Evidence (16496 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
Filter claims →
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 |
The bottleneck is often not model capability but missing project memory.
Assertion made in the abstract without accompanying quantitative evidence in the abstract.
Reconstructing this context can consume an estimated 5,000-20,000 tokens per session.
Statement in paper abstract presenting an estimate (no detailed method or sample described in the abstract).
Existing benchmarks for time-series forecasting focus solely on prediction error metrics; the decision utility of advanced forecasting (foundation) models remains unverified.
Authors' literature/benchmark review and critique presented in the paper.
Resource utilization in cloud data centers remains at low levels due to conservative over-provisioning to guarantee service reliability.
Stated as background motivation in the paper (literature/operational observation); asserted by authors as common industry phenomenon.
With the instruction files, 26.35% of the projects decreased their merge rate.
Reported proportion of projects showing a decrease in merge rate after creating instruction files based on the pre/post comparison of projects in the dataset (148 projects, 15,549 PRs).
Xie et al. (2026) show experimentally that job candidates are less satisfied with firms using AI evaluators than with human experts due to perceived loss of control; the negative effect is stronger for individuals with an internal locus of control.
Experimental study on recruitment using control theory as described (sample size not provided).
In the healthcare sector, Chou et al. (2026a, 2026b) identify AI anxiety as a multifaceted hurdle to adoption; emotional affect and outcome expectations are essential influences on usage intentions (two-stage SEM-ANN approach).
Two-stage SEM–ANN modeling grounded in social cognitive theory as reported; empirical data specifics not provided in text.
Liu et al. (2026a, 2026b) find experimentally that the severity of AI service failure in hotel contactless services significantly decreases customers' forgiveness willingness, but high levels of brand attachment mitigate this negative effect.
Experimental studies in hotel contactless service contexts (details and sample sizes not provided in the text).
Allowing AI to take the lead in strategic decision-making without human wisdom may be inappropriate due to AI's inability to navigate tacit knowledge and ethical nuances in Chinese management wisdom.
Argumentative claim based on cited literature (e.g., De Cremer and Kasparov, 2021; Del Giudice et al., 2023) and authors' synthesis.
Developers reject fixes for (a) incorrect implementation (e.g., incomplete, wrong approach), (b) fixes that do not pass CI pipelines and fail tests, (c) fixes for which the agent is unable to perform the implementation (e.g., no code generated, sessions lost), and (d) fixes whose priority is low.
Observed categories from the qualitative analysis of the 306 non-merged PRs described in the study.
The qualitative findings identify 14 reasons divided into four high-level categories for rejecting AI-agent fixes.
Result of the paper's qualitative analysis on the representative sample (306 non-merged PRs).
From a first exploration of the AIDev dataset, 46.41% of the fixes proposed by the agents Copilot, Devin, Cursor, and Claude are rejected.
Empirical analysis of the AIDev dataset reported by the authors; agents named explicitly (Copilot, Devin, Cursor, Claude).
One in three Scheduled Tribe (ST) graduates work in farm or elementary occupations untouched by AI.
Occupational distribution from PLFS 2025 after mapping AI-exposure indices; reported share of ST graduates in farm/elementary (AI-unexposed) occupations in the 83,000-employed-graduate sample.
One in four Scheduled Caste (SC) graduates work in farm or elementary occupations untouched by AI.
Occupational distribution from PLFS 2025 after mapping AI-exposure indices; reported share of SC graduates in farm/elementary (AI-unexposed) occupations in the 83,000-employed-graduate sample.
Graduates from the Scheduled Castes and the Scheduled Tribes are 0.24--0.37 standard deviations less exposed than upper-caste graduates within the same district.
Within-district comparisons using three occupational AI-exposure indices mapped to PLFS 2025; reported standardized exposure differences for SC and ST graduates relative to upper-caste graduates in the 83,000-employed-graduate sample.
Existing evaluation frameworks mask critical architectural gaps and inefficiencies of complex MAS by failing to account for the marginal utility of increased computational cost.
Comparative analysis of performance versus computational cost across evaluated systems showing limited marginal gains despite higher cost (authors' analysis across experiments).
Across traditional reasoning datasets and tasks with interactive multi-step workflows (e.g., BrowseComp-Plus), automatically generated MAS consistently underperform Chain-of-Thought with Self-Consistency (CoT-SC) despite being up to 10x more expensive.
Systematic empirical evaluation comparing automatically generated MAS to CoT-SC across multiple task suites including traditional reasoning datasets and interactive multi-step workflows such as BrowseComp-Plus (experimental comparisons reported in the paper).
Empirical support for MAS superiority relies primarily on comparisons with SAS baselines using benchmarks that prioritize isolated reasoning tasks, which do not adequately assess MAS advantages.
Critical literature review and analysis of prior empirical evaluations (authors' claim about the composition and limitations of existing benchmarks).
Right now, across the world, AI agents recompute a document's KV cache from scratch each time a document is read, so many agents redundantly re-run the compute-intensive prefill step on identical text.
Authors' empirical/operational observation of current agent behavior described in the paper (no explicit sample size reported).
Even a high-quality initial model can be undermined over time because human creators, relying on the model, produce more homogeneous content that harms subsequent training and lowers model performance.
Dynamic theoretical modeling and qualitative analysis in the paper demonstrating increased reliance on AI-assisted content creation and its effect on training data distribution.
In a dynamic model, an initially good AI model induces greater human reliance on AI-assisted creation, which homogenizes content and creates a feedback loop that degrades the model's own performance — a phenomenon termed the "curse of precision."
Dynamic theoretical model and analysis presented in the paper showing feedback effects across training rounds; no empirical sample reported.
More innovative creators are especially harmed under the strong-IP regime — a phenomenon the paper terms the "originality penalty."
Analytical result derived from the static game model in the paper highlighting differential effects by creator innovativeness; theoretical characterization labeled "originality penalty."
A regime of strong intellectual property rights, modeled as a static Stackelberg game, also fails to provide adequate creative incentives (it underpowers creative incentives).
Theoretical analysis using a static Stackelberg-game model developed in the paper; analytical results show reduced creator incentives under this regime.
A "free-for-all" model based on fair use fails because it does not compensate creators for their contributions.
Analytical / conceptual argument presented in the paper (no empirical sample); model-based reasoning showing creators receive no compensation under a free-for-all regime.
More than 70% of respondents cite organisational resistance as a barrier to digital adoption.
Industry MRO digital survey reported in the paper (more than 70% reported); method = secondary evidence from an industry MRO digital survey. Sample size not stated in abstract.
Over 80% of respondents cite data limitations as a barrier to scaling digital implementations.
Industry MRO digital survey reported in the paper (over 80% reported); method = secondary evidence from an industry MRO digital survey. Sample size not stated in abstract.
Only 6% of MROs have scaled digital and analytics across the enterprise.
Industry MRO digital survey reported in the paper (6% reported); method = secondary evidence from an industry MRO digital survey. Sample size not stated in abstract.
Unequal access to GenAI tools in higher education may exacerbate employability gaps and inequities among students.
Concern identified and discussed in the literature as summarized in the review article (conceptual/literature-based; no new empirical evidence reported).
GenAI raises concerns including passive dependence, weakened critical thinking, uncertain authorship, academic integrity breaches, algorithmic bias, unequal access, and employability gaps.
Synthesis of concerns reported in prior studies and discussions in the literature (review article); no new empirical data provided.
Agents do not reroute along the methodological axes humans use to bias their estimates.
Analytic comparison showing that when biased prompts are applied, agents change methodological choices but not along the particular methodological axes (choices) that human analysts exploit to bias estimates.
Benchmarking multiple state-of-the-art open and closed source VLMs on our evaluation framework demonstrates substantial limitations in current engineering reasoning capabilities.
Empirical claim based on the paper's benchmarking experiments using the EngVQA dataset and the 8-stage framework (models and detailed results not provided in the excerpt).
Existing benchmarks primarily evaluate final answers and provide limited assessment of intermediate reasoning processes.
Claim in paper contrasting EngVQA's process-oriented evaluation with prior benchmarks (literature/benchmark review claim; no specific benchmarks or quantitative comparison provided in the excerpt).
Failures in engineering reasoning by AI systems may produce physically invalid yet superficially plausible solutions, posing risks for engineering education, scientific assistance, and technical decision-making.
Argumentative claim in the paper highlighting potential risks of reasoning failures in high-stakes engineering contexts (motivational/background statement).
Datasets are rarely standardized or shared.
Review synthesis and commentary across included studies and supplementary documents indicating limited data standardization and sharing.
Agents performed more weakly on a task requiring novel bioinformatics reasoning.
Reported ABC-Bench results indicating relatively lower agent scores on the task characterized by novel bioinformatics reasoning (authors' summary in the abstract).
Penerapan AI menimbulkan isu etika dan keamanan data yang memerlukan tata kelola AI yang bertanggung jawab.
Sistematis studi literatur yang menelaah 33 sumber ilmiah, laporan lembaga internasional, dan kebijakan terkait (n=33).
AI meningkatkan risiko pengangguran pada sektor yang pekerjaannya bersifat rutin.
Sistematis studi literatur yang menelaah 33 sumber ilmiah, laporan lembaga internasional, dan kebijakan terkait (n=33).
Penerapan AI menyebabkan kesenjangan keterampilan (skill gap) antara kebutuhan pasar dan kemampuan tenaga kerja.
Sistematis studi literatur yang menelaah 33 sumber ilmiah, laporan lembaga internasional, dan kebijakan terkait (n=33).
Neither MCP nor A2A defines the shared workspace in which humans and agents perform accountable work together.
Analytical claim by the authors contrasting existing standards with the missing specification of a shared human-agent workspace; no empirical evaluation provided.
In current practice the human judgement is recorded, if at all, in application code, chat threads, ticket comments, and tribal memory.
Descriptive statement about current recording practices; presented without empirical study or counts in the provided text.
The technical surface for this collaboration remains weakly specified.
Asserted by the authors as an assessment of current technical standards and interfaces; no audit or measurement cited in the provided text.
Macro-level correlation between Frey-Osborne (2013) and Eloundou-era rankings is Spearman rho = -0.750, p = 0.020 (against the original Oxford Martin appendix), indicating inversion.
Reported Spearman correlation and p-value comparing macro-level rankings between the original Frey-Osborne appendix and the paper's Eloundou-era results.
Tool-Mediated Physical (macro M2) has mean OAI = 0.054.
Reported macro-level mean OAI computed after projecting DWA OAI values into the 7-macro typology.
Automated evaluators on which the community currently relies -- LLM-as-a-judge, artificial metrics, and even state-of-the-art (SOTA) models -- agree weakly with expert judgment.
Comparisons/correlations between automated evaluator outputs and human expert ratings showing weak agreement.
LLMs falter most in pluralistic fields like the social sciences that demand context-aware interpretation and evolving theories.
Field-level performance comparisons showing lower agreement/quality of LLM-generated ideas in pluralistic fields (social sciences) relative to more constrained fields (life sciences etc.).
Senior social scientists are the harshest critics, and their skepticism is well-earned.
Subgroup analysis by seniority and field indicating senior social scientists give lower ratings (more critical) than other subgroups; interpretation provided by authors.
Non-reasoning LLMs collapse into a narrow 'hivemind' of similar ideas.
Comparative analysis of idea outputs from different LLM classes showing reduced diversity/similarity concentration for non-reasoning models (as described in results).
Board power disparity weakens the positive relationship between AI competitive actions and operational efficiency.
Interaction tests in the authors' empirical models using governance measures (power disparity) and NLP-identified AI actions from S&P 500 firms' press releases (2010–2022); reported as a negative conditional effect on operational efficiency.
This regulatory pressure creates a direct conflict between multi-stakeholder transparency and corporate data privacy.
Paper's conceptual argument describing a tension between transparency requirements and proprietary data protection; no empirical study provided.
Regulatory compliance demands have surpassed the capacity of manual corporate reporting.
Assertion in paper (conceptual observation about reporting capacity); no empirical measurement or sample size reported.