pith. sign in

arxiv: 2605.16283 · v2 · pith:XG4YTQCInew · submitted 2026-04-12 · 💻 cs.CY · cs.AI

Can the Recovery Mechanism Survive AI? Skill Formation, Labor, and What Current Measurement Misses

Aysa Xuemo Fan This is my paper

Pith reviewed 2026-05-25 06:33 UTC · model grok-4.3

classification 💻 cs.CY cs.AI
keywords generative AIskill formationlabor economicscognitive taxonomyproductive struggleevidence gapseducation technologyepistemic agency
0
0 comments X

The pith

Generative AI may break the historical cycle in which education raised cognitive ceilings to adapt to technologies that displace workers, because it now operates at the top of that ceiling.

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

The paper argues that past technologies displaced workers but societies recovered by using education to expand what people could do. Generative AI differs because it can handle the complex cognitive tasks that education previously prepared workers for. The author develops a stock-versus-flow framework to show that current economic data reflects worker augmentation while education data shows strain on the pipeline that forms new skills. A gap analysis of existing studies finds that knowledge dimensions of cognition go unmeasured, small experiments show AI boosts performance without boosting learning, and no research connects professional and student settings. An extended taxonomy of judgment under uncertainty, epistemic identity, and epistemic agency is applied to cases to separate AI uses that preserve learning from those that erode it.

Core claim

AI's societal risk lies not in replacing teachers but in eliminating the productive struggle through which the next generation's capacity forms. This follows from the stock-versus-flow divergence between labor-market augmentation and developmental-pipeline strain, from the systematic absence of learning-outcome measures across studies, and from the taxonomy distinguishing interaction patterns that maintain versus remove the epistemic work required for skill formation.

What carries the argument

The stock-versus-flow framework, which separates current-worker data from the developmental pipeline that produces future workers, together with an extended cognitive taxonomy of judgment under uncertainty, epistemic identity, and epistemic agency.

Load-bearing premise

The three small studies that measured learning outcomes and the reanalysis showing performance gains without learning gains are representative of the broader evidence base and that no studies exist that connect professional and student populations.

What would settle it

A study with several hundred or more participants in a student population that directly measures knowledge retention or epistemic agency and finds sustained learning gains from AI use comparable to non-AI conditions.

Figures

Figures reproduced from arXiv: 2605.16283 by Aysa Xuemo Fan.

Figure 1
Figure 1. Figure 1: The historical pattern: each technology automated a cognitive floor, and education retreated to the floor above. Gen [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The inverted Bloom’s Taxonomy. Students dele [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Anderson and Krathwohl’s (2001) two￾dimensional taxonomy. Experts operate in the high￾knowledge, high-process zone (bottom-right). Students using AI may operate at high cognitive process levels but low knowledge levels (top-right), performing above their knowledge base. The Uncontrollable Adoption Even if schools resist AI (and many try), the adoption is unstoppable. Schools can restrict AI in classrooms. … view at source ↗
Figure 3
Figure 3. Figure 3: Original analysis of Anthropic Economic Index public data (Handa et al. 2025c). (a) Mean AI exposure by O*NET [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Anderson and Krathwohl’s (2001) two￾dimensional taxonomy. Experts operate in the high￾knowledge, high-process zone (bottom-right). Students using AI may operate at high cognitive process levels but low knowledge levels (top-right), performing above their knowledge base. even copying from a textbook required navigating the ma￾terial and connecting fragments. These frictions were them￾selves learning mechani… view at source ↗
Figure 5
Figure 5. Figure 5: Six AI interaction patterns from Shen and Tamkin [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

Throughout the modern era, when new technologies displaced workers, societies adapted through the same mechanism: education raised the cognitive ceiling, producing workers capable of tasks machines could not yet reach. Generative AI may be the first technology to break this cycle, because it now operates at the top of that ceiling. Drawing on labor economics, deployment data from millions of AI conversations across multiple platforms, original reanalysis of two public datasets, and skill-formation experiments, this paper develops three contributions. First, a stock-versus-flow framework showing that economic data and education data tell divergent stories about the same technology: augmentation dominates current workers, but the developmental pipeline producing the next generation is under strain. Second, a systematic gap analysis of the evidence base, revealing that the knowledge dimension of cognition is unmeasured across all major studies, that the three studies measuring learning outcomes (each $n < 200$) consistently find AI improves performance without improving learning ($d = 1.21$ in our cross-platform reanalysis), and that no study bridges professional and student populations. Third, an extended cognitive taxonomy (judgment under uncertainty, epistemic identity, and epistemic agency) applied to three cases from the evidence to distinguish AI interaction patterns that preserve learning from structurally similar ones that erode it. The paper argues that AI's societal risk lies not in replacing teachers but in eliminating the productive struggle through which the next generation's capacity forms, and proposes a research and design agenda targeting what current measurement systems miss.

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 claims that generative AI breaks the historical education-adaptation cycle to technological displacement because it operates at the cognitive ceiling. It advances three contributions: (1) a stock-versus-flow framework showing augmentation for current workers but strain on the developmental pipeline, based on labor economics, millions of AI conversation logs, and public dataset reanalysis; (2) a gap analysis concluding that the knowledge dimension is unmeasured in all major studies, that three learning-outcome studies (each n<200) find performance gains without learning gains (d=1.21 in cross-platform reanalysis), and that no study bridges professional and student populations; (3) an extended cognitive taxonomy (judgment under uncertainty, epistemic identity, epistemic agency) applied to three cases to distinguish learning-preserving from eroding AI interactions. The central argument is that AI's risk lies in eliminating productive struggle for the next generation's capacity, with a proposed research and design agenda.

Significance. If the central claims hold, the work would be significant for labor economics, education technology, and AI policy by synthesizing large-scale deployment data with skill-formation evidence to identify measurement gaps and pipeline risks. The use of millions of conversation logs and reanalysis of public datasets is a strength for grounding the stock-flow divergence; the taxonomy offers a novel lens for distinguishing interaction patterns. However, the interpretive synthesis from small-n studies limits the strength of the generalization to societal risk.

major comments (3)
  1. [Gap analysis section] Gap analysis (abstract and associated section): the claim that the three learning-outcome studies (each n<200) and cross-platform reanalysis (d=1.21) establish systematic evidence gaps and that no study bridges professional/student populations is load-bearing for the pipeline-strain argument, yet the manuscript provides no explicit search protocol, inclusion criteria, or power analysis; small samples raise risks of selection effects and low power that must be addressed to support the absence-of-bridging-studies conclusion.
  2. [Reanalysis subsection] Cross-platform reanalysis (abstract): the d=1.21 effect size for performance without learning gains is presented without reported confidence intervals, dataset identifiers, or robustness checks against confounding variables such as task difficulty or participant selection; this detail is required to evaluate whether the finding generalizes beyond the cited studies.
  3. [Taxonomy application to cases] Extended cognitive taxonomy (third contribution): the new constructs (judgment under uncertainty, epistemic identity, epistemic agency) are applied to cases without operational definitions, measurement items, or differentiation from existing constructs in cognitive psychology or education research; this weakens the distinction between preserving and eroding interaction patterns.
minor comments (2)
  1. [Abstract] Abstract: clarify the relationship between the 'original reanalysis of two public datasets' and the 'three studies measuring learning outcomes' to avoid potential overlap or double-counting.
  2. [Discussion or agenda section] The paper flags its own evidence gaps; this self-acknowledgment is a strength but should be paired with explicit recommendations for the minimal sample sizes or study designs needed to close them.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for these constructive comments, which identify opportunities to enhance transparency and precision. We address each major point below, indicating revisions where they strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [Gap analysis section] Gap analysis (abstract and associated section): the claim that the three learning-outcome studies (each n<200) and cross-platform reanalysis (d=1.21) establish systematic evidence gaps and that no study bridges professional/student populations is load-bearing for the pipeline-strain argument, yet the manuscript provides no explicit search protocol, inclusion criteria, or power analysis; small samples raise risks of selection effects and low power that must be addressed to support the absence-of-bridging-studies conclusion.

    Authors: We agree that explicit methodological details will improve rigor. The revised manuscript will add a subsection describing the targeted literature search (key terms, databases, and time frame focused on AI performance vs. learning outcome studies), inclusion criteria (empirical studies reporting both metrics), and a note on sample-size limitations including risks of selection bias and low power. We will also qualify the 'no bridging studies' statement as based on our review of prominent works rather than an exhaustive systematic review. These additions address the concern while preserving the observation that integrative professional-student studies remain absent from the current evidence base. revision: partial

  2. Referee: [Reanalysis subsection] Cross-platform reanalysis (abstract): the d=1.21 effect size for performance without learning gains is presented without reported confidence intervals, dataset identifiers, or robustness checks against confounding variables such as task difficulty or participant selection; this detail is required to evaluate whether the finding generalizes beyond the cited studies.

    Authors: We will expand the reanalysis subsection to report the 95% confidence interval around d=1.21, provide full identifiers and access details for the two public datasets, and include robustness checks (e.g., leave-one-out sensitivity and controls for task difficulty and selection). The effect size was obtained via fixed-effects aggregation of the three studies; adding these elements will allow readers to assess generalizability directly. revision: yes

  3. Referee: [Taxonomy application to cases] Extended cognitive taxonomy (third contribution): the new constructs (judgment under uncertainty, epistemic identity, epistemic agency) are applied to cases without operational definitions, measurement items, or differentiation from existing constructs in cognitive psychology or education research; this weakens the distinction between preserving and eroding interaction patterns.

    Authors: The taxonomy functions as a conceptual lens rather than a measurement instrument in this paper. We will insert a brief differentiation paragraph situating the constructs relative to established notions such as metacognition and epistemic cognition, while clarifying that the three case applications are illustrative. Operational definitions and items are explicitly listed as future research needs in the proposed agenda; the current contribution is the distinction they enable between interaction patterns. revision: partial

Circularity Check

0 steps flagged

No significant circularity; contributions rest on external data reanalysis and evidence review

full rationale

The paper presents three contributions derived from deployment data across platforms, reanalysis of two public datasets, and skill-formation experiments. The stock-versus-flow framework, systematic gap analysis (including the d=1.21 effect size from cross-platform reanalysis of existing studies), and application of the extended cognitive taxonomy to three cases are all grounded in external sources rather than self-defined quantities or fitted parameters renamed as predictions. No equations, self-citation chains, uniqueness theorems, or ansatzes imported from prior author work are described. The central claim about AI breaking the education-adaptation cycle is supported by the gap analysis of the evidence base, which draws on independent studies rather than reducing to the paper's own inputs by construction. This is a self-contained argument against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 3 invented entities

Based on abstract only; the paper relies on domain assumptions from labor economics about historical adaptation mechanisms and introduces a new taxonomy without providing independent evidence for its components.

axioms (2)
  • domain assumption Throughout the modern era, when new technologies displaced workers, societies adapted through the same mechanism: education raised the cognitive ceiling.
    Invoked in the opening paragraph as the baseline recovery mechanism that AI may break.
  • domain assumption The knowledge dimension of cognition is unmeasured across all major studies.
    Stated as part of the systematic gap analysis in the second contribution.
invented entities (3)
  • judgment under uncertainty no independent evidence
    purpose: Component of extended cognitive taxonomy to distinguish AI interaction patterns that preserve versus erode learning.
    Introduced in the third contribution and applied to three cases from the evidence.
  • epistemic identity no independent evidence
    purpose: Component of extended cognitive taxonomy to distinguish AI interaction patterns that preserve versus erode learning.
    Introduced in the third contribution and applied to three cases from the evidence.
  • epistemic agency no independent evidence
    purpose: Component of extended cognitive taxonomy to distinguish AI interaction patterns that preserve versus erode learning.
    Introduced in the third contribution and applied to three cases from the evidence.

pith-pipeline@v0.9.0 · 5796 in / 1573 out tokens · 35499 ms · 2026-05-25T06:33:27.564404+00:00 · methodology

Review history (2 revisions) →

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

48 extracted references · 48 canonical work pages · 1 internal anchor

  1. [1]

    Computers and Education: Artificial Intelligence , pages=

    Directive, metacognitive, or a blend of both? A comparison of AI-generated feedback types on student engagement, confidence, and outcomes , author=. Computers and Education: Artificial Intelligence , pages=. 2026 , publisher=

    work page 2026

  2. [2]

    2001 , publisher=

    A taxonomy for learning, teaching, and assessing: A revision of Bloom's taxonomy of educational objectives: complete edition , author=. 2001 , publisher=

    work page 2001

  3. [3]

    Anthropic

    Anthropic Education Report: How Educators Use Claude , author=. Anthropic. Available online: https://www. anthropic. com/news/anthropic-education-report-how-educators-use-claude (accessed on 15 December 2025) , year=

    work page 2025

  4. [4]

    Anthropic , year =

    Handa, Kunal and Bent, Drew and Tamkin, Alex and McCain, Miles and Durmus, Esin and Stern, Michael and Schiraldi, Mike and Huang, Saffron and Ritchie, Stuart and Syverud, Steven and others , title =. Anthropic , year =

    work page

  5. [5]

    2025 , url =

    Huang, Saffron and Seethor, Bryan and Durmus, Esin and Handa, Kunal and McCain, Miles and Stern, Michael and Ganguli, Deep , title =. 2025 , url =

    work page 2025

  6. [6]

    2025 , url =

    Handa, Kunal and Stern, Michael and Huang, Saffron and Hong, Jerry and Durmus, Esin and McCain, Miles and Yun, Grace and Alt, AJ and Millar, Thomas and Tamkin, Alex and Leibrock, Jane and Ritchie, Stuart and Ganguli, Deep , title =. 2025 , url =

    work page 2025

  7. [7]

    2026 , url =

    Swanson, Kristen and Bent, Drew and Ludwig, Zoe and Dakan, Rick and Feller, Joe , title =. 2026 , url =

    work page 2026

  8. [8]

    2026 , url =

    Appel, Ruth and Massenkoff, Maxim and McCrory, Peter and McCain, Miles and Heller, Ryan and Neylon, Tyler and Tamkin, Alex , title =. 2026 , url =

    work page 2026

  9. [9]

    and Bjork, Robert A

    Bjork, Elizabeth L. and Bjork, Robert A. and others , title =. Psychology and the Real World: Essays Illustrating Fundamental Contributions to Society , volume =

    work page

  10. [10]

    Journal of University Teaching and Learning Practice , volume =

    Corbin, Thomas and Bearman, Margaret and Boud, David and Crawford, Nicole and Dawson, Phillip and Fawns, Tim and Henderson, Michael and Lodge, Jason and Luo, Jiahui and Matthews, Kelly and Nicola-Richmond, Kelli and Nieminen, Juuso Henrik and Pepperell, Nicole and Swiecki, Zachari and Tai, Joanna and Walton, Jack , title =. Journal of University Teaching ...

    work page 2025

  11. [11]

    Impact of

    Darvishi, Ali and Khosravi, Hassan and Sadiq, Shazia and Ga. Impact of. Computers & Education , volume =. 2024 , publisher =

    work page 2024

  12. [12]

    Computers & Education , volume =

    Deng, Ruiqi and Jiang, Maoli and Yu, Xinlu and Lu, Yuyan and Liu, Shasha , title =. Computers & Education , volume =. 2025 , publisher =

    work page 2025

  13. [13]

    1938 , doi =

    Dewey, John , title =. 1938 , doi =

    work page 1938

  14. [14]

    2019 , isbn =

    Frey, Carl Benedikt , title =. 2019 , isbn =

    work page 2019

  15. [15]

    , title =

    Goldin, Claudia and Katz, Lawrence F. , title =. 2008 , isbn =

    work page 2008

  16. [16]

    Journal of Marketing Education , pages=

    Generative AI’s impact on critical thinking: Revisiting Bloom’s taxonomy , author=. Journal of Marketing Education , pages=. 2024 , publisher=

    work page 2024

  17. [17]

    2016 , publisher=

    The rise and fall of American growth: The US standard of living since the civil war , author=. 2016 , publisher=

    work page 2016

  18. [18]

    2016 , publisher=

    Handbook of epistemic cognition , author=. 2016 , publisher=

    work page 2016

  19. [19]

    arXiv preprint arXiv:2503.04761 , year=

    Which economic tasks are performed with ai? evidence from millions of claude conversations , author=. arXiv preprint arXiv:2503.04761 , year=

    work page arXiv

  20. [20]

    British Journal of Educational Technology , year=

    A systematic review of generative AI in education: Empirical insights from a human--AI interaction perspective , author=. British Journal of Educational Technology , year=

    work page

  21. [21]

    Artificial Intelligence Applications in K-12 , pages=

    Assessing in the age of AI: is critical thinking the answer? , author=. Artificial Intelligence Applications in K-12 , pages=. 2024 , publisher=

    work page 2024

  22. [22]

    2026 , url =

    Maxim Massenkoff and Eva Lyubich and Peter McCrory and Ruth Appel and Ryan Heller , title =. 2026 , url =

    work page 2026

  23. [23]

    arXiv preprint arXiv:2306.10052 , year=

    Assigning AI: Seven approaches for students, with prompts , author=. arXiv preprint arXiv:2306.10052 , year=

    work page arXiv

  24. [24]

    Human factors , volume=

    Complacency and bias in human use of automation: An attentional integration , author=. Human factors , volume=. 2010 , publisher=

    work page 2010

  25. [25]

    2026 , url =

    Schwartz, Matthew , title =. 2026 , url =

    work page 2026

  26. [26]

    Journal of consumer psychology , volume=

    Metacognitive experiences in consumer judgment and decision making , author=. Journal of consumer psychology , volume=. 2004 , publisher=

    work page 2004

  27. [27]

    Perspectives on Psychological Science , volume=

    Learning versus performance: An integrative review , author=. Perspectives on Psychological Science , volume=. 2015 , publisher=

    work page 2015

  28. [28]

    Computers in Human Behavior , volume=

    Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry , author=. Computers in Human Behavior , volume=. 2024 , publisher=

    work page 2024

  29. [29]

    Yan, Lixiang and Greiff, Samuel and Lodge, Jason M. and Ga. Distinguishing performance gains from learning when using generative. Nature Reviews Psychology , volume =. 2025 , publisher =

    work page 2025

  30. [30]

    Promises and challenges of generative artificial intelligence for human learning , journal =

    Yan, Lixiang and Greiff, Samuel and Teuber, Ziwen and Ga. Promises and challenges of generative artificial intelligence for human learning , journal =. 2024 , doi =

    work page 2024

  31. [31]

    , title =

    Zhai, Chunpeng and Wibowo, Santoso and Li, Lily D. , title =. Smart Learning Environments , volume =. 2024 , publisher =

    work page 2024

  32. [32]

    International Review of Research in Open and Distributed Learning , volume =

    Jordan, Katy , title =. International Review of Research in Open and Distributed Learning , volume =. 2014 , doi =

    work page 2014

  33. [33]

    1989 , isbn =

    Curriculum and Evaluation Standards for School Mathematics , publisher =. 1989 , isbn =

    work page 1989

  34. [34]

    Polanyi, Michael , title =

    work page

  35. [35]

    Skinner, B. F. , title =. Science , volume =. 1958 , doi =

    work page 1958

  36. [36]

    1975 , isbn =

    Tinbergen, Jan , title =. 1975 , isbn =

    work page 1975

  37. [37]

    Proceedings of the Conference on Language Modeling (

    Huang, Saffron and Durmus, Esin and McCain, Miles and Handa, Kunal and Tamkin, Alex and Hong, Jerry and Stern, Michael and Somani, Arushi and Zhang, Xiuruo and Ganguli, Deep , title =. Proceedings of the Conference on Language Modeling (. 2025 , note =

    work page 2025

  38. [38]

    and Mueller, Jared and McEachen, William and Mitchell, Wes and Carter, Shan and Clark, Jack and Kaplan, Jared and Ganguli, Deep , title =

    Tamkin, Alex and McCain, Miles and Handa, Kunal and Durmus, Esin and Lovitt, Liane and Rathi, Ankur and Huang, Saffron and Mountfield, Alfred and Hong, Jerry and Ritchie, Stuart and Stern, Michael and Clarke, Brian and Goldberg, Landon and Sumers, Theodore R. and Mueller, Jared and McEachen, William and Mitchell, Wes and Carter, Shan and Clark, Jack and K...

    work page 2024

  39. [39]

    Beware of metacognitive laziness:

    Fan, Yizhou and Tang, Luzhen and Le, Huixiao and Shen, Kejie and Tan, Shufang and Zhao, Yueying and Shen, Yuan and Li, Xinyu and Ga. Beware of metacognitive laziness:. British Journal of Educational Technology , volume =. 2025 , publisher =

    work page 2025

  40. [40]

    2026 , month =

    Karamanis, Minas , title =. 2026 , month =

    work page 2026

  41. [41]

    arXiv preprint arXiv:2601.20245 , year =

    Shen, Judy Hanwen and Tamkin, Alex , title =. arXiv preprint arXiv:2601.20245 , year =

    work page arXiv

  42. [42]

    Navigating the Jagged Technological Frontier:

    Dell'Acqua, Fabrizio and McFowland III, Edward and Mollick, Ethan R and Lifshitz-Assaf, Hila and Kellogg, Katherine and Rajendran, Saran and Krayer, Lisa and Candelon, Fran. Navigating the Jagged Technological Frontier:. Harvard Business School Technology & Operations Mgt.\ Unit Working Paper , number =

    work page

  43. [43]

    The Quarterly Journal of Economics , volume =

    Brynjolfsson, Erik and Li, Danielle and Raymond, Lindsey , title =. The Quarterly Journal of Economics , volume =. 2025 , publisher =

    work page 2025

  44. [44]

    Proceedings of the Eighth

    Ericson, Petter and Carli, Rachele and Tucker, Jason and Dignum, Virginia , title =. Proceedings of the Eighth. 2025 , publisher =

    work page 2025

  45. [45]

    Fesler, Lily and Martinez, JP and Agnew, Chris and Loeb, Susanna , title =

    work page

  46. [46]

    arXiv preprint arXiv:2603.15245 , year =

    Kumar, Aakriti and Poungpeth, Nalin and Yang, Diyi and Lambert, Bruce and Groh, Matthew , title =. arXiv preprint arXiv:2603.15245 , year =

    work page arXiv

  47. [47]

    LLAMA LIMA: A Living Meta-Analysis on the Effects of Generative AI on Learning Mathematics

    Strohmaier, Anselm and B. arXiv preprint arXiv:2601.18685 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Proceedings of the 16th International Learning Analytics and Knowledge Conference (

    McNichols, Hunter and Ikram, Fareya and Lan, Andrew , title =. Proceedings of the 16th International Learning Analytics and Knowledge Conference (. 2026 , publisher =

    work page 2026