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 |
Existing evidence is time-sensitive and heterogeneous: rapidly evolving models, heterogeneous study designs, and many short-term lab/microtask studies limit direct comparability and long-run inference.
Meta-observation from the review: documented methodological limitations across the literature (variation in models, tasks, metrics; prevalence of short-term studies).
Real‑time and LLM‑based methods improve responsiveness but raise governance, transparency, and reproducibility challenges that BLS must manage (audit trails, uncertainty communication).
Operational tradeoff discussion in the paper identifying governance risks; no case studies or incident analyses provided.
Distinguishing automation versus augmentation using causal methods changes policy responses (e.g., income support versus reskilling).
Policy implication drawn from conceptual separation of substitution and complementarity effects; logical inference rather than empirical demonstration in the paper.
Methodological caveats across the literature (heterogeneity of tasks/measures, publication bias, short-term studies) limit the generalizability of current findings.
Meta-level critique within the synthesis noting study heterogeneity, likely publication/short-term biases, and variable domain-specific performance dependent on user expertise and workflows.
Standard productivity metrics are likely to undercount the value generated by AI-augmented ideation; quality-adjusted measures of creative output are required.
Measurement critique based on the mismatch between existing productivity statistics and the kinds of upstream idea-generation gains observed in empirical studies; supported by the review's methodological discussion.
The authors were able to fully reproduce the reported results for 49% of CHI papers that had publicly shared study data and analysis code.
Empirical reproduction attempts performed by the authors on the population of CHI papers that publicly shared study data and analysis code (sample defined as 'all CHI papers that had publicly shared study data and analysis code' — exact number/time window not specified in the summary).
Realized value from AI methods (ML, predictive analytics, anomaly detection, XAI) is conditional: these technical methods deliver capabilities only when combined with strong data governance, standardized processes, and change management.
Thematic synthesis across the systematic review (2020–2025) showing repeated case-study and practitioner-report evidence that technical gains failed to scale without governance, process standardization, and organizational change efforts.
Evaluation of the equivalency system should use metrics such as concordance between claimed competencies and verified inputs, predictive validity versus labor-market integration outcomes, and false positive/negative rates in automated decisions.
Methodological recommendation in the paper outlining specific evaluation metrics; this is a prescriptive claim (no empirical implementation reported).
The hybrid estimator (GA+SQP) is computationally more intensive than single-stage MLE/local optimization, implying a trade-off between estimation reliability and runtime cost.
Reported runtime and computational cost comparisons in estimation experiments: the paper notes longer runtimes for GA+SQP versus standard optimizers while documenting improvements in objective values and convergence behavior.
Despite laboratory and pilot successes, many engineered bioprocesses remain at bench or pilot scale and require techno‑economic validation before industrial competitiveness can be established.
Review aggregate noting scale and validation status of case studies (many reported at lab or pilot fermenter scale) and explicit references to the need for TEA and LCA for industrial assessment.
Results and implications are limited by the sample and context: evidence comes from law students on a single issue-spotting exam using one brief training intervention, so generalizability to experienced professionals, other tasks, or other models is untested.
Authors’ reported sample (164 law students) and explicit caution about generalizability in the study summary; the intervention and outcome are specific to one exam and one ~10-minute training.
Some mechanism-specific estimates are imprecise due to the sample size; confidence intervals for those estimates are wide.
Authors report wide confidence intervals for mechanism decomposition (principal stratification) results based on the randomized sample of 164 students.
Overall, the protocol reframes AI governance in finance as a rights‑centered institutional design problem with direct economic consequences for market structure, credit allocation, compliance costs, and incentives shaping AI model development.
High-level synthesis claim made by the author, supported by the corpus audit (~4,200 texts), 12 years of legal research, doctrinal/comparative analysis, and the economics implications section.
Machine learning, recommender systems, NLP, computer vision, causal inference, reinforcement learning, federated learning/differential privacy/secure computation, and algorithmic governance tools are co-deployed in modern ad-tech.
Technical methods inventory drawn from literature and industry reports; no new experimental sample reported.
Personalization now spans data infrastructures, real-time bidding markets, recommender systems, creative generation, attribution pipelines, privacy tools, and governance regimes — all tightly coupled.
Survey of technical components and industry practice (system-analysis level); descriptive synthesis of common ad-tech stacks and interdependencies; no single-sample empirical audit provided.
AI has transformed personalized digital advertising from a narrow prediction task into a complex socio-technical infrastructure.
System-level conceptual analysis and literature synthesis presented in the paper; no single empirical dataset or sample size reported (review of industry components such as RTB, recommender systems, identity graphs).
Applying differential privacy to model updates provides a bounded formal guarantee on information leakage, but DP noise budgets and communication constraints create accuracy and latency trade-offs that must be managed.
Analytical treatment of DP's impact on learning (trade-off modeling) and qualitative simulation examples showing accuracy degradation under DP noise; no numeric privacy-utility curves from field deployments provided.
There is no consensus in the literature on net job effects — studies diverge on whether AI produces net job gains.
Direct finding from the review: the 17 peer‑reviewed studies produce heterogeneous results on net employment impacts (some positive, some negative, some neutral).
Effects of AI adoption are heterogeneous across industries, firm sizes, regions, and worker characteristics (education, experience, occupation).
Microdata and firm-level studies exploiting cross-sectional and panel variation, quasi-experimental designs leveraging differential adoption across firms/regions, and comparative institutional analyses showing variation by context.
The effects of K_T adoption are heterogeneous across industries, firms, countries, and cohorts — early adopters and capital-rich firms/countries gain most — implying important transition dynamics for political economy.
Cross-country comparisons, industry- and firm-level panel heterogeneity analyses, and case studies demonstrating variation in adoption timing and gains; model simulations emphasizing transition path dependence.
Aggregate productivity (output per worker or per unit of inputs) can rise while labor’s share and employment decline due to substitution toward K_T.
Macro growth-accounting exercises decomposing output growth into contributions from labor, traditional capital, and technological capital; model simulations showing productivity gains coexisting with falling labor shares under substitution elasticities.
HCI research explores how people rely on AI advice, but it largely overlooks replicating realistic decision-making scenarios.
Finding from the paper's analytical review examining decision-making tasks used in prior HCI studies and assessing their validity relative to application-grounded contexts.
Recent empirical studies show critical concerns that people over-rely on AI advice without analytically engaging with it.
Summary claim based on the paper's analytical review of recent empirical studies in the human-AI reliance literature (number of studies not specified in abstract).
Developing countries face structural asymmetries, infrastructural deficits, and human capital gaps that constrain algorithmic performance even where technical sophistication is high.
Cross-study synthesis noting recurrent findings in the reviewed literature that institutional and infrastructural constraints in developing economies limit algorithmic performance (reported across the 68 studies).
AI intensifies market concentration, reinforcing winner-takes-most dynamics through data-driven network effects.
Synthesis of market-structure and industrial-organization studies in the SLR reporting evidence of increased concentration and network/data advantages favoring incumbents.
AI displaces routine occupations.
Synthesis of empirical and modeling studies within the 78-study SLR reporting occupational/task-level substitution effects for routine activities.
Overall, LLM assistance did not produce measurable advantages for human-supervised verification and was associated with reduced detection of major errors, meaning expert human judgment remains indispensable for reliable empirical verification.
Synthesis of experimental findings comparing human-only, AI-assisted, and AI-led conditions; summary concludes no measurable advantages for AI-assistance and reduced major-error detection, and emphasizes continued importance of human expertise.
AI-led teams detected fewer errors across all categories than human or AI-assisted teams.
Reported error-detection comparisons across experimental conditions; summary states AI-led teams detected fewer errors across all categories.
AI-led (autonomous ChatGPT with minimal human oversight) teams achieved only a 37% reproduction rate.
Reported reproduction outcome for AI-led condition in randomized experiment; summary gives 37% reproduction rate for autonomous AI teams.
Verifying results of published social sciences research is expensive, costing hundreds of dollars per study.
Authors' statement in paper background/intro summarizing prior evidence or cost estimates for computational reproducibility efforts; no specific cost study or sample size reported in the provided summary.
Spatial analysis accounting for spatial interdependence yields a total abatement effect of 15.6%.
Spatial econometric / spatial analysis reported in the study that adjusts for spatial interdependence and reports a total abatement (policy effect) of 15.6% (details and sample size not provided in abstract).
The policy reduces urban CO2 emissions by 6.0% on average.
Quasi-experimental analysis exploiting China's staggered establishment of National AI Innovation Pilot Zones (AIPZ) as a natural experiment; reported average treatment effect on urban CO2 emissions in the study (sample size not reported in abstract).
The magnitude of the negative association between AI exposure and weekly working hours grows over time, reaching its largest value in 2025.
Time-varying estimates from the event-study framework reported in the paper showing increasingly larger negative effects in later post-2022 years, with the largest estimate in 2025.
Industries with higher levels of AI exposure experienced larger declines in weekly working hours in 2023, 2024 and 2025.
Exposure-based event-study empirical analysis comparing industry-level weekly working-hour trends between 2020 and 2025 using the constructed AI exposure index; the paper reports statistically significant negative associations in 2023–2025.
AI search might satisfy information needs inside the intermediary while weakening the referral bargain that has linked search, traffic, and content production on the open web.
Interpretation/inference based on observed low outbound click rates from ChatGPT, shift in click destinations away from ad-supported sites, and measured reduction in traditional search use following ChatGPT Search expansion (Comscore clickstream analysis and rollout exploitation).
Search-referral losses from ChatGPT are largest for informational categories.
Category-specific analysis of search-referral changes using Comscore desktop clickstream and comparison across content categories; paper reports largest referral losses concentrated in informational categories.
Wider ChatGPT Search access cuts traditional search use by 9.4%.
Quasi-experimental estimate exploiting expansions in ChatGPT Search access (rollout variation) combined with URL-level Comscore U.S. desktop clickstream to measure changes in traditional search use; reported estimated reduction of 9.4%.
ChatGPT produces outbound clicks in only 5.2% of conversation sessions, far below Google's referral ratio.
Analysis of URL-level Comscore U.S. desktop clickstream comparing ChatGPT conversation sessions to Google search sessions; measured outbound-click rate per conversation/session (paper reports 5.2% for ChatGPT and states this is far below Google's referral ratio).
In developed regions, DIA–DIT synergy produces negative spatial spillovers on neighbouring areas' green productivity.
Spatial Durbin model results reported in the paper showing negative spillover coefficients for developed regions; summary provides no numeric coefficients or sample size.
The positive effect of DIA–DIT synergy on GP exhibits diminishing marginal returns once the synergy passes a certain threshold.
Threshold models reported in the paper identify a synergy threshold beyond which marginal returns to GP decline; no numeric threshold or sample size provided in the summary.
Dimensional diagnosis identified that 69% of hallucination failures were prompt-induced interpretation errors—these were invisible in aggregate scoring.
Result from the paper's sales-intelligence case study reporting failure-mode breakdown (percentage reported: 69%).
The interaction between renewable energy and CO2 emissions is negative and significant (RE × CO2 = −0.041, p < 0.001), implying high emissions undermine renewable energy benefits for green growth.
GMM interaction term RE × CO2 estimated on the 18-country panel (2000–2023); reported coefficient −0.041 with p < 0.001.
Negative renewable-energy shocks have a statistically significant but smaller long-run negative effect on green growth (−0.012, p = 0.015).
Long-run negative-shock coefficient from CS-PMG-NARDL on 18 G20 countries (2000–2023); reported coefficient −0.012 with p = 0.015.
The error-correction term indicates stable long-run adjustment: ECT = −0.145, p < 0.001 (ARDL) and ECT = −0.115, p = 0.024 (NARDL).
Estimated error-correction terms from CS-PMG-ARDL and CS-PMG-NARDL models on the 18-country panel (2000–2023); reported ECT values and p-values.
Reframing AI as a manifestation of accumulation crisis and hegemonic instability challenges accounts that treat it as an autonomous driver of capitalist renewal.
Theoretical critique and reframing based on Marxian crisis theory and related literatures; no empirical sampling or quantified tests described in the excerpt.
These patterns indicate a flight toward financial expansion characteristic of hegemonic autumn.
Interpretation of co-occurring capital expenditure and financial indicators using Marxian/hegemonic transition frameworks; no quantitative evidence or sample size provided in the excerpt.
The recent surge in artificial intelligence (AI) investment functions less as the basis of a new productive regime than as a crisis response within financialised capitalism.
Analytical argumentation and interpretation of contemporary investment patterns; no empirical sample size or formal causal identification reported in the provided text.
Contemporary capitalism is characterised by persistent overaccumulation, declining profitability, and intensified financialisation under conditions of hegemonic instability.
Theoretical synthesis drawing on Marxian crisis theory, social structures of accumulation, and theories of hegemonic transition; no specific sample size or quantitative dataset reported in the provided text.
AI has caused a decrease in the labor share of income.
Estimated impacts reported in paper indicate a decline in labor share associated with higher AI exposure; stated as a result of the analysis.
Naively persisting entire conversation histories is token-inefficient and counterproductive because irrelevant context degrades generation quality.
Argumentation in the paper supported by empirical finding that full-history persistence reduced task completion; also conceptual token-efficiency rationale.