Pith. sign in

REVIEW 4 major objections 6 minor 213 references

Off-the-shelf generative AI raises students' unaided test scores by 0.27 SD immediately, and most of that gain still shows up one week later without AI.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 06:16 UTC pith:KZD7FDJ4

load-bearing objection Clean lab RCT: off-the-shelf AI raises unaided test scores ~0.27 SD immediately and one week later, with delayed essay gains that stick for augmentation users; external validity is the real limit, not internal identification. the 4 major comments →

arxiv 2607.08849 v1 pith:KZD7FDJ4 submitted 2026-07-09 econ.GN cs.HCq-fin.EC

Experimental Evidence on the Learning Impact of Generative AI

Zara Contractor , Germ\'an Reyes This is my paper
classification econ.GN cs.HCq-fin.EC
keywords generative AIstudent learningrandomized experimentknowledge retentionaugmentation versus automationessay qualityhigher educationhuman capital
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether giving college students free access to ordinary generative AI helps or hurts real learning, not just task output. In a randomized, proctored lab experiment, undergraduates spent up to 35 minutes learning an unfamiliar technical topic and writing an analytical essay, either with AI allowed or forbidden; they then took unaided knowledge tests and wrote unaided essays immediately and about a week later. AI access raised immediate test scores by 0.27 standard deviations, and roughly three-quarters of that gain remained one week later when no one had AI. Essay quality barely moved while AI was available, but unaided essays a week later improved in style and relevance to the prompt. Those delayed writing gains were concentrated among students who used AI as a tutor (augmentation) rather than as a ghostwriter (automation); automation users' short-run quality boosts vanished once AI was removed. Two mechanisms show up in the data: treated students reallocated time away from drafting and toward reading and searching, and they reported more enjoyment. The design speaks to the chatbots students actually use, not customized tutoring systems, and holds total learning time roughly fixed.

Core claim

Random assignment to off-the-shelf generative AI during a short learning-and-essay phase raises unaided knowledge-test scores by 0.27 SD immediately and by a similar amount about one week later without AI. Higher-order essay quality improves mainly after AI is removed, and those delayed gains are larger for students who use AI to explain concepts than for those who use it to generate text.

What carries the argument

The augmentation-versus-automation classification of ChatGPT conversation logs, which separates students who use AI as a tutor (explain, clarify, feedback) from those who use it to produce draft text. That split organizes both short-run versus retained effects and students' own mental models of how AI affects learning.

Load-bearing premise

That a fixed-time lab session with elite undergraduates on low-prior-knowledge topics, followed by one-week retention, identifies the learning effect students would get when they choose how long to study and often use AI to finish faster.

What would settle it

A field experiment in ordinary coursework where students choose total study time and AI access either raises or lowers total learning once time reallocation is allowed, or where one-week retention gains disappear when topics are already familiar and stakes are real grades.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper reports a proctored, in-person RCT with 211 Middlebury undergraduates who learn an unfamiliar technical topic and write an analytical essay under AI-allowed or AI-forbidden conditions, then take unaided knowledge tests and write unaided essays immediately and about one week later. Random assignment raises ChatGPT use by ~67 pp. The main finding is that AI access raises unaided test scores by 0.27 SD immediately (ITT 6.7 pp on a 56.3% control mean) and by a similar 0.27 SD one week later without AI (~76% of the immediate effect). Session One essays show more AI-detected text and little quality change; Session Two unaided essays improve mainly in writing style/clarity and relevance, with larger delayed quality gains among LLM-classified “augmentation” users than “automation” users. Mechanisms include a shift of time from drafting toward reading/searching and higher reported enjoyment. The paper also reports beliefs and open-ended causal narratives about AI and learning.

Significance. If the estimates hold, this is among the cleanest experimental answers to whether off-the-shelf generative AI builds durable academic human capital rather than only short-run task performance. Strengths include: random assignment with a large first stage; multi-method compliance monitoring; unaided immediate and one-week retention assessments; dual human and AI essay grading plus objective linguistic and AI-detection measures; and a transparent literature meta-comparison. The augmentation-versus-automation heterogeneity and the time-reallocation/enjoyment mechanisms are policy-relevant and map onto students’ own mental models. The design deliberately studies unrestricted chatbots against a no-AI counterfactual, which is more informative for real student use than many scaffolded-tutor designs. External validity to ordinary coursework with endogenous study time remains the main limit, which the authors already flag.

major comments (4)
  1. Table 7, Panel D and §4.4: AI access raises any integrity violation by 12.6 pp (p=0.005). The back-of-envelope that cheating can explain ~2.2 pp of the 6.7 pp Session One test ITT (~one-third) is load-bearing for interpreting the knowledge gains as learning rather than test-taking contamination. Please report the main test-score ITT/TOT for Sessions One and Two after excluding proctor-flagged and self-reported violators (and a joint “any violation” sample), and clarify whether Session Two retention survives that restriction. If the retention effect is robust, state that prominently; if not, revise the learning interpretation accordingly.
  2. Table 8 and §5.3: The automation/augmentation split is constructed from LLM labels of treated students’ ChatGPT logs and is endogenous among users (different prompting, time use, and AI-detection rates). The differential fade-out is informative as descriptive heterogeneity, but several passages read as if use mode is a causal treatment. Please reframe Table 8 as non-causal heterogeneity among treated users, report balance of baseline ability/AI experience across use types, and avoid language that implies random assignment of automation vs augmentation.
  3. Table 6, columns 4–6 and Abstract: Session Two overall quality is +0.31 points (0.20 SD, p=0.143); only writing style and relevance are significant. The abstract’s claim that essay quality “improves in style and relevance” is accurate, but the introduction and §5.2 sometimes elevate this to broader “higher-order skills.” Please align the main text with the dimension-level pattern, report multiple-testing-adjusted inference for the five dimensions (or pre-specify primary essay outcomes), and avoid treating the imprecise overall quality index as established.
  4. Conclusion and §2: The design holds total learning time roughly fixed (~33 minutes). The paper correctly notes that ordinary AI use often saves time. Because the central policy claim is about learning impact of AI access, please add a short quantitative discussion of how large a reduction in time-on-task would be needed to offset the 0.27 SD retention gain under alternative assumptions, so readers can map the lab ITT to settings with endogenous study time.
minor comments (6)
  1. Figure 5 / Appendix B.7: State more clearly which of the 22 literature estimates are ITT vs TOT and whether all use unassisted outcomes only, so the grand mean of 0.18 SD is interpretable.
  2. Table 4: Self-assessed knowledge is essentially flat while objective scores rise. A brief discussion of why subjective knowledge does not track the test gains would help (calibration, ceiling of the 0–10 scale, or different construct).
  3. Appendix Table A1 and take-up: White students are less likely to use AI among the treated. Given the first-stage is not universal, a short note on whether TOT is driven by particular subgroups would be useful.
  4. §3.3 / double-lasso: Report the selected controls for the main test-score specifications (or an appendix table) so readers can see what residual imbalance is being adjusted.
  5. Figure 8 and Appendix C: The narrative coding is interesting but long relative to the experimental contribution; consider moving more of the causal-graph material to the appendix and keeping one summary figure in the main text.
  6. Typos/clarity: “whereasautomation” missing space in the abstract; ensure consistent Session One/Two capitalization; check that N varies slightly across tables (essay missingness) are explained once in a note.

Circularity Check

0 steps flagged

No circularity: central claims are experimental ITTs from random assignment, not quantities forced by fitted parameters or self-definition.

full rationale

The paper’s load-bearing results are intent-to-treat (and 2SLS TOT) estimates of random assignment to off-the-shelf AI access on unaided knowledge tests and essays (Eq. 1; Tables 4–6, 8). The 0.27 SD immediate and retention effects are differences in measured outcomes between AI-allowed and AI-forbidden arms; they are not derived from a structural parameter fitted to the same outcomes, nor defined in terms of those outcomes. Augmentation vs. automation is an LLM classification of ChatGPT conversation logs used only for heterogeneity (Appendix B.6; Table 8), not an input that forces the main ITT. Mechanisms (time mix, enjoyment, integrity) and belief/narrative analyses are separately measured descriptive outcomes. Self-citations (Contractor and Reyes 2026 on campus adoption/usage) supply context only and do not underwrite identification. There is no self-definitional loop, fitted-input-as-prediction, uniqueness import, or renaming of a known result as a first-principles derivation. The design is self-contained against its own experimental benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

This is an empirical RCT, not a theory paper. Load-bearing premises are design and measurement choices rather than free physical constants or invented particles. The main claim rests on random assignment identifying the ITT of AI access, on tests/essays measuring learning, and on LLM labels capturing meaningful use types for heterogeneity.

free parameters (3)
  • Double-lasso control selection and strata fixed effects
    Precision and residual imbalance adjustment depend on the covariate pool and lasso selection; not a fitted structural constant, but a modeling choice that can move point estimates slightly.
  • LLM conversation classification thresholds/prompts for augmentation vs automation
    Heterogeneity results depend on Claude Opus labels of chat logs into Automation/Augmentation/Mixed/Other; different prompts or models could reassign users.
  • Essay quality aggregation (human average + AI grader average)
    Main essay outcomes average human Prolific graders and Claude scores on a 0–10 rubric; weights and grader selection affect quality estimates.
axioms (5)
  • domain assumption Random assignment to AI-allowed vs AI-forbidden identifies the causal effect of AI access (ITT) under SUTVA and no differential attrition.
    Standard RCT identification; balance and attrition checks support it (§3.2–3.3).
  • domain assumption Unaided multiple-choice tests and analytical essays measure factual/conceptual knowledge and higher-order skills relevant to learning.
    Outcome construct validity is assumed throughout §§3.4–5.
  • domain assumption One week without resources is a meaningful retention horizon for skill accumulation claims.
    Session Two is the persistence test; longer horizons are not observed (§2.2, §5).
  • ad hoc to paper LLM labels of ChatGPT logs into augmentation vs automation recover economically meaningful use modes.
    Used for Table 8 heterogeneity; validated with prompt patterns, time use, and AI-detection rates (§5.3, App. B.6).
  • standard math Linear models with heteroskedasticity-robust SEs and double-lasso controls recover average treatment effects of interest.
    Equation (1) and 2SLS TOT (§3.3).
invented entities (1)
  • Augmentation vs automation user types (student-level) independent evidence
    purpose: Partition treated AI users to explain differential persistence of learning and essay quality.
    Operational taxonomy built from conversation logs; related to prior automation/augmentation language but newly applied as experimental moderators here.

pith-pipeline@v1.1.0-grok45 · 55538 in / 3170 out tokens · 33655 ms · 2026-07-13T06:16:36.636518+00:00 · methodology

0 comments
read the original abstract

We study how generative AI affects student learning in a randomized experiment. In proctored, in-person sessions, undergraduates learn about an unfamiliar topic and write an analytical essay with or without access to off-the-shelf generative AI, then complete unaided assessments immediately and one week later. We measure learning with knowledge tests (factual and conceptual understanding) and open-ended essays (higher-order skills). AI access raises immediate test scores by 0.27 standard deviations. These gains persist one week later. Essay quality, by contrast, changes little while students have AI access but improves in style and relevance one week later, when students write unaided. These delayed gains are larger among augmentation users-who use AI to explain concepts rather than generate text-whereas automation users' short-run quality gains vanish once AI is removed. We find evidence for two mechanisms behind the learning gains: students shift time away from drafting text and toward reading and searching for information, and they report greater learning enjoyment.

Figures

Figures reproduced from arXiv: 2607.08849 by Germ\'an Reyes, Zara Contractor.

Figure 1
Figure 1. Figure 1: Experimental Sessions Timelines Panel A. Session One timeline 0 min 60 min Welcome & Instructions Baseline Test Learning Phase (35 mins max) Post-Learning Survey Post-Learning Test Panel B. Session Two timeline 0 min 45 min Welcome & Instructions Endline Test (10 Questions) Essay (20 mins max) Randomized order Exit Survey Notes: This figure shows the timeline of the two experimental sessions. Panel A shows… view at source ↗
Figure 2
Figure 2. Figure 2: The Impact of AI Access on Generative AI Usage During the Learning Phase [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Types of Generative AI Use During the Learning Phase [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of Test Performance by Treatment Group [PITH_FULL_IMAGE:figures/full_fig_p030_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect Sizes Across AI-and-Learning Experiments [PITH_FULL_IMAGE:figures/full_fig_p031_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effects of AI Access on Essay Quality and Linguistic Features [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Actual and Perceived Treatment Effects on Test Performance [PITH_FULL_IMAGE:figures/full_fig_p033_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The Average Narrative About AI’s Effect on Learning [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗

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

213 extracted references · 2 canonical work pages

  1. [1]

    Review of Economic Studies , year =

    Andre, Peter and Haaland, Ingar and Roth, Christopher and Wiederholt, Mirko and Wohlfart, Johannes , title =. Review of Economic Studies , year =

  2. [2]

    2024 , institution =

    Emi, Bradley and Spero, Max , title =. 2024 , institution =. 2402.14873 , archiveprefix =

    Pith/arXiv arXiv 2024

  3. [3]

    2025 , month =

    Masrour, Elyas , title =. 2025 , month =

    2025

  4. [4]

    and Spero, Max , title =

    Masrour, Elyas and Emi, Bradley N. and Spero, Max , title =. Proceedings of the 1st Workshop on Detecting AI Generated Content (GenAIDetect), COLING , pages =. 2025 , url =

    2025

  5. [5]

    International Conference on Learning Representations (ICLR) , year =

    Thai, Katherine and Emi, Bradley and Masrour, Elyas and Iyyer, Mohit , title =. International Conference on Learning Representations (ICLR) , year =

  6. [6]

    2025 , url =

    Jabarian, Brian and Imas, Alex , title =. 2025 , url =

    2025

  7. [7]

    Journal of Economic Perspectives , volume=

    Automation and New Tasks: How Technology Displaces and Reinstates Labor , author=. Journal of Economic Perspectives , volume=

  8. [8]

    Science , volume=

    What Can Machine Learning Do? Workforce Implications , author=. Science , volume=

  9. [9]

    2025 , month =

    Becker, Joel and Rush, Nate and Barnes, Beth and Rein, David , title =. 2025 , month =. 2507.09089 , archiveprefix =

    Pith/arXiv arXiv 2025

  10. [10]

    2025 , month =

    Building an. 2025 , month =

    2025

  11. [11]

    and Hitzig, Zo\"e and Ong, Christopher and Shan, Carl Yan and Wadman, Kevin , title =

    Chatterji, Aaron and Cunningham, Thomas and Deming, David J. and Hitzig, Zo\"e and Ong, Christopher and Shan, Carl Yan and Wadman, Kevin , title =. 2025 , type =

    2025

  12. [12]

    The Generative

    Str. The Generative. 2026 , url =

    2026

  13. [13]

    VoxDevLit , volume =

    Education Technology , author =. VoxDevLit , volume =

  14. [14]

    2025 , month = aug, howpublished =

    Narayanan, Arvind , title =. 2025 , month = aug, howpublished =

    2025

  15. [15]

    2025 , doi =

    Kestin, Greg and Miller, Kelly and Klales, Anna and Milbourne, Timothy and Ponti, Gregorio , journal =. 2025 , doi =

    2025

  16. [16]

    The Impact of Generative

    Lee, Hao-Ping (Hank) and Sarkar, Advait and Tankelevitch, Lev and Drosos, Ian and Rintel, Sean and Banks, Richard and Wilson, Nicholas , booktitle =. The Impact of Generative. 2025 , publisher =

    2025

  17. [17]

    Reading Between the Lines: Modeling User Behavior and Costs in

    Mozannar, Hussein and Bansal, Gagan and Fourney, Adam and Horvitz, Eric , booktitle =. Reading Between the Lines: Modeling User Behavior and Costs in. 2024 , publisher =

    2024

  18. [18]

    2024 , eprint =

    Empirical evidence of large language model's influence on human spoken communication , author =. 2024 , eprint =

    2024

  19. [19]

    2025 , eprint =

    Ammari, Tawfiq and Chen, Meilun and Zaman, S M Mehedi and Garimella, Kiran , title =. 2025 , eprint =

    2025

  20. [20]

    The New Yorker , year =

    Hsu, Hua , title =. The New Yorker , year =

  21. [21]

    , title =

    Walsh, James D. , title =. New York Magazine , year =

  22. [22]

    2025 , url =

    Kunal Handa and Drew Bent and Alex Tamkin and Miles McCain and Esin Durmus and Michael Stern and Mike Schiraldi and Saffron Huang and Stuart Ritchie and Steven Syverud and Kamya Jagadish and Margaret Vo and Matt Bell and Deep Ganguli , title =. 2025 , url =

    2025

  23. [23]

    2026 , url =

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

    2026

  24. [24]

    The Quarterly Journal of Economics , volume=

    The (perceived) returns to education and the demand for schooling , author=. The Quarterly Journal of Economics , volume=. 2010 , publisher=

    2010

  25. [25]

    The Review of Economic Studies , volume=

    Determinants of college major choice: Identification using an information experiment , author=. The Review of Economic Studies , volume=. 2015 , publisher=

    2015

  26. [26]

    Generative

    Contractor, Zara and Reyes, Germ. Generative

  27. [27]

    Handbook of the Economics of Education , editor =

    Technology and Education: Computers, Software, and the Internet , author =. Handbook of the Economics of Education , editor =

  28. [28]

    Journal of Economic Literature , volume=

    Upgrading education with technology: Insights from experimental research , author=. Journal of Economic Literature , volume=. 2020 , publisher=

    2020

  29. [29]

    Journal of Memory and Language , volume=

    How many words do we read per minute? A review and meta-analysis of reading rate , author=. Journal of Memory and Language , volume=. 2019 , publisher=

    2019

  30. [30]

    Journal of Reading , volume=

    Reading rate: Theory, research, and practical implications , author=. Journal of Reading , volume=. 1992 , publisher=

    1992

  31. [31]

    Handbook 1: Cognitive domain , author=

    Taxonomy of educational objectives: The classification of educational goals. Handbook 1: Cognitive domain , author=. 1956 , publisher=

    1956

  32. [32]

    2023 , month =

    Nam, Jane , title =. 2023 , month =

    2023

  33. [33]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , title =. Advances in Neural Information Processing Systems , volume =

  34. [34]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages =

    Chiang, Cheng-Han and Lee, Hung-yi , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages =. 2023 , publisher =

    2023

  35. [35]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =

    Liu, Yang and Iter, Dan and Xu, Yichong and Wang, Shuohang and Xu, Ruochen and Zhu, Chenguang , title =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =. 2023 , publisher =

    2023

  36. [36]

    , title =

    Carter, Susan Payne and Greenberg, Kyle and Walker, Michael S. , title =. Economics of Education Review , year =

  37. [37]

    The Quarterly Journal of Economics , year =

    Malamud, Ofer and Pop-Eleches, Cristian , title =. The Quarterly Journal of Economics , year =

  38. [38]

    Technology and Child Development: Evidence from the

    Cristia, Julian and Ibarrar. Technology and Child Development: Evidence from the. American Economic Journal: Applied Economics , year =

  39. [39]

    and Rush, Mark and Yin, Lu , title =

    Figlio, David N. and Rush, Mark and Yin, Lu , title =. Journal of Labor Economics , year =

  40. [40]

    and Fox, Lindsay and Loeb, Susanna and Taylor, Eric S

    Bettinger, Eric P. and Fox, Lindsay and Loeb, Susanna and Taylor, Eric S. , title =. American Economic Review , year =

  41. [41]

    and Ladd, Helen F

    Vigdor, Jacob L. and Ladd, Helen F. and Martinez, Erika , title =. Economic Inquiry , year =

  42. [42]

    and Goodman, Sarena and Smith, Jonathan , title =

    Dettling, Lisa J. and Goodman, Sarena and Smith, Jonathan , title =. The Review of Economics and Statistics , year =

  43. [43]

    , title =

    Caldwell, Jane E. , title =. CBE---Life Sciences Education , year =

  44. [44]

    Education and Information Technologies , year =

    Lewin, Cathy and Somekh, Bridget and Steadman, Stephen , title =. Education and Information Technologies , year =

  45. [45]

    Journal of the European Economic Association , volume =

    Expertise , author =. Journal of the European Economic Association , volume =

  46. [46]

    Generative

    Bastani, Hamsa and Bastani, Osbert and Sungu, Alp and Ge, Haosen and Kabakc. Generative. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2422633122 , note =

    doi:10.1073/pnas.2422633122

  47. [47]

    Belloni, Alexandre and Chernozhukov, Victor and Hansen, Christian , year = 2014, month = may, journal =. High-

    2014

  48. [48]

    , year = 2026, journal =

    Bick, Alexander and Blandin, Adam and Deming, David J. , year = 2026, journal =. The

    2026

  49. [49]

    Generative

    Brynjolfsson, Erik and Li, Danielle and Raymond, Lindsey , year = 2025, month = may, journal =. Generative

    2025

  50. [50]

    Endoscopist

    Budzy. Endoscopist. The Lancet Gastroenterology & Hepatology , volume =

  51. [51]

    Cui, Zheyuan (Kevin) and Demirer, Mert and Jaffe, Sonia and Musolff, Leon and Peng, Sida and Salz, Tobias , year = 2026, journal =. The

    2026

  52. [52]

    Writing Code vs

    Demirer, Mert and Musolff, Leon and Yang, Liyuan , institution =. Writing Code vs. Shipping Code: Productivity Effects Across Generations of

  53. [53]

    Dell'Acqua, Fabrizio and McFowland, Edward and Mollick, Ethan R. and. Navigating the. Organization Science , doi =

  54. [54]

    and Hauser, Oliver P

    Doshi, Anil R. and Hauser, Oliver P. , journal =. Generative

  55. [55]

    Meincke, Lennart and Nave, Gideon and Terwiesch, Christian , journal =

  56. [56]

    and Kushlev, Kostadin , journal =

    Moon, Kibum and Green, Adam E. and Kushlev, Kostadin , journal =. Homogenizing Effect of Large Language Models (

  57. [57]

    The Creative Link Between Words and Ideas Is Weakening in the

    Moon, Kibum and Kushlev, Kostadin and Bank, Andrew and. The Creative Link Between Words and Ideas Is Weakening in the. doi:10.31234/osf.io/jsz58_v6 , url =

    work page doi:10.31234/osf.io/jsz58_v6

  58. [58]

    Minnesota Law Review , volume =

    Lawyering in the Age of Artificial Intelligence , author =. Minnesota Law Review , volume =

  59. [59]

    Educational Evaluation and Policy Analysis , volume =

    How Big Are Effect Sizes in International Education Studies? , author =. Educational Evaluation and Policy Analysis , volume =

  60. [60]

    Educational Researcher , volume =

    Interpreting Effect Sizes of Education Interventions , author =. Educational Researcher , volume =

  61. [61]

    and Lavy, Victor , journal =

    Angrist, Joshua D. and Lavy, Victor , journal =. Using

  62. [62]

    Kirabo and Mackevicius, Claire L

    Jackson, C. Kirabo and Mackevicius, Claire L. , journal =. What Impacts Can We Expect from School Spending Policy?

  63. [63]

    The Promise of Tutoring for

    Nickow, Andre and Oreopoulos, Philip and Quan, Vincent , journal =. The Promise of Tutoring for

  64. [64]

    and Rockoff, Jonah E

    Chetty, Raj and Friedman, John N. and Rockoff, Jonah E. , journal =. Measuring the Impacts of Teachers

  65. [65]

    Economics of Education Review , volume =

    The Economic Value of Higher Teacher Quality , author =. Economics of Education Review , volume =

  66. [66]

    Baird, Matthew and Carpanelli, Mar and Xu, Brian and Xu, Kevin , journal =. Firms'

  67. [67]

    Academy of Management Journal , volume =

    When and How Artificial Intelligence Augments Employee Creativity , author =. Academy of Management Journal , volume =

  68. [68]

    , institution =

    Dell'Acqua, Fabrizio and Ayoubi, Charles and Lifshitz, Hila and Sadun, Raffaella and Mollick, Ethan and Mollick, Lilach and Han, Yi and Goldman, Jeff and Nair, Hari and Taub, Stewart and Lakhani, Karim R. , institution =. The Cybernetic Teammate: A Field Experiment on Generative

  69. [69]

    Does Generative

    Cruces, Guillermo and. Does Generative

  70. [70]

    and Sting, Fabian J

    Lehmann, Matthias and Cornelius, Philipp B. and Sting, Fabian J. , year = 2025, month = mar, number =

    2025

  71. [71]

    Experimental

    Noy, Shakked and Zhang, Whitney , year = 2023, month = jul, journal =. Experimental

    2023

  72. [72]

    Peng, Sida and Kalliamvakou, Eirini and Cihon, Peter and Demirer, Mert , year = 2023, month = feb, number =. The. 2302.06590 , doi =

    Pith/arXiv arXiv 2023

  73. [73]

    Rav. Higher. 2025 , month = feb, journal =

    2025

  74. [74]

    Perceptions and

    St. Perceptions and. Computers and Education: Artificial Intelligence , volume =

  75. [75]

    Harvard undergraduate survey on generative

    Hirabayashi, Shikoh and Jain, Rishab and Jurkovi. Harvard undergraduate survey on generative

  76. [76]

    , journal =

    Goldsmith-Pinkham, Paul and Tan, Chenhao and Zentefis, Alexander K. , journal =. Human-

  77. [77]

    Kanazawa, Kyogo and Kawaguchi, Daiji and Shigeoka, Hitoshi and Watanabe, Yasutora , journal =

  78. [78]

    Psychological Methods , volume =

    A General Approach to Causal Mediation Analysis , author =. Psychological Methods , volume =

  79. [79]

    Social Sciences & Humanities Open , volume =

    Barcaui, Andr. Social Sciences & Humanities Open , volume =

  80. [80]

    2412.16429 , archivePrefix =

    Pith/arXiv arXiv

Showing first 80 references.