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
Falsifiability condition for Ghost GDP: If monetary velocity does not decline (or instead rises) as the labor share falls, the Ghost GDP channel is unsupported by the data.
Explicit falsification condition provided in the paper based on the model link labor share -> velocity -> consumption; suggested empirical test using monetary-velocity proxies and labor-share series from FRED.
Empirically, top-quintile households account for roughly 47–65% of U.S. consumption.
Calibration and reported quantitative scenarios in the paper using U.S. consumption concentration data (constructed from U.S. consumption/income micro- and macro-data sources referenced in the methods section).
Because the sample is small and purposive and the design is qualitative, insights are rich but not statistically representative or quantified across the broader research landscape.
Authors' stated study limitations in the paper acknowledging small purposive sample (n=16) and qualitative design.
The study's data come from semi-structured interviews with 16 expert practitioners across biosecurity, cybersecurity, education, and labor.
Study methods reported in the paper: qualitative data source explicitly stated as 16 semi-structured interviews across listed domains.
The authors released their code and data for reproducibility at https://github.com/blocksecteam/ReEVMBench/.
Statement in the paper indicating public release of code and dataset at the provided GitHub URL.
Crystallization Efficiency (CE) is defined as Useful_Crystallized_Knowledge / (Human_Effort × Time).
Operational formalism and metric definitions presented in the paper (explicit formula provided). This is a proposed metric, not an empirically validated measure.
The paper proposes operational patterns (Dual-Workspace Pattern separating live interaction workspace and persistent knowledge workspace) and a Spiral Development Model (iterative interaction → crystallization → validation → redeployment).
Operational framework section describing patterns and workflows; illustrated in the case study implementation.
The Knowledge Crystallization Cycle formalizes operations (extract, synthesize, validate, integrate) and proposes efficiency and quality metrics including Crystallization Efficiency (CE), Fidelity, Reuse Rate, and Freshness/Volatility Score.
Operational formalism section of the paper presenting metric definitions and proposed calculations (e.g., CE = Useful_Crystallized_Knowledge / (Human_Effort × Time)). These are proposed metrics, not validated at scale.
The paper introduces a Three-Layer Cognitive Architecture that organizes agent knowledge by volatility and degree of personalization (stable/core knowledge; institutionalized heuristics/patterns; volatile/session-level tacit details).
Architectural specification presented in the paper (conceptual design document). No experimental validation beyond the illustrative case study.
Nurture-First Development (NFD) reframes agent creation from a one-time engineering task into a continuous, conversational growth process.
Conceptual formalization in the paper (architectural and operational descriptions). No large-scale empirical test reported; supported by theoretical argumentation and illustrative examples.
Findings are based on a student sample rating decontextualized messages, so external validity to industry communication or real project logs is uncertain and requires replication.
Study sample consisted of 81 students in team-based software projects labeling decontextualized statements; authors explicitly note this limitation as a caveat.
Many apparent correlations between predictors and sentiment labels do not remain significant after global multiple-testing correction.
Correlation analyses across many predictors with explicit application of multiple-testing correction procedures; many initial signals failed to survive correction.
The paper does not provide quantitative estimates of time saved per report, cost reductions, or effects on employment/wages; such economic impacts remain to be quantified.
Caveats noted in the paper: absence of quantitative estimates for time/cost/employment effects and a call for field trials and economic modeling. This is explicitly stated in the summary.
The paper used a clinically grounded, multi-level evaluation framework that separately assessed raw AI drafts (automatic metrics + clinician review) and radiologist-AI collaborative final reports (how radiologists edit and downstream clinical effects), including comparisons across radiologist experience levels.
Methodology section summarized in the paper: multi-level assessment covering AI drafts and radiologist-edited collaborative reports; combination of automatic metrics and radiologist-/clinician-centered evaluations; experience-level stratified analyses (novice/intermediate/senior).
CBCTRepD is a report-generation system trained on this curated paired dataset to produce bilingual CBCT radiology draft reports intended for radiologist-in-the-loop (co-authoring) workflows.
System description in the paper: CBCTRepD built using the curated dataset; authors state purpose is to generate clinically usable drafts for radiologist editing. (Model architecture and training hyperparameters are not specified in the provided text.)
The authors curated a paired CBCT–report dataset of approximately 7,408 CBCT studies covering 55 oral and maxillofacial disease entities that is bilingual and includes diverse acquisition settings.
Data curation described in the paper: stated dataset size (~7,408 studies), coverage of 55 disease entities, bilingual reports, and inclusion of a range of acquisition settings to increase heterogeneity and clinical realism. (Exact languages, provenance of studies, and dataset split details are not specified in the provided text.)
Evaluation was performed on five different material setups.
Experimental evaluation described in the summary: performance reported as averaged across five material setups. The summary does not list per-setup names or trial counts.
The simulation models samples as collections of spheres with per-sphere procedurally generated dislodgement-force thresholds derived from Perlin noise to introduce spatial heterogeneity and diversity.
Simulation/modeling description in the paper: discrete-sphere representation of sample; each sphere assigned a dislodgement threshold; spatial variation produced via Perlin noise. This is a concrete modeling choice reported in the methods.
The paper uses a mixed-methods approach combining a systematic literature review with an empirical practitioner survey to assess perceptions, adoption, and impact of AI-driven tools.
Methodological statement in the paper; survey design covers tool usage, perceived benefits, challenges, and expectations.
Empirical work (experiments and measurements) is needed to quantify how much value interpretive traces add to downstream outputs, how RATs affect platform incentives, and what governance frameworks fairly allocate resulting rents.
Concluding recommendation in the paper stating the research gaps; not an empirical claim but a stated need.
The current presentation of RATs is speculative and illustrative; empirical validation, scalability, and ethical safeguards remain to be developed.
Limitations section of the paper explicitly states the speculative nature and lack of empirical evaluation.
Implementation of RATs requires instrumentation at the browser/platform level or via plugins and must address privacy/consent, storage/ownership, sharing controls, and interoperable trace formats.
Design and implementation considerations enumerated in the paper; this is a requirements statement rather than an empirical claim.
Analytical approaches compatible with RATs include sequence/trajectory mining, network analysis of associations/co-read graphs, embedding/clustering of trajectories, qualitative inspection of reflections, and experimental (A/B or RCT) evaluation of downstream effects.
Methods section of the paper listing suggested analytical techniques; these are proposed methods rather than applied analyses.
The approach shifts some computational burden to obtaining MCMC samples of the parameter posterior, requiring access to (or ability to compute) MCMC samples before surrogate training.
Method description: training data are MCMC-drawn parameter vectors; the paper notes this practical requirement and trade-off (MCMC cost vs. avoiding repeated expensive forward-model evaluations).
Because this is a conceptual/systems-architecture paper, it does not present new empirical performance benchmarks.
Explicit statement in the paper's Data & Methods section that no new empirical benchmarks are presented.
The evaluated models consist of an MLP baseline and a GNN tailored to exploit relational/spatial structure among beams/antennas.
Model descriptions provided in the methods section: two supervised-learning architectures (MLP and GNN) used for beam prediction experiments.
Using Federated Learning (FL) with orbital planes as distributed learners and HAPS for aggregation avoids centralization of raw channel data.
Method description: federated-learning architecture with clients mapped to orbital planes and HAPS performing coordination/aggregation; explicitly states no central pooling of raw channel samples.
DPS was empirically evaluated across diverse reasoning domains (mathematical reasoning, planning, and visual-geometry) to test generality.
Paper reports experiments on those three categories of tasks; they are listed as the evaluated tasks in the methods/experiments section.
DPS uses the inferred per-prompt state distributions as a predictive prior to select prompts estimated to be most informative, avoiding exhaustive candidate rollouts for filtering.
Method and selection mechanism described: predictive prior ranking/filtering replaces rollout-heavy candidate evaluation. (Procedure described in paper; empirical comparisons reported.)
Dynamics-Predictive Sampling (DPS) models each prompt’s "extent of solving" under the current policy as a latent state in a dynamical system (a hidden Markov model) and performs online Bayesian inference on historical rollout reward signals to estimate that state.
Methodological description in the paper: DPS uses an HMM representation of per-prompt solving progress and applies online Bayesian updates using past rollout rewards. (No numerical sample size needed for this modeling claim.)
The paper does not present large-scale empirical validation; its evidence is primarily theoretical exposition, a constructed illustrative example, and a literature survey.
Explicit description of methods and data in the paper (analysis type: theoretical exposition + illustrative example; no experimental sample reported).
The authors recommend specific measurement metrics and empirical research priorities (e.g., MAPE, stockout frequency, inventory turns, lead times, fill rates, total supply chain cost, service-level volatility, resilience measures; causal studies like diff-in-diff or randomized interventions).
Explicit recommendations in the paper's measurement and research agenda sections.
The study's small sample size and qualitative design limit external generalizability and prevent causal effect size estimation; potential selection and reporting biases exist due to purposive sampling and interview-based data.
Authors explicitly state these limitations in the paper's limitations section.
The study is a qualitative multi-case study of five medium-to-large organizations, using semi-structured interviews across procurement, production planning, inventory management, and distribution, analyzed via cross-case comparison.
Methods section description provided by the authors (sample size n = 5, sectors, interview-based primary data, cross-case analysis).
The paper is primarily theoretical and prescriptive: it synthesizes literature and proposes a framework and design guidelines rather than reporting large-scale empirical datasets or causal identification of economic outcomes.
Meta-claim about the paper's methods explicitly stated in the Data & Methods summary; based on the paper's methodological description.
Key measurable outcomes to assess Human–AI teams include accuracy/efficiency, robustness to novel cases, decision consistency, trust/misuse rates, training costs, and inequity indicators.
Prescriptive list of metrics offered by the authors as part of the research agenda and evaluation guidance; not empirically derived from a dataset in the paper.
Empirical evaluation strategies for Human–AI teams should include randomized interventions, field trials, lab experiments, phased rollouts (difference-in-differences), and structural models that allow interaction terms between human skill and AI quality.
Methodological recommendation in the paper; suggested study designs rather than implemented analyses.
Research priorities include empirical measurement of task‑level automation rates, firm and industry productivity effects, wage impacts across occupations, and diffusion patterns.
Paper's stated research agenda and identification of measurement gaps; based on methodological critique of current evidence base.
Measuring these productivity gains will be challenging because quality improvements, faster iteration, and creative outputs are harder to price/observe than lines of code.
Methodological argument about measurement difficulty; based on conceptual considerations, not empirical validation.
Measuring AI's economic impact requires new metrics that account for decision-value uplift, reduced tail-risk exposures, and dynamic gains from continuous learning; causal identification will require experiments or staggered rollouts.
Methodological recommendation backed by conceptual discussion of measurement challenges; no implementation of such measurement approaches is reported in the paper.
Performance and evaluation should be measured using forecast accuracy, decision lift/value added, latency, and false positive/negative rates.
Paper-prescribed evaluation metrics; presented as recommended practice rather than derived from empirical testing within the paper.
Core AI techniques for these frameworks include supervised/unsupervised ML, NLP for unstructured text, anomaly detection for control/transaction monitoring, and reinforcement/prescriptive models for recommendations.
Methodological claim listing standard ML/NLP/anomaly-detection techniques and prescriptive approaches; statement of methods rather than an empirical comparison of alternatives.
Next‑gen frameworks use large-scale structured (transactions, ledgers, KPIs) and unstructured sources (reports, news, contracts, call transcripts) to power models.
Descriptive claim listing data types the paper recommends; presented as design input requirements rather than empirically validated data-integration projects.
There is a need for quantitative studies and microdata on firm-level RM practices, AI adoption, and performance outcomes to measure effect sizes and causal pathways.
Stated research gaps and limitations in the review (lack of primary empirical quantification; heterogeneity across contexts).
The review's conclusions are limited by reliance on published literature (potential bias toward successful implementations), lack of primary empirical quantification (no effect sizes), and heterogeneity across organizational contexts limiting direct generalizability.
Explicit limitations stated in the paper summarizing scope and method (qualitative literature review, secondary evidence only).
Heterogeneity in system designs and deployment contexts complicates cross-site comparisons.
Limitations section and observed variation in platform architectures, degrees of automation, and governance across sites reported via descriptive data and interviews.
Non-random selection of institutions limits causal inference and external generalizability of the study's findings.
Study limitations explicitly state non-random site selection and heterogeneous deployments; methodological note that causal claims are constrained.
The study uses a quantitative, cross-sectional survey-based research design of managers and educational administrators and employs descriptive statistics, correlation, and regression analyses.
Methods described in the summary explicitly state research design and analytical techniques; this is a methodological claim rather than an empirical substantive finding. (Sample size not provided in summary.)
Estimation/calibration, stability assessment, and global sensitivity methods used: parameters calibrated/estimated on 2016–2023 data; equilibrium located; Jacobian eigenvalues computed for local stability; variance-based global sensitivity analysis performed over parameter space.
Methods section: description of parameter estimation/calibration, equilibrium computation, Jacobian-based stability analysis, and variance-based global sensitivity analysis.
The main empirical conclusions are based on a short annual panel (2016–2023) and a stylized aggregate interaction model; results should be interpreted with caution due to potential omitted variables, aggregation bias, and limited sample size.
Explicit limitations listed in the paper: short time series (eight annual observations), national aggregate data, simplified model structure, no firm/sector heterogeneity, possible endogeneity/measurement issues.