Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Brief training raises GenAI adoption and exam grades for law students; access alone does not.

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

2026-07-15 14:52 UTC pith:TJLQACCL

load-bearing objection Clean three-arm RCT isolating brief training under optional LLM access; main ITT results are real, but the adoption-vs-effectiveness mechanism claim is under-identified by the paper’s own bounds. the 3 major comments →

arxiv 2603.04982 v3 pith:TJLQACCL submitted 2026-03-05 cs.CY cs.AIcs.HC

Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis

Benjamin M. Chen , Hong Bao This is my paper
classification cs.CY cs.AIcs.HC
keywords generative AIuser traininglegal analysisissue-spottingadoptionprincipal stratificationproductivitylaw students
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 tests whether a short training session can turn optional generative AI access into better professional work. In a randomized mock contract-law exam, 164 students either had no large language model, optional access without guidance, or optional access plus a brief video and quiz on how to use the model for issue-spotting. Untrained access produced shorter answers, more case misstatements, and no score gain. Training raised reported adoption from 26% to 41% and lifted mean grade point by 0.27 relative to untrained users, while also improving accurate statement of legal rules. The authors argue that training mainly expands who chooses to use the tool rather than only improving those already inclined to use it, and that without instruction higher-ability users sit out while lower-ability users adopt unproductively. Realizing productivity gains therefore requires both access and instruction.

Core claim

On a timed issue-spotting exam, optional LLM access without training did not improve performance and was associated with shorter answers and more case misstatements. Adding a short training intervention raised self-reported use (41% versus 26%) and raised mean grade point by 0.27 relative to untrained LLM access, while reducing rule misstatements. Principal-stratification bounds under strict mean dominance place the adoption channel above the effectiveness channel, though intervals are wide.

What carries the argument

Principal stratification that partitions participants into always-users, never-users, and induced users, then partially identifies the adoption effect for induced users and the effectiveness effect for always-users under monotone treatment response, exclusion for never-users, support bounds, and mean dominance.

Load-bearing premise

The claim that training works mainly by inducing new users rather than by improving existing users rests on assumptions (no defiers, exclusion for never-users, and mean dominance) whose resulting bounds have wide overlapping confidence intervals.

What would settle it

A larger pre-registered experiment that measures pre-treatment ability independently, records actual tool logs rather than self-report, and finds either no adoption increase among high-ability students or no grade-point gain once adoption is held fixed would overturn the central claim.

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

3 major / 5 minor

Summary. The paper reports a three-arm randomized experiment (N=164 law students) on a contract-law issue-spotting exam: no GenAI, optional DeepSeek access, or optional access plus a brief training video and quiz. Untrained access does not improve (and on some margins worsens) performance relative to no access; training raises self-reported adoption (41% vs 26%, p=0.044), mean grade point by 0.27 relative to untrained access (p=0.027), and reduces rule misstatements (p=0.014). Principal-stratification bounds under monotonicity, never-user exclusion, and mean dominance are used to argue that the training effect operates primarily through adoption rather than effectiveness among always-users, though the paper notes that bootstrap CIs are wide. The authors further interpret ability-quartile patterns as evidence that, without training, higher-ability students opt out while lower-ability students adopt unproductively.

Significance. If the main ITT contrasts hold, the paper supplies clean experimental evidence that a low-cost, task-specific training intervention can reverse the null or negative effect of optional GenAI access on a judgment-intensive professional task. That finding is directly relevant to legal education, firm training policy, and the broader debate on whether GenAI levels or amplifies skill differences. Strengths include pre-registration of main hypotheses, random assignment, blinded grading, transparent flagging of non-pre-registered tests, and an explicit (if partial) identification strategy for the adoption-versus-effectiveness channels. The contribution is empirical and policy-relevant rather than theoretical; the stylized model in §2 is illustrative scaffolding.

major comments (3)
  1. The abstract and §5.5/§6 claim that training "operates primarily through adoption rather than effectiveness" and that "the adoption lower bound (1.06) exceeds the effectiveness upper bound (0.42) at strict mean dominance." Table 3 shows that under γ=0 the adoption lower-bound 95% bootstrap CI is [-0.364, 1.740] and the effectiveness upper-bound CI is [-0.049, 0.938]; the intervals overlap substantially. The ordering is therefore a point-estimate comparison under an untestable dominance assumption that equalizes the bounds already at γ≈0.64. The main ITT results do not depend on this decomposition; the mechanism language should be rewritten to match the paper's own identification limits (e.g., "point estimates under strict mean dominance favor adoption, but CIs do not reject effectiveness as the larger channel").
  2. Adoption is measured solely by post-exam self-report (§5.1). With optional access and no server logs or prompt records, differential reporting (e.g., trained participants more willing to admit use) cannot be ruled out and would inflate both the first-stage and the principal-stratum proportions. A robustness check that treats non-response or that bounds the adoption contrast under plausible misreporting rates would strengthen the central claim; alternatively the limitation should be stated more prominently when interpreting the 15 pp difference.
  3. Table 4 and the accompanying discussion in §6 treat post-exam grade-point quartiles as a proxy for pre-treatment ability and conclude that training induces adoption among high-ability students (0% → 42% in Q4). Quartile cell sizes are small (n=6 in Group 2 Q4), the analysis is not pre-registered, and the ability measure is the outcome itself, so selection and reverse causality are possible. The claim that "higher-ability practitioners opt out while lower-ability users adopt but unproductively" is load-bearing for the paper's challenge to the leveling narrative; it should be either supported by a pre-treatment ability measure or clearly labeled as exploratory and post-hoc.
minor comments (5)
  1. Several two-tailed tests on length and case citations (§5.3–5.4) are flagged as non-pre-registered; keep that discipline consistently in the abstract and introduction so that the reader does not over-weight them.
  2. Figure 4 density overlay and the grade-point bar chart (Figure 5) would be clearer with exact sample sizes and a note that grade point is discrete (1–4.3).
  3. The theoretical model in §2 uses ΔY_T and L(c,e_T) without stating units or how they map to the grade-point scale; a short sentence linking the model objects to the empirical outcomes would help.
  4. Typo / cross-reference: "Error! Reference source not found." appears in §6 before Table 4.
  5. Appendix preference/helpfulness results are null; a one-sentence summary in the main text would prevent the reader from wondering whether attitudes moved.

Circularity Check

0 steps flagged

No circularity: RCT contrasts and principal-stratification bounds are self-contained empirical results, not forced by definition or self-citation.

full rationale

The paper's load-bearing claims are direct experimental contrasts (adoption rates 41% vs 26%, grade-point difference 0.27, rule-misstatement reduction) from a three-arm RCT of 164 law students, plus partial-identification bounds under stated assumptions (monotonicity, never-user exclusion, mean dominance). The stylized productivity model in §2 is illustrative only and is not used to generate or constrain the empirical estimates. Principal stratification follows Frangakis–Rubin with support and dominance restrictions; the resulting bounds (Table 3) are computed from observed conditional means and stratum proportions, not from parameters fitted to the target effects themselves. No equation equates a claimed prediction to a fitted input by construction; no uniqueness theorem or ansatz is imported via self-citation; references are external (Acemoglu, Autor, Choi/Schwarcz, etc.). The mechanism ordering (adoption lower bound > effectiveness upper bound at γ=0) is a point-estimate comparison under an untestable assumption whose CIs overlap—an identification limitation, not circularity. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 6 axioms · 1 invented entities

The paper is primarily empirical. Load-bearing background assumptions are standard experimental and partial-identification conditions rather than free physical parameters or newly invented particles. The stylized productivity model in §2 is illustrative; the quantitative claims rest on the RCT and the principal-stratification bounds.

free parameters (1)
  • mean-dominance tolerance γ = 0 (strict) and grid up to 1.665
    Sensitivity parameter that relaxes E[Y(1,1)|C] ≥ E[Y(1,1)|A] − γ; results are reported for a grid of γ values, with the strict case γ=0 used for the headline ordering.
axioms (6)
  • domain assumption Random assignment of participants to the three experimental arms (SUTVA and no interference).
    Invoked throughout §3–5 to identify average treatment contrasts and stratum proportions.
  • domain assumption Monotone treatment response: training never decreases DeepSeek use (no defiers).
    Proposition 1 and subsequent bounds in §5.5; standard principal-stratification assumption.
  • domain assumption Exclusion restriction for never-users: training does not affect outcomes of participants who never use DeepSeek.
    Used to point-identify E[Y(0,0)|N] in Proposition 2.
  • domain assumption Global support restriction: grade points lie in [1.0, 4.3] under all potential outcomes.
    Used to construct sharp bounds in Proposition 3.
  • ad hoc to paper Mean dominance (up to γ): trained induced users perform at least as well as trained always users minus γ.
    Introduced in Proposition 4 to tighten bounds; not independently verified.
  • domain assumption Self-reported DeepSeek use is a valid measure of actual adoption during the exam.
    Primary adoption outcome in §5.1; no server logs or keystroke data are reported.
invented entities (1)
  • always / induced / never users (principal strata) no independent evidence
    purpose: Decompose the training effect into an adoption channel (induced users) versus an effectiveness channel (always users).
    Standard Frangakis–Rubin principal stratification labels applied to this experiment; not a new physical entity but a modeling construct whose proportions are only partially identified.

pith-pipeline@v1.1.0-grok45 · 24273 in / 3179 out tokens · 30702 ms · 2026-07-15T14:52:37.219281+00:00 · methodology

0 comments
read the original abstract

Can targeted user training unlock the productive potential of generative artificial intelligence in professional settings? We study this question using a randomized experiment in which 164 law students completed an issue-spotting examination under one of three conditions: no GenAI access, optional access to a large language model (LLM), or LLM access with a brief training intervention. Untrained LLM access proved counterproductive: relative to participants without any LLM access, untrained users wrote significantly shorter answers, committed more case misstatements, and scored marginally lower, though most differences fall short of conventional significance. Training reversed this pattern. Trained participants adopted the LLM at higher rates (41% vs. 26%; p = 0.044), scored 0.27 grade points higher than untrained users--roughly one fine grade--(p = 0.027), and stated applicable rules more accurately (p = 0.014). Principal stratification analysis suggests training operates primarily through adoption rather than effectiveness--the adoption lower bound (1.06) exceeds the effectiveness upper bound (0.42) at strict mean dominance--though confidence intervals are wide. More broadly, these findings challenge the view that GenAI primarily benefits lower-skilled workers: without training, higher-ability practitioners opt out while lower-ability users adopt but unproductively. Realizing GenAI's productivity gains requires investment in both access and instruction.

Figures

Figures reproduced from arXiv: 2603.04982 by Benjamin M. Chen, Hong Bao.

Figure 1
Figure 1. Figure 1: A Screenshot of the Training Video 3 The supplementary examination was taken by a small number of students, none of whom were eligible for this study, and has not been made available as part of the University of Hong Kong Law Library’s collection of past year papers. 4 This material was created using Youyan 3D, a video generation platform [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Information Displayed on Electronic Screens in Classrooms Alt Text: Resource access notices displayed on electronic screens in examination classrooms. The left panel shows instructions for Group 1, permitting access to course notes and Westlaw only. The right panel shows instructions for Groups 2 and 3, additionally permitting access to DeepSeek. At the end of the mock examination, answers were submitted e… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution Density of Grade Point by Groups Alt text: Density plot titled “Distribution Density of Grade Point” showing three overlapping kernel density curves for Grade Point (x-axis, range 0–5) against Density (y-axis, range 0–0.55). Curves represent three groups: “No AI” (black solid line, peak at ~2.0, highest density), “AI without Training” (dark gray line, peak at ~2.0), and “AI with Training” (lig… view at source ↗
Figure 5
Figure 5. Figure 5: Average Grade Point by Group Alt text: Bar chart showing mean grade point by experimental group, with 95% confidence intervals. Group 1 (No AI, dark grey bar): 2.29; Group 2 (AI without Intervention, medium grey bar): 2.25; Group 3 (AI with Intervention, light grey bar): 2.52. The difference between Groups 2 and 3 is statistically significant (d = 0.270; p = 0.027*, one-tailed t-test). The difference betwe… view at source ↗
Figure 6
Figure 6. Figure 6: Average Number of Missed Issues by Group Alt text: Bar chart showing mean number of examination issues missed by group, with 95% confidence intervals. Group 1 (No AI, dark grey bar): 1.10; Group 2 (AI without Intervention, medium grey bar): 1.23; Group 3 (AI with Intervention, light grey bar): 0.98. No statistically significant differences are detected across groups on this measure [PITH_FULL_IMAGE:figure… view at source ↗
Figure 7
Figure 7. Figure 7: Average Complexity by Group Alt text: Bar chart showing mean Flesch–Kincaid grade level of answers by group, with 95% confidence intervals, used as a measure of answer complexity. Group 1 (No AI, dark grey bar): 12.84; Group 2 (AI without Intervention, medium grey bar): 12.38; Group 3 (AI with Intervention, light grey bar): 12.13. No statistically significant differences are detected across groups. Length … view at source ↗
Figure 8
Figure 8. Figure 8: Average Length by Group Alt text: Bar chart showing mean word count of answers by group, with 95% confidence intervals. Group 1 (No AI, dark grey bar): 1,059.67; Group 2 (AI without Intervention, medium grey bar): 895.74; Group 3 10 These tests were not pre-registered [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Average Number of Rule Misstatements by Group Alt text: Bar chart showing mean number of rule misstatements by group, with 95% confidence intervals. Group 1 (No AI, dark grey bar): 2.47; Group 2 (AI without Intervention, medium grey bar): 2.58; Group 3 (AI with Intervention, light grey bar): 2.16. The difference between Groups 2 and 3 is statistically significant (d = −0.424; p = 0.014*, one-tailed t-test)… view at source ↗
Figure 10
Figure 10. Figure 10: Average Number of Case Misstatements by Group Alt text: Bar chart showing mean number of case misstatements by group, with 95% confidence intervals. Group 1 (No AI, dark grey bar): 0.02; Group 2 (AI without Intervention, medium grey bar): 0.19; Group 3 11 This test was not pre-registered [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Average Permission by Group Alt text: Bar chart showing mean permission score by group, with 95% confidence intervals. Participants rated their agreement with allowing LLM access in law examinations on a 1–5 scale. Group 1 (No AI, dark grey bar): 3.57; Group 2 (AI without Intervention, medium grey bar): 3.23; Group 3 (AI with Intervention, light grey bar): 3.38. No statistically significant differences ar… view at source ↗
Figure 12
Figure 12. Figure 12: Average Helpfulness by Group Alt text: Bar chart showing mean helpfulness score by group, with 95% confidence intervals. Participants rated their agreement that LLM access would help them in law examinations on a 1–5 scale. Group 1 (No AI, dark grey bar): 3.12; Group 2 (AI without Intervention, medium grey bar): 3.25; Group 3 (AI with Intervention, light grey bar): 3.29. No statistically significant diffe… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Eskwai for Students: Generative AI Assistant for Legal Education in Ghana

    cs.CL 2026-05 unverdicted novelty 5.0

    Eskwai for Students is a RAG-based AI system for Ghanaian legal questions, deployed over 30 months to 3.1K students who submitted 32K queries, with evaluation of helpfulness and identification of ethical concerns in u...

Reference graph

Works this paper leans on

32 extracted references · 6 canonical work pages · cited by 1 Pith paper

  1. [1]

    The Simple Macroeconomics of AI

    “The Simple Macroeconomics of AI.” Economic Policy 40 (121): 13–58. https://doi.org/10.1093/epolic/eiae042. Acemoglu, Daron, and David Autor

    doi:10.1093/epolic/eiae042

  2. [2]

    https://doi.org/10.1016/S0169-7218(11)02410-5

    Elsevier. https://doi.org/10.1016/S0169-7218(11)02410-5. Acemoglu, Daron, and Pascual Restrepo

    doi:10.1016/s0169-7218(11)02410-5

  3. [3]

    Automation and New Tasks: How Technology Displaces and Reinstates Labor

    “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33 (2): 3–30. https://doi.org/10.1257/jep.33.2.3. Acemoglu, Daron, and Pascual Restrepo

    doi:10.1257/jep.33.2.3

  4. [4]

    Robots and Jobs: Evidence from US Labor Markets

    “Robots and Jobs: Evidence from US Labor Markets.” The Journal of Political Economy 128 (6): 1–57. https://doi.org/10.1086/705716. Agrawal, Ajay K., John McHale, and Alexander Oettl

    doi:10.1086/705716

  5. [5]

    Working Paper No

    Enhancing Worker Productivity Without Automating Tasks: A Different Approach to AI and the Task- Based Model. Working Paper No. 34781. National Bureau of Economic Research. https://doi.org/10.3386/w34781. American Bar Association

    work page doi:10.3386/w34781

  6. [6]

    Angrist, Joshua D., Guido W

    https://www.americanbar.org/groups/law_practice/resources/law- technology-today/2025/the-legal-industry-report-2025/. Angrist, Joshua D., Guido W. Imbens, and Donald B. Rubin

    2025

  7. [7]

    Identification of Causal Effects Using Instrumental Variables

    “Identification of Causal Effects Using Instrumental Variables.” Journal of the American Statistical Association 91 (434): 444–55. https://doi.org/10.1080/01621459.1996.10476902. Autor, David H

    doi:10.1080/01621459.1996.10476902 1996

  8. [8]

    Working Paper No

    Applying AI to Rebuild Middle Class Jobs. Working Paper No. 32140. National Bureau of Economic Research. https://doi.org/10.3386/w32140. Autor, David H., Frank Levy, and Richard J. Murnane

    work page doi:10.3386/w32140

  9. [9]

    The Skill Content of Recent Technological Change: An Empirical Exploration

    “The Skill Content of Recent Technological Change: An Empirical Exploration.” The Quarterly Journal of Economics 118 (4): 1279–333. https://doi.org/10.1162/003355303322552801. Bauer, Emmanuel, Dominik Stammbach, Nianlong Gu, and Elliott Ash

    doi:10.1162/003355303322552801

  10. [10]

    General Purpose Technologies ‘Engines of Growth’?

    “General Purpose Technologies ‘Engines of Growth’?” Journal of Econometrics 65 (1): 83–108. https://doi.org/10.1016/0304- 4076(94)01598-T. Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond

    doi:10.1016/0304-

  11. [11]

    AI Hallucination Cases

    Generative AI at Work. Working Paper No. 31161. National Bureau of Economic Research. https://doi.org/10.3386/w31161. Charlotin, Damien. n.d. “AI Hallucination Cases.” Accessed January 28,

    doi:10.3386/w31161

  12. [12]

    Lawyering in the Age of Artificial Intelligence

    “Lawyering in the Age of Artificial Intelligence.” Minnesota Law Review 109 (1): 147–218. https://doi.org/10.24926/265535.4225. Choi, Jonathan H., and Daniel B. Schwarcz

    work page doi:10.24926/265535.4225

  13. [13]

    Understanding the Duty of Competence for Attorneys Using Generative Ai

    “Understanding the Duty of Competence for Attorneys Using Generative Ai.” North Carolina Journal of Law & Technology 27: 1–26. https://doi.org/10.2139/ssrn.5053423. Cui, Kevin Zheyuan, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng, and Tobias Salz

    work page doi:10.2139/ssrn.5053423

  14. [14]

    The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers

    “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.” Preprint, Massachusetts Institute of Technology, Department of Economics, February. https://economics.mit.edu/sites/default/files/2024- 01/draft_copilot_experiments.pdf. Dahl, Matthew, Varun Magesh, Mirac Suzgun, and Daniel E. Ho

    2024

  15. [15]

    Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models

    “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models.” Journal of Legal Analysis 16 (1): 64–93. https://doi.org/10.1093/jla/laae003. Dell’Acqua, Fabrizio, Edward McFowland, Ethan R. Mollick, et al

    doi:10.1093/jla/laae003

  16. [16]

    Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality

    “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” SSRN Electronic Journal, ahead of print. https://doi.org/10.2139/ssrn.4573321. Denniston, Alex, and Peter Duffy

    doi:10.2139/ssrn.4573321

  17. [17]

    GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models

    “GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.” arXiv:2303.10130. Preprint, arXiv, August

    Pith/arXiv arXiv

  18. [18]

    Engel, Christoph, and Johannes Kruse

    https://doi.org/10.48550/arXiv.2303.10130. Engel, Christoph, and Johannes Kruse

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.10130

  19. [19]

    Principal Stratification in Causal Inference

    “Principal Stratification in Causal Inference.” Biometrics 58 (1): 21–29. https://doi.org/10.1111/j.0006- 341X.2002.00021.x. Freitas, Julian

    Pith/arXiv arXiv doi:10.1111/j.0006- 2002

  20. [20]

    Why People Resist Embracing AI

    “Why People Resist Embracing AI.” Harvard Business Review, January. https://hbr.org/2025/01/why-people-resist-embracing-ai. Humlum, Anders, and Emilie Vestergaard

    2025

  21. [21]

    Confidence Intervals for Partially Identified Parameters

    “Confidence Intervals for Partially Identified Parameters.” Econometrica 72 (6): 1845–57. https://doi.org/10.1111/j.1468-0262.2004.00555.x. Lam, Kwok-Yan, Victor C. W. Cheng, and Zee Kin Yeong

    doi:10.1111/j.1468-0262.2004.00555.x 2004

  22. [22]

    Building a Better Lawyer: Experimental Evidence That Artificial Intelligence Can Increase Legal Work Efficiency

    “Building a Better Lawyer: Experimental Evidence That Artificial Intelligence Can Increase Legal Work Efficiency.” Journal of Empirical Legal Studies 21 (4): 979–1022. https://doi.org/10.1111/jels.12396. Noy, Shakked, and Whitney Zhang

    doi:10.1111/jels.12396

  23. [23]

    Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence

    “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science 381 (6654): 187–92. https://doi.org/10.1126/science.adh2586. Peng, Sida, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer

    doi:10.1126/science.adh2586

  24. [24]

    The Impact of AI on Developer Productivity: Evidence from GitHub Copilot

    “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.” arXiv:2302.06590. Preprint, arXiv, February

    Pith/arXiv arXiv

  25. [25]

    Rubin, Donald B

    https://doi.org/10.48550/arXiv.2302.06590. Rubin, Donald B

    Pith/arXiv arXiv doi:10.48550/arxiv.2302.06590

  26. [26]

    Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies

    “Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies.” Journal of Educational Psychology 66 (5): 688–701. https://doi.org/10.1037/h0037350. Rubin, Donald B

    doi:10.1037/h0037350

  27. [27]

    Inference and Missing Data

    “Inference and Missing Data.” Biometrika 63 (3): 581–92. https://doi.org/10.1093/biomet/63.3.581. Rubin, Donald B

    doi:10.1093/biomet/63.3.581

  28. [28]

    Estimating Causal Effects from Large Data Sets Using Propensity Scores

    “Estimating Causal Effects from Large Data Sets Using Propensity Scores.” Annals of Internal Medicine 197 (8_Part_2): 757–63. https://doi.org/10.7326/0003-4819-127-8_Part_2-199710151-00064. Savelka, Jaromir

    work page doi:10.7326/0003-4819-127-8_part_2-199710151-00064

  29. [29]

    Unlocking Practical Applications in Legal Domain: Evaluation of GPT for Zero-Shot Semantic Annotation of Legal Texts

    “Unlocking Practical Applications in Legal Domain: Evaluation of GPT for Zero-Shot Semantic Annotation of Legal Texts.” Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law, June 19, 447–51. https://doi.org/10.1145/3594536.3595161. Schwarcz, Daniel, and Jonathan H. Choi

    doi:10.1145/3594536.3595161

  30. [30]

    AI Tools for Lawyers: A Practical Guide

    “AI Tools for Lawyers: A Practical Guide.” Minnesota Law Review Headnotes 108 (1): 1–39. https://minnesotalawreview.org/wp-content/uploads/2023/10/FL1-Choi- Schwarcz.pdf. Schwarcz, Daniel, Sam Manning, Patrick James Barry, David R. Cleveland, J. J. Prescott, and Beverly Rich

    2023

  31. [31]

    AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice

    “AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice.” Preprint, SSRN. https://doi.org/10.2139/ssrn.5162111. Splawa-Neyman, Jerzy, D. M. Dabrowska, and T. P. Speed

    work page doi:10.2139/ssrn.5162111

  32. [32]

    On the Application of Probability Theory to Agricultural Experiments. Essay on Principles. Section 9

    “On the Application of Probability Theory to Agricultural Experiments. Essay on Principles. Section 9.” Statistical Science 5 (4). https://doi.org/10.1214/ss/1177012031. Thomson Reuters

    doi:10.1214/ss/1177012031