Evidence (8807 claims)
Search and filter individual claims pulled from the papers. Looking for a specific finding ("what's the effect on wages?"), you're in the right place. Want to compare whole outcome categories against each other instead? Use the Evidence Explorer.
The board below groups claims two ways: by broad theme (nine paper-level topics) and by outcome category (the 34 claim-level outcomes that the Explorer and Syntheses also use).
Browse by theme
Nine broad, paper-level topics. Click one to filter the claims below.
Adoption
9875 claims
Filter claims →
Productivity
8807 claims
Filtered →
Governance
7870 claims
Filter claims →
Human-AI Collaboration
7560 claims
Filter claims →
Org Design
4892 claims
Filter claims →
Innovation
4781 claims
Filter claims →
Labor Markets
4004 claims
Filter claims →
Skills & Training
3308 claims
Filter claims →
Inequality
2332 claims
Filter claims →
Claims by outcome category
Counts by direction of finding. These are the same 34 outcome categories the Explorer compares and the Syntheses are written for. A linked row has a published synthesis.
| Outcome | Positive | Negative | Mixed | Null | Total |
|---|---|---|---|---|---|
| Other | 870 | 233 | 116 | 1066 | 2363 |
| Governance & Regulation | 976 | 451 | 218 | 133 | 1809 |
| Organizational Efficiency | 949 | 224 | 144 | 88 | 1416 |
| Technology Adoption Rate | 764 | 287 | 141 | 122 | 1325 |
| Research Productivity | 501 | 152 | 74 | 362 | 1101 |
| Output Quality | 542 | 216 | 69 | 69 | 896 |
| Decision Quality | 387 | 198 | 94 | 54 | 740 |
| Firm Productivity | 513 | 67 | 101 | 27 | 714 |
| AI Safety & Ethics | 249 | 303 | 73 | 36 | 667 |
| Market Structure | 190 | 192 | 134 | 27 | 548 |
| Task Allocation | 243 | 77 | 91 | 36 | 452 |
| Innovation Output | 291 | 33 | 55 | 20 | 401 |
| Skill Acquisition | 206 | 72 | 65 | 21 | 364 |
| Employment Level | 133 | 63 | 115 | 22 | 335 |
| Fiscal & Macroeconomic | 153 | 79 | 52 | 32 | 323 |
| Task Completion Time | 206 | 37 | 12 | 15 | 272 |
| Firm Revenue | 179 | 52 | 29 | 5 | 266 |
| Consumer Welfare | 130 | 76 | 47 | 13 | 266 |
| Inequality Measures | 48 | 137 | 51 | 6 | 242 |
| Worker Satisfaction | 101 | 81 | 25 | 13 | 220 |
| Error Rate | 84 | 110 | 11 | 5 | 210 |
| Wages & Compensation | 98 | 47 | 30 | 10 | 185 |
| Regulatory Compliance | 88 | 73 | 17 | 7 | 185 |
| Automation Exposure | 66 | 64 | 33 | 16 | 182 |
| Team Performance | 105 | 29 | 30 | 11 | 176 |
| Training Effectiveness | 109 | 22 | 14 | 21 | 168 |
| Developer Productivity | 114 | 21 | 14 | 8 | 158 |
| Job Displacement | 12 | 90 | 24 | 1 | 127 |
| Hiring & Recruitment | 57 | 9 | 9 | 5 | 80 |
| Skill Obsolescence | 6 | 56 | 9 | 1 | 72 |
| Social Protection | 43 | 17 | 8 | 2 | 70 |
| Creative Output | 35 | 21 | 9 | 4 | 70 |
| Labor Share of Income | 18 | 21 | 17 | 1 | 57 |
| Worker Turnover | 15 | 16 | — | 4 | 35 |
| Industry | — | — | — | 1 | 1 |
Productivity
Remove filter
The productivity gap is attributable to organizational and behavioral factors.
Theoretical analysis, generalization, and synthesis of corporate cases and empirical reports linking observed micro-behaviors to macro-level productivity outcomes.
There is a gap between anticipated macroeconomic efficiency gains (aggregate labor productivity) and observed micro-level outcomes following AI adoption.
Comparison of aggregate productivity trends (BLS series/AAPC calculations) with micro-level evidence drawn from corporate case studies and empirical reports documenting localized impacts of AI.
The CAD has implications for knowledge-work stratification and AI platform governance.
Argumentative/policy discussion in the paper linking the CAD to potential stratification among knowledge workers and governance considerations for AI platforms.
The probabilistic model demonstrates that manual context attachment leads to a combinatorial collapse in task-success probability as corpus size and task conjunctivity grow.
Results from the paper's probabilistic model (analytic/theoretical demonstration based on fan effect reasoning); no empirical sample reported.
For knowledge-intensive workers whose intellectual capital spans tens of thousands of files, the CAD constitutes a qualitative threshold in AI usefulness: below it, the cognitive burden of context curation falls on the human, reproducing the inefficiencies AI is meant to eliminate.
Theoretical argument grounded in the paper's conceptual discussion about large personal/organizational corpora (stated scale: tens of thousands of files) and the user burden of manual context attachment.
There exists a finer-grained divide at the level of individual interaction — the Context Access Divide (CAD) — whereby two users with nominally equivalent agent access may experience qualitatively different AI utility depending on whether the system can autonomously retrieve context (Dynamic Context Retrieval) or requires manual document attachment (Manual Attachment).
Conceptual argument and definitional framing in the paper introducing the CAD as a novel dimension of inequality; comparison of two interaction modalities (Dynamic Context Retrieval vs Manual Attachment).
Free overrides also cut sales by 1.19%.
Randomized field experiment comparing free-overrides arm to control; effect reported as 1.19% reduction in sales.
Free overrides reduce inventory by 1.95%.
Randomized field experiment comparing free-overrides arm to control; effect reported as 1.95% reduction in inventory.
The testing tool raised cost by 42 to 68 percent without improving functional score or reliability, even on interface visible criteria.
Comparison between runs with and without the testing tool showing reported cost increase (42–68%) and no improvement in functional score or reliability.
Container deployment was the dominant defect, failing first try in 44 percent of runs.
Criterion-level analysis of failure modes across the 90 runs reporting first-try failure frequency for container deployment.
Monopoly production of AI restricts its deployment, slowing the transition and impact of AI.
Theoretical model comparing monopolistic AI producer behavior to competitive deployment; result is derived analytically. No empirical sample reported.
Wages of labor that is substituted for by AI decrease in both absolute and relative terms.
Analytical economic model / comparative statics predicting wage declines for labor substituted by AI. No empirical sample reported.
Computational theorising, synthetic task simulations, real LLM agent traces, and robustness analyses show that human-imitation forms often underperform when they add lossy handoffs, correlated deliberation, and verification burdens.
Empirical and simulation-based methods listed in the paper (computational theorising, synthetic task simulations, analysis of real LLM agent traces, robustness checks). The excerpt does not report sample sizes, numeric effect sizes, or statistical tests.
The AI premium is absent in emerging markets, including China.
Geographic cross-sectional analysis indicating no significant AI premium in emerging market firms (explicitly mentioning China).
Unemployment, idle labour and weakening domestic demand are holding back growth.
Empirical results reported in the paper indicate a negative relationship between unemployment (and related indicators) and economic growth in the 27-country panel (2008–2020).
AI poses environmental challenges.
The abstract lists environmental challenges as one of the potential trade-offs identified by the systematic review of 194 articles.
AI can contribute to widening inequality.
Abstract reports the review identifies widening inequality as a potential trade-off of AI, based on synthesis of 194 articles.
AI can give rise to job displacement.
The abstract states the review finds potential trade-offs including job displacement across the surveyed literature (194 articles).
Only a small percentage of the human-labelled bugs are detected as being likely associated with LLM-generated code.
Comparison between a set of human-labelled bugs and detector-flagged LLM-generated code to assess co-occurrence (human-labelled bug sample size not provided).
Comments exhibit a relatively low proportion of grammatically correct sentences.
Linguistic/grammatical analysis of comments flagged as likely LLM-generated (detector-based proxy analysis).
Code detected as likely to be generated by LLMs decreased over time (2021–2025).
Detector-based proxy analysis on active company- and community-maintained repositories from 2021 to 2025 using various tools and techniques to detect LLM-generated code.
In open-ended collaboration and bargaining, the same manipulation substantially degrades performance.
Experimental manipulation of agreeableness in LLMs on open-ended research collaboration and competitive bargaining tasks; authors report substantial performance degradation in these domains. Abstract lacks numeric metrics, sample sizes, and statistical significance details.
In the same repositories, agent-authored contributions concentrate repository-level friction roughly twice as much as human ones (intraclass correlation 0.30 versus 0.16); this gap holds after controlling for codebase size, age, task shape, process maturity, and merge path.
Comparison of intraclass correlations (ICC) between agent-authored and human-authored pull requests using multilevel models with controls for codebase characteristics and process variables. Dataset includes >930,000 agent-authored PRs (human sample size not specified in excerpt).
The paper reframes humans not as passive users, but as core system components whose competencies, limitations, and adaptive capacities constrain the performance envelope of optimized AI systems.
Framing/interpretive claim derived from the paper's perspective and literature synthesis (conceptual; no empirical support provided in text).
Organizational structures, bias susceptibility, retraining constraints, and interface design co-determine system stability, error propagation, and optimization ceilings.
Conceptual claim based on synthesis of literature across organizational adoption and ML lifecycle management (no empirical tests or sample sizes reported).
Human interfaces define throughput limits in areas such as prompt engineering, data-stream curation, adjudication of model outputs, and the orchestration of hybrid automation workflows including robotics, scraping, and digitization.
Theoretical assertion supported by the paper's systems-oriented analysis and literature synthesis (no empirical measurement or sample size provided).
Despite accelerating advances in AI capabilities, human capital remains the enduring and dominant system constraint.
Argument and synthesis of emerging research across human-AI interaction, ML lifecycle management, organizational adoption, and adult learning theory (conceptual synthesis; no empirical sample size reported).
GAGI is a necessary complement to GDP-based monitoring: any macroeconomic monitoring instrument that tracks only aggregate output will systematically miss the distributional harm that automation can cause even while reported growth remains strong.
Argument combining conceptual critique of GDP with empirical demonstration on G7 data using the GAGI index (authors' normative policy recommendation).
The divergence between welfare-adjusted prosperity (GAGI) and headline GDP widens sharply after 2022, temporally coincident with the after-effects of COVID and the acceleration of generative-AI deployment, though this evidence alone does not demonstrate causation.
Temporal pattern observed in the authors' G7 2010–2026 empirical series (associational observation; authors explicitly note lack of causal identification).
Applying GAGI to the G7 economies over 2010-2026 shows that welfare-adjusted prosperity has diverged persistently and increasingly from headline GDP growth.
Empirical analysis performed on G7 countries over 2010–2026 (sample: 7 economies; time series comparison of GAGI vs. GDP per capita).
What is missing from the macroeconomic monitoring toolkit is an operational monitoring trigger: a statistic minimal enough to compute annually from public data, transparent enough to audit without modelling assumptions, and normalised so that year-on-year, cross-country change is legible to a regulator.
Normative/methodological claim by the authors arguing for a practical monitoring statistic (no empirical test; statement of need).
GDP per capita is blind to two first-order determinants of lived prosperity: income/wealth distribution and inflation impact.
Conceptual/definitional argument presented by the authors in the paper (no empirical test reported).
Capital distortion negatively moderates the effect of industrial robots on firm-level TFP (i.e., capital distortion reduces the positive impact of robots on TFP).
Moderation/interaction analysis using the same panel data of Chinese listed firms and industrial robots (2006–2019); reported tests of interaction between robot application and measures of capital distortion.
Longer system responses and more information-providing turns negatively affect user satisfaction.
Statistical modeling of user satisfaction using features of multi-turn interactions (response length, number of information-providing turns) derived from the 49 participant sessions; models show negative associations reported in the paper.
Accuracy of developer+LLM assessments against expert ground truth is low.
Comparison of participant/LLM assessment outcomes to expert-annotated ground truth for the 148 NFRs; reported low accuracy in the paper.
Existing benchmarks typically report accuracy for a single model on a single run, which systematically understates real-world LLM capabilities—particularly under heterogeneous data distributions—because (i) different models get different questions correct according to their specializations, and (ii) given a budget, multiple generations can be sampled and selectively retained.
Argument supported by the paper's empirical Capability Frontier analysis (21 models, 16 benchmarks) and supporting simulations demonstrating specialization and gains from multiple samples/selection.
We surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents (e.g., shortcuts).
Empirical analysis of failures and shortcuts observed when evaluating more capable agents on CORE-Bench Hard (case-study observations).
When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version; this accuracy-centric approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance (construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration).
Argument and framing in the paper supported by conceptual analysis and the CORE-Bench Hard case study (qualitative reasoning and empirical examples).
GenAI adoption carries risks including overreliance on models, misalignment between model outputs and human needs, and uneven performance across tasks and contexts.
Reported adverse effects and risks identified in the reviewed literature (task-level experiments and applied studies summarized by the paper).
Traditional dynamic pricing models in large-scale e-commerce suffer from limited interpretability, poor utilization of unstructured information, and misalignment with long-term business objectives such as cumulative Gross Merchandise Value (GMV), Return on Investment (ROI) and milestone achievement.
Author statement in paper abstract/introduction asserting limitations of existing dynamic pricing models (literature/motivation claim); no empirical test or sample size reported in the provided text.
Re-asking the same GPQA-Diamond questions in free-response rather than multiple-choice form reopens the all-wrong tail: beta is 0.127; a five-judge panel had kappa 0.73 to 0.92, locating co-failure in answer format rather than subject.
Empirical comparison of GPQA-Diamond questions in multiple-choice vs free-response format, with human adjudication by a five-judge panel and reported inter-rater agreement (kappa range).
On execution-graded code, the observed all-wrong rate beta is 0.079.
Empirical measurement on execution-graded code tasks reported in the paper (dataset/method described in the study).
Across 67 models from 21 providers on open-ended mathematics, observed beta is 0.052 versus 0.023 under the full 67-model Gaussian copula (tetrachoric-calibrated single-factor model underprices the all-wrong tail), about 2.5 times underpricing with 90% CI 1.7 to 3.4 and k = 17.
Empirical evaluation across 67 models from 21 providers on open-ended mathematics; comparison between observed all-wrong rate and rate predicted by a full 67-model Gaussian copula / tetrachoric-calibrated single-factor model; confidence interval reported and k specified (k = 17 as reported).
For any policy whose output is one member model answer, accuracy cannot exceed one minus beta, where beta is the rate at which every model is wrong on the same query.
Theoretical derivation/proof given in the paper (a formal upper bound linking ensemble accuracy to the all-wrong rate beta).
This analysis (identifying robust trait combinations across the full tested space) was infeasible with the mechanistic model alone.
Comparative statement: inability to perform such exhaustive search with APSIM due to computational expense, contrasted with capability of the emulator.
AI has a significant negative influence on value chain upgrading in labor-intensive equipment manufacturing industries.
Industry-type heterogeneity analysis within the same 30-province panel (2010–2022) showing a statistically significant negative coefficient for labor-intensive subsectors.
There is a Verifier-Goodharting Floor on flywheel ceilings under imperfect rewards (a formal result showing a lower bound / floor imposed by reward imperfection and verifier limitations).
Formal theoretical result in the paper (derivation showing that imperfect/verifier-mediated rewards create a floor limiting flywheel / self-improvement ceilings).
There exists a Determinism-Efficiency Bound on chain-task success (one of three formal results pinning down the regime).
Formal theoretical result presented in the paper (a derived bound relating environment determinism to chain-task efficiency/success).
With per-step determinism δ < 1, k-step chain success degrades as δ^k (long-chain agent execution fails exponentially in environments designed for human tolerance).
Formal theoretical statement in the paper: a mathematical characterization of per-step determinism and k-step chain success (Derivation / formal result presented by the authors).
Infrastructural deficiencies, unstable electricity supply, limited technical expertise, and high implementation costs remain major barriers to AI adoption.
Survey responses from 522 participants reporting barriers to AI adoption; descriptive analysis identifying these factors as major barriers.