Evidence (8974 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
10085 claims
Filter claims →
Productivity
8974 claims
Filtered →
Governance
8062 claims
Filter claims →
Human-AI Collaboration
7749 claims
Filter claims →
Org Design
5057 claims
Filter claims →
Innovation
4896 claims
Filter claims →
Labor Markets
4088 claims
Filter claims →
Skills & Training
3372 claims
Filter claims →
Inequality
2377 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 | 882 | 244 | 117 | 1097 | 2424 |
| Governance & Regulation | 1010 | 469 | 229 | 135 | 1875 |
| Organizational Efficiency | 977 | 235 | 149 | 90 | 1462 |
| Technology Adoption Rate | 781 | 299 | 143 | 128 | 1362 |
| Research Productivity | 506 | 155 | 74 | 363 | 1110 |
| Output Quality | 555 | 219 | 71 | 70 | 915 |
| Decision Quality | 395 | 200 | 95 | 54 | 751 |
| Firm Productivity | 523 | 67 | 101 | 27 | 724 |
| AI Safety & Ethics | 262 | 309 | 75 | 36 | 688 |
| Market Structure | 195 | 201 | 135 | 30 | 566 |
| Task Allocation | 248 | 77 | 96 | 38 | 464 |
| Innovation Output | 300 | 34 | 55 | 20 | 411 |
| Skill Acquisition | 207 | 75 | 65 | 21 | 368 |
| Employment Level | 138 | 67 | 119 | 24 | 350 |
| Fiscal & Macroeconomic | 156 | 80 | 53 | 33 | 329 |
| Task Completion Time | 211 | 38 | 13 | 16 | 280 |
| Firm Revenue | 183 | 52 | 29 | 5 | 270 |
| Consumer Welfare | 131 | 77 | 48 | 13 | 269 |
| Inequality Measures | 50 | 141 | 54 | 9 | 254 |
| Worker Satisfaction | 104 | 85 | 25 | 13 | 227 |
| Error Rate | 87 | 112 | 11 | 5 | 215 |
| Automation Exposure | 69 | 69 | 37 | 20 | 198 |
| Wages & Compensation | 102 | 49 | 31 | 11 | 193 |
| Team Performance | 115 | 30 | 30 | 11 | 187 |
| Regulatory Compliance | 88 | 74 | 17 | 7 | 186 |
| Training Effectiveness | 109 | 22 | 14 | 21 | 168 |
| Developer Productivity | 116 | 21 | 15 | 8 | 161 |
| Job Displacement | 12 | 92 | 26 | 1 | 131 |
| Hiring & Recruitment | 57 | 12 | 9 | 5 | 83 |
| Skill Obsolescence | 6 | 59 | 10 | 2 | 77 |
| Social Protection | 43 | 17 | 8 | 2 | 70 |
| Creative Output | 35 | 21 | 9 | 4 | 70 |
| Labor Share of Income | 18 | 23 | 17 | 1 | 59 |
| Worker Turnover | 15 | 16 | — | 4 | 35 |
| Industry | — | — | — | 1 | 1 |
Productivity
Remove filter
This study provides the first cross-class synthesis covering raw materials, work-in-process, and finished goods within a unified evaluative framework, positioning machine learning and deep reinforcement learning methods alongside classical policy families and quantifying the boundary conditions for each approach.
Author-stated theoretical contribution and scope of the review (coverage of raw materials, WIP, finished goods and methods).
A random-effects model estimated by restricted maximum likelihood was applied to pool percentage cost-reduction effect sizes across 18 studies admissible to quantitative synthesis.
Methods reported in the paper: random-effects meta-analysis using REML across 18 studies eligible for quantitative pooling.
A systematic review and meta-analytic synthesis of 31 peer-reviewed studies published between 2004 and 2025 was conducted following the PRISMA 2020 protocol.
Study methods reported in the paper: systematic review following PRISMA 2020; sample of 31 peer-reviewed studies dated 2004–2025.
The study uses PyQu to quantify changes across five quality attributes for Python code.
Methodological description: application of PyQu (an ML-based quality assessment tool for Python) to measure five quality attributes before and after refactoring edits.
From the observed diffs, we derive a taxonomy of 24 recurring change operations.
Manual/automated analysis of diffs from the studied agentic refactoring PRs to identify and categorize recurring change operations into a 24-item taxonomy.
We will release the reanalysis pipeline to support replication.
Authors' statement of intent in the paper to release code/pipeline for replication.
In offensive cybersecurity, the marginal benefit of Skills collapses: the spread between the no-Skills and full-Skills conditions is only 8.9 percentage points (p = 0.71, χ²; p = 0.25, Cochran–Armitage trend test; five of six pairwise Cohen's h values fall below the 0.2 small-effect threshold).
Statistical re-analysis of the 180-run CTF study comparing no-Skills vs full-Skills conditions: reported spread = 8.9 percentage points; reported p-values from χ² and Cochran–Armitage trend tests; reported Cohen's h comparisons.
Those four documentation conditions correspond almost exactly to a No-Skills, Experiential-Skills, Curated-Skills, and Comprehensive-Skills ablation.
Authors map the four documentation-line-count conditions from the re-analyzed study to skill-ablation categories (No/Experiential/Curated/Comprehensive) as part of their interpretive re-analysis.
We re-analyze a recently published 180-run controlled study of an MCP-grounded autonomous Capture-the-Flag (CTF) agent under four documentation conditions (55, 1,478, 1,976, and 4,147 lines).
Authors' re-analysis of an existing controlled study consisting of 180 runs and four documentation conditions with the stated line counts; this is a descriptive claim about the re-analysis dataset and experimental conditions.
Across 660 trials with Claude Code, code cleanliness does not change the agent's pass rate.
Empirical evaluation: 660 trials run using Claude Code on the minimal-pair repos with hidden tests; reported comparison of pass rates between clean and messy repo variants showing no change.
Each output is scored with a unified rubric covering task completion, correctness, compliance, and clarity.
Measurement approach stated in the abstract (unified rubric with listed dimensions).
The study uses three LLM systems: ChatGPT, Claude, and Grok.
Method description in the paper's abstract naming the three LLMs evaluated.
The evaluation covers four task types: summarization, planning, explanation, and coding.
Method description in the paper's abstract listing the four task types used for evaluation.
The study compares three prompt conditions: a raw prompt, a checklist-improved prompt, and a clarifying-question prompt.
Experimental design described in the paper (three prompt conditions stated in the abstract).
Large language models (LLMs) are widely used for open-ended tasks.
Stated as background/context in the paper's introduction; no quantitative data reported in the abstract.
We conduct extensive experiments on public datasets, in simulated auction environments, and through large-scale online deployment on Taobao.
Statement of experimental methodology describing the types of evaluations performed (public datasets, simulated auctions, and online deployment).
Reported empirical values are transformed through transparent indicators such as relative growth, CAGR, growth multipliers, stock-flow ratios, concentration ratios, and HHI.
Methodological description and application in the paper listing these specific indicators used to summarize public data on AI investment, adoption, robots, compute, and labour-market reallocation.
The study uses a conceptual-empirical quantitative diagnostic design rather than a causal econometric model.
Explicit methodological statement in the paper describing the design choice and rejecting causal econometric modeling in favor of diagnostics using public institutional data and transparent indicators.
The agentic economy is not yet a completed global order, but its transition pressure is measurable enough to require a distinct economic vocabulary, reproducible diagnostics, and future sector-level measurement.
Synthesis of diagnostic indicators (AI investment/adoption trends, robot stock, compute-energy coupling, labour reallocation measures) showing measurable transition pressures; conclusion drawn from the conceptual-empirical diagnostic.
Following PRISMA 2020 guidelines, searches across Google Scholar, Web of Science, Scopus, ScienceDirect, and CNKI yielded 1,562 initial records, of which 21 studies published between 2019 and 2026 met inclusion criteria.
Methodological description of the systematic literature review reported in the paper: initial records = 1,562; included studies = 21; publication years 2019–2026.
Small and medium-sized enterprises (SMEs) constitute over 98.5% of businesses in many economies including China.
Descriptive statistic reported in the paper's background/intro; source of the statistic not specified within the summary provided.
This study analyzes developments through April 2026.
Explicit timeframe statement in the paper's summary/introduction.
The authors provide source code for their framework on GitHub to encourage further research.
Statement in the paper that the source code is available on GitHub; verifiable by visiting the repository (link not provided in the excerpt).
Heuristics such as TSP and PNN are commonly used as inexpensive approximations for customer trajectories.
Descriptive claim about common practice cited in the paper; used as motivation for proposing the RL approach (no quantitative survey evidence provided in the excerpt).
We conducted a randomized controlled experiment in which participants—analogs of early-career knowledge workers—were assigned to self-study a technical domain using either traditional resources or large-language-model (LLM) assistance.
Statement of experimental design in the paper (randomized controlled experiment assigning participants to either traditional resources or LLM assistance; participants described as analogs of early-career knowledge workers).
Behavioral findings from any single framework therefore warrant cross-configuration validation before being claimed as general.
Prescriptive conclusion derived from the observed cross-configuration heterogeneity in the paper's empirical results.
Framework identity accounts for more of the between-configuration variation than LLM family: for mean turns, framework explains 64% of the between-configuration variance against the LLM's 10%.
Variance decomposition / explained-variance analysis reported for 'mean turns' across configurations (reported percentages: 64% vs 10%).
The analysis separates framework effects from LLM effects by holding each layer fixed in turn and measures one behavior–outcome effect per configuration to examine agreement across configurations.
Methods description in the paper: experimental design holding LLM or framework fixed to disentangle effects.
This study analyzes 64,380 SWE-bench runs from 126 agent configurations spanning 43 frameworks, where each configuration pairs an LLM with a framework supplying tools and workflow.
Dataset and experimental design reported in the paper: 64,380 runs; 126 configurations; 43 frameworks.
The formal semantics and proof-checked admission model are specified and under active development, with evaluation of the verified core reserved for future work.
Author statement in the paper about the current development status and that evaluation of the verified core is deferred to future work.
Reward is non-positive in the CybORG CAGE-2 environment, so all configurations operate in a failure-mitigation mode.
Environment specification reported in the paper (CybORG CAGE-2 modeled as a POMDP with non-positive reward structure).
The evaluation spanned five model families, six models, and twelve configurations, totaling 3,475 episodes with token-level cost accounting.
Methods description in the paper reporting the experimental design and sample counts.
These factors evolve over time, have inter-dependencies across multiple resource dimensions, and generally do not lend themselves to closed-form analysis.
Methodological observation motivating simulation/sequence-based evaluation; asserted in the paper's rationale.
The study used a structured questionnaire (five-point Likert) administered to employees in AI-enabled organizations across various sectors and analyzed the data using SPSS (descriptive statistics, reliability analysis, correlation analysis, regression analysis).
Methods section summary provided in the paper (survey instrument description and analytical techniques).
We evaluate PRISM across 35 enterprise conversational agents over a three-week deployment period on the Yellow.ai V3 platform.
Statement in abstract: evaluation across 35 agents over a three-week deployment on Yellow.ai V3 platform (empirical deployment described).
The study investigates the non-linear impact of AI on economic growth in 19 G20 countries (2005–2023) using the Generalized Method of Moments (GMM) with both linear and quadratic models.
Methodological description provided in the paper: panel dataset covering 19 G20 countries over 2005–2023 and estimation via GMM with linear and quadratic specifications.
IIQ is positioned as a deployment-oriented measurement framework: a formal proposal for tracking AI embedding in workflows, not a direct measure of model capability or a substitute for causal productivity evaluation.
Explicit positioning statement in paper: authors state scope and limits of IIQ as deployment/usage metric rather than capability or causal productivity estimator (conceptual/positioning).
The study uses World Bank Enterprise Survey firm-level data from 2007 to 2024 and employs feasible generalized least squares (FGLS), robust ordinary least squares (OLS), and high-dimensional fixed effects (HDFE) linear regression techniques.
Direct methodological statement in the paper's abstract/summary. This is a descriptive factual claim about data and methods.
AI deployment has limited effects on retrial rates.
Same randomized field experiment; retrial rates (repeat customer contacts) were measured and reported as showing limited/no substantive change under AI deployment.
The findings are based on India-focused samples.
Paper explicitly notes the sample/context is India-focused.
PRIF was developed and validated using mixed-method design: interviews with 30 risk advisors, case studies, and analysis of 30 forensic reports, with validation via thematic coding, risk metrics, and Delphi panel refinement.
Reported methods in the paper: mixed-method design including 30 risk advisor interviews and analysis of 30 forensic reports; validation methods named (thematic coding, risk metrics, Delphi panel).
The overall impact of prompt design on readability remains limited.
Reported results from prompt-dimension experiments indicating that while some prompt elements influence readability, the aggregate effect size of prompt engineering on overall readability was limited.
Current LLMs produce code with overall readability comparable to human-written code.
Comparison of readability scores (from the paper's readability model) between LLM-generated code and human-written code across 5,869 scenarios; reported summary conclusion that overall readability is comparable.
The analysis proceeded through within-case coding and cross-case pattern matching across five dimensions: intelligence source, AI mechanism, decision domain, economic implication, and boundary condition.
Method section describing coding and analytical procedures applied to the archival corpus across the four cases.
The empirical corpus comprises annual reports, 10-K filings, earnings releases, and official corporate materials published mainly between 2024 and 2026, complemented by recent peer-reviewed literature.
Paper's data description listing document types and time window for archival evidence; number of documents not enumerated.
The study adopts a qualitative comparative multiple-case design using four theoretically sampled cases: Walmart, Unilever, Sprinklr, and DoubleVerify.
Methodological statement in the paper describing case selection and study design.
The study includes Natural Language Processing (NLP) analysis of 5 million consumer contacts.
Methodological statement in the paper specifying the NLP data volume.
The study includes surveys of 800 marketers.
Methodological statement in the paper specifying the survey sample size.
The study includes AI adoption audits from 120 organizations.
Methodological statement in the paper specifying the audits sample size.
LLM-generated solutions contain roughly the same number of ideas as participant-generated solutions.
Comparative analysis of idea counts within solutions reported in the paper; phrased as 'roughly the same number of ideas' (no numeric effect size provided in the abstract).