (2025, June 17).Message from CEO Andy Jassy: Some thoughts on Generative AI. About Amazon.https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jas sy-on-generative-ai/

    work page 2025

  18. [18]

    (2025c, June 26).AI is doing up to 50% of the work at Salesforce, CEO Marc Benioff says.https://www.cnbc.com/2025/06/26/ai-salesforce-benioff.html

    CNBC. (2025c, June 26).AI is doing up to 50% of the work at Salesforce, CEO Marc Benioff says.https://www.cnbc.com/2025/06/26/ai-salesforce-benioff.html

    work page 2025

  19. [19]

    (2026, January 29).Top engineers at Anthropic, OpenAI say AI now writes 100% of their code, with big implications for the future of software development jobs.https://fortune

    Fortune. (2026, January 29).Top engineers at Anthropic, OpenAI say AI now writes 100% of their code, with big implications for the future of software development jobs.https://fortune. com/2026/01/29/100-percent-of-code-at-anthropic-and-openai-is-now-ai-written-b oris-cherny-roon/ 17

    work page 2026

  20. [20]

    Senate Commerce Committee

    U.S. Senate Commerce Committee. (2026, March 3).Less Hype, More Help: AI That Improves Safety, Productivity, and Care.https://www.commerce.senate.gov/2026/3/less-hype-mor e-help-ai-that-improves-safety-productivity-and-care

    work page 2026

  21. [21]

    Less Hype, More Help: AI That Improves Safety, Productivity, and Care

    Siemens. (2026, March 3).Testimony before the U.S. Senate Commerce Subcommittee hearing “Less Hype, More Help: AI That Improves Safety, Productivity, and Care”.https://www.comm erce.senate.gov/services/files/C9ACEB55-01EB-4F71-8E06-E6015C4C3886

    work page 2026

  22. [22]

    (2026, March 6).The Employment Situation - February 2026

    BLS. (2026, March 6).The Employment Situation - February 2026. U.S. Bureau of Labor Statistics.https://www.bls.gov/news.release/archives/empsit_03062026.htm

    work page 2026

  23. [23]

    Terry, S.J.(2023).The macro impact of short-termism.Econometrica, 91(5), 1881-1912.https: //doi.org/10.3982/ECTA15420

    work page doi:10.3982/ecta15420 2023

  24. [24]

    Green, F., Felstead, A., Gallie, D., Inanc, H., & Jewson, N. (2016). The declining volume of workers’ training in Britain.British Journal of Industrial Relations, 54(2), 422-448.https: //doi.org/10.1111/bjir.12130

    work page doi:10.1111/bjir.12130 2016

  25. [25]

    Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation- augmentation paradox.Academy of Management Review, 46(1), 192-210.https://doi.org/ 10.5465/amr.2018.0072

    work page doi:10.5465/amr.2018.0072 2021

  26. [26]

    J., & Aguirre, T

    Manning, S. J., & Aguirre, T. (2026).How adaptable are American workers to AI-induced job displacement?(NBER Working Paper 34705).https://www.nber.org/papers/w34705

    work page 2026

  27. [27]

    (2025, May 16).AI labor displacement and the limits of worker retraining.https: //www.brookings.edu/articles/ai-labor-displacement-and-the-limits-of-worker-ret raining/

    Brookings. (2025, May 16).AI labor displacement and the limits of worker retraining.https: //www.brookings.edu/articles/ai-labor-displacement-and-the-limits-of-worker-ret raining/

    work page 2025

  28. [28]

    F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P

    Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts.Transactions of the Association for Computational Linguistics, 12, 157-173.https://direct.mit.edu/tacl/article/doi/10.1 162/tacl_a_00638/119630/Lost-in-the-Middle-How-Language-Models-Use-Long

    work page 2024

  29. [29]

    (2025).State of AI-assisted software development.https://dora.dev/dora-r eport-2025

    Google DORA. (2025).State of AI-assisted software development.https://dora.dev/dora-r eport-2025

    work page 2025

  30. [30]

    (2026a, January 8)

    Sonar. (2026a, January 8). Sonar data reveals critical verification gap in AI coding.https: //www.sonarsource.com/company/press-releases/sonar-data-reveals-critical-verif ication-gap-in-ai-coding/

    work page

  31. [31]

    (2024, November 18).Does GitHub Copilot improve code quality? Here’s what the data says

    GitHub. (2024, November 18).Does GitHub Copilot improve code quality? Here’s what the data says. The GitHub Blog.https://github.blog/news-insights/research/does-github-cop ilot-improve-code-quality-heres-what-the-data-says/

    work page 2024

  32. [32]

    (2025).Information technology: Agencies need to plan for modernizing critical decades-old legacy systems(GAO-25-107795).https://files.gao.gov/reports/GAO-25-107795/index

    GAO. (2025).Information technology: Agencies need to plan for modernizing critical decades-old legacy systems(GAO-25-107795).https://files.gao.gov/reports/GAO-25-107795/index. html

    work page 2025

  33. [33]

    Villalobos, P., Ho, A., Sevilla, J., Besiroglu, T., Heim, L., & Hobbhahn, M. (2024).Will we run out of data? Limits of LLM scaling based on human-generated data.Epoch AI.https: //epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-g enerated-data

    work page 2024

  34. [34]

    (2024).AI models collapse when trained on recursively generated data.Nature, 631, 755-759.h t t p s : //www.nature.com/articles/s41586-024-07566-y 18

    Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., & Anderson, R. (2024).AI models collapse when trained on recursively generated data.Nature, 631, 755-759.h t t p s : //www.nature.com/articles/s41586-024-07566-y 18

    work page 2024

  35. [35]

    (2025).Demystifying synthetic data in LLM pre-training: A sys- tematic study of scaling laws, benefits, and pitfalls

    Kang, F., Ardalani, N., Kuchnik, M., Emad, Y., Elhoushi, M., Sengupta, S., Li, S.-W., Raghaven- dra, R., Jia, R., & Wu, C.-J. (2025).Demystifying synthetic data in LLM pre-training: A sys- tematic study of scaling laws, benefits, and pitfalls. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing(pp. 10739-10758). Associ...

    work page 2025

  36. [36]

    (2025).The 2025 Foundation Model Transparency Index

    Wan, A., Klyman, K., Kapoor, S., Maslej, N., Longpre, S., Xiong, B., Liang, P., & Bommasani, R. (2025).The 2025 Foundation Model Transparency Index. Stanford Center for Research on Foundation Models.https://crfm.stanford.edu/fmti/December-2025/paper.pdf

    work page 2025

  37. [37]

    (2025, January 31)

    Reuters. (2025, January 31). Taiwan bans government agencies from using DeepSeek, citing security concerns.Taipei Times.https://www.taipeitimes.com/News/taiwan/archives/202 5/01/31/2003831128

    work page 2025

  38. [38]

    CNBC. (2025, May 16).Musk’s xAI says Grok’s ’white genocide’ posts resulted from change that violated ’core values’.https://www.cnbc.com/2025/05/15/musks-xai-grok-white-genocid e-posts-violated-core-values.html

    work page 2025

  39. [39]

    (2025, January 21).Microsoft and OpenAI evolve partnership to drive the next phase of AI

    Microsoft. (2025, January 21).Microsoft and OpenAI evolve partnership to drive the next phase of AI. The Official Microsoft Blog.https://blogs.microsoft.com/blog/2025/01/21/micros oft-and-openai-evolve-partnership-to-drive-the-next-phase-of-ai/

    work page 2025

  40. [40]

    (2025, September 22).Nvidia to invest up to $100 billion in OpenAI, linking two artificial intelligence titans

    Reuters. (2025, September 22).Nvidia to invest up to $100 billion in OpenAI, linking two artificial intelligence titans. Investing.com.https://www.investing.com/news/stock-marke t-news/nvidia-to-invest-100-billion-in-openai-4249616 19

    work page 2025