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
Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence.
Model analysis (categorization of dynamical behaviors) presented in the paper.
This feedback can give rise to distinct dynamical regimes.
Analytical results derived from the minimal dynamical model described in the paper.
We introduce a minimal model with three variables -- human cognition, data quality, and model capability.
Model development in the paper (mathematical/minimal dynamical model); presented as a constructed model rather than empirical measurement.
Humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining.
Conceptual framing and theoretical proposal in the paper; model formulation rather than empirical data.
Prior work has studied cognitive offloading in humans and model collapse in recursive training, but these effects are typically considered in isolation.
Literature review / related-work statement in paper; references to prior research (qualitative, no sample size stated).
Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning.
Background/literature observation cited in paper (qualitative claim), no empirical sample or quantified data reported in text provided.
Modeling fiscal policy as a government problem (instead of an abstract planner) implies a tax changes the firm's automation first-order condition, raises revenue only on the remaining automation base, and requires specifying rebates and administrative losses.
Explicit governmental optimization and budget-accounting setup in the model: taxes enter firms' automation first-order conditions; revenue is computed on post-tax automation activity and rebates/administration are modeled.
The central analytic object is the derivative of household consumption demand and the collective wage bill with respect to automation.
Paper's stated modeling focus: comparative-static derivatives linking automation to household consumption demand and aggregate wages; used to characterize incidence and welfare effects.
Automation reallocates income and ownership claims.
Theoretical model with heterogeneous households who hold capital/equity claims; equilibrium determines wages and returns and shows changes in income and ownership shares when automation increases.
Across studies, causal modeling reveals that cognitive alignment systematically drives attentional coordination in successful collaboration, while mismatches between effort and attention characterize unproductive regulation.
Synthesis of causal inference results from the three studies using time-series measures (JME, JVA) and episode-based analyses across the pooled dataset (182 dyads total).
There is substantial heterogeneity in the productivity effects across settings.
Meta-analytic heterogeneity assessment reported in the paper (subgroup/moderator analyses indicate variability by context). The paper states 'substantial heterogeneity across settings.'
Program outcomes are moderated by a person's prior occupational skill set, their area of work, and features of the local economy.
Heterogeneity analyses across subgroups defined by prior occupational skill composition, industry/area of work, and local labor-market conditions in the WIOA administrative data (2017-2023) show variation in outcomes.
These findings challenge the notion of a universal technological dividend from AI (i.e., AI does not automatically deliver uniform productivity gains across firms).
Overall interpretation/synthesis of heterogeneous empirical results from the panel and cluster analyses showing variation in productivity effects across firm types.
AI adoption yields asymmetric productivity gains depending on firms' resource constraints and competitive environments (i.e., heterogeneity rather than a homogeneous effect).
Heterogeneity analysis using multidimensional clustering (firm size, age, market competitiveness, digital infrastructure) applied to the panel dataset; reported differential effects across clusters.
AI adoption affects Total Factor Productivity (TFP) of firms.
Panel regression analysis using the stated panel dataset examining relationship between AI adoption and firm-level TFP.
Overall conclusion: AI offers substantial benefits to financial institutions, but ethical considerations and strategic workforce planning are essential for sustainable integration.
Synthesis/interpretation by the authors drawing on their empirical results (positive effects on ROA, efficiency, risk-adjusted returns, customer satisfaction, reduced compliance costs/breaches) and identified challenges (algorithmic bias, workforce displacement).
These divergences carry direct implications for policy interventions.
Interpretation/conclusion drawn from the divergence between RL Feasibility Index and existing measures (policy implication claimed by authors).
While Agentic AI enhances economic performance, its benefits are mediated by structural conditions and are unevenly distributed across countries (i.e., reinforcing core–periphery inequalities).
Combined findings from fixed-effects regressions, mediation analysis, and observed heterogeneity between developed and emerging economies in the 2015–2024 panel.
No single governance setting dominates across all contexts; moderate governance becomes increasingly competitive as the learner accumulates experience within the governed action space.
Empirical finding reported from experiments with the contextual-bandit learner operating under different governance constraints and learning over time; comparative performance over learning horizon described in the paper. Sample size / trial counts not provided in the excerpt.
This workload-buffering effect (governance improving performance while reducing fatigue) contradicts the usual framing of governance as pure overhead.
Interpretation and comparison of empirical manufacturing results against prior framing in literature (qualitative claim within the paper). No sample size provided.
Governance is not a binary switch but a tunable design variable: tighter constraints predictably convert autonomous AI assignments into supervised collaborations, with domain-specific costs and benefits.
Empirical finding reported from experiments using the HAAS benchmark across the two domains (software engineering and manufacturing); qualitative and/or quantitative comparisons of allocations under varying governance constraints. Paper does not state sample size in the provided text.
Workload-aware blended pricing reorders the leaderboard substantially: 7 of 10 top-ranked endpoints under the chat preset (3:1 input:output) fall out of the top 10 under the retrieval-augmented preset (20:1).
Comparison of endpoint rankings under two workload presets (chat preset 3:1 and retrieval-augmented preset 20:1); statement gives counts (7 of top 10 change).
Modeled joules per correct answer varies by a factor of 6.2 across endpoints.
Modeled energy estimate combined with task accuracy to compute joules per correct answer across 78 endpoints.
Across 78 endpoints, the same model on different endpoints differs in tail latency by an order of magnitude.
Empirical tail-latency measurements across 78 endpoints serving 12 model families.
The same model on different endpoints differs in fingerprint similarity to first party by up to 12 points.
Empirical measurement of fingerprint (output-distribution) similarity to a first-party reference across the same set of endpoints (78 endpoints, 12 model families).
Across 78 endpoints serving 12 model families, the same model on different endpoints differs in mean accuracy by up to 12.5 points on math and code.
Empirical measurement across 78 endpoints and 12 model families comparing mean accuracy on math and code tasks.
The rise of digital agents will transform the foundations of production, labour markets, institutional arrangements and the international distribution of economic power.
Synthesis and theoretical projection across sections of the paper; presented as a broad conclusion without reported empirical quantification in the provided text.
There is a fundamental asymmetry between economic and social reproduction: digital agents can compensate for productive functions of the population but are unable to substitute the population's functions of social reproduction.
Theoretical argument and conceptual distinction in the paper; no empirical study measuring substitution in social reproduction provided.
These patterns suggest that AI adoption is associated with expected efficiency gains that shape both firms' pricing behaviour and their macroeconomic expectations.
Interpretation based on observed increases in productivity/profitability and different pricing/inflation expectations among adopters vs non-adopters in survey and DID analyses.
AI adoption leads both to job displacement and job creation, including the emergence of new occupational categories.
Abstract states the review examines empirical evidence on both job displacement and creation and the emergence of new occupations; no numeric counts or sample sizes provided in abstract.
The study identifies short-term transitional risks and long-term productivity gains associated with AI integration in the workforce.
Abstract states the paper evaluates both short-term risks and long-term productivity gains from AI integration based on the reviewed literature; no empirical quantification given in abstract.
AI-driven automation and augmentation are reshaping employment landscapes, with emphasis on sector-level disruption, skill transformation, and socioeconomic consequences.
Abstract states this as a conclusion of the review drawing on interdisciplinary empirical literature; no specific studies or sample sizes cited in abstract.
The accelerating deployment of artificial intelligence across industries has fundamentally altered the structure of global labour markets.
Statement in abstract summarizing a systematic review of interdisciplinary literature (economics, computer science, organizational behaviour, public policy); no specific sample size reported in abstract.
The magnitude of AI’s effect on potential GDP varied across industries and depended on the level of digital maturity, human resources, and institutional conditions.
Decompositional analysis across aggregated industry data and scenario-based modeling drawing on sectoral sources and reviews.
Failures are structured by task family and execution surface, with HR, management, and multi-system business workflows as persistent bottlenecks and local workspace repair comparatively easier but unsaturated.
Error-mode analysis across the 105 tasks and evaluated models reported in experiments; authors identify task-family-level patterns (HR, management, multi-system workflows) and relative ease of local workspace repair.
Whether LLM-based assistants improve or degrade code quality remains unresolved: existing studies report contradictory outcomes contingent on context and evaluation criteria.
Review finds mixed/contradictory findings across included studies regarding code quality effects.
The system tends to be factually correct when it answers but often omits information (i.e., 'the system is right when it answers — it just leaves things out').
Interpretation combining reported factual accuracy (85.5%) with low completeness (0.40) from benchmark results.
The study establishes statistically significant relationships between organizational AI adoption and compensation dynamics.
Econometric estimates (difference-in-differences and propensity score matched comparisons) using the combined datasets listed in the paper and controlling for industry, firm size, geography, occupation characteristics, and macroeconomic variables.
The study establishes statistically significant relationships between organizational AI adoption and changes in occupational structures.
Same econometric approach (difference-in-differences and propensity score matching) applied to combined datasets (Anthropic Economic Index, Census Business Trends and Outlook Survey, Federal Reserve regional surveys, labor market analytics), with controls for industry, firm size, location, occupation-level characteristics, and macroeconomic environment.
The study establishes statistically significant relationships between organizational AI adoption and changes in employment patterns in the United States during 2022–2025.
Econometric analysis using multiple large-scale data sources (Anthropic Economic Index, U.S. Census Bureau Business Trends and Outlook Survey, Federal Reserve regional surveys, labor market analytics) and methods described as difference-in-differences estimation and propensity score matching controlling for industry (NAICS 2-digit), firm size, geography, occupation characteristics, and macro conditions.
The paper extends paradox theory to conceptualise the Creativity Paradox in the context of GenAI.
Theoretical extension and conceptual development within the paper (no empirical tests reported).
Within that n=11 subset, 9 of 11 agents shift by at least 2 ranks between composite and benchmark-only rankings.
Comparison of rank positions between composite and benchmark-only rankings on the 11-agent subset; reported count of agents that moved at least 2 ranks.
The four factors capture largely complementary information (n=50; ρ_max = 0.61 for Adoption-Ecosystem, all others |ρ| ≤ 0.37).
Correlation analysis among the four factor scores computed on the 50-agent sample; reported maximum inter-factor Pearson/Spearman correlation coefficients.
Provisioned Throughput delivers the lowest latency at low concurrency but saturates its reserved capacity above approximately 20 concurrent users.
Empirical measurements from the instrumented system across concurrency up to 50 users and tier comparisons; the paper reports the observed saturation point near ~20 concurrent users.
Delegating tasks to genAI can be individually beneficial in the short term even as widespread adoption degrades future model performance (creating a social dilemma).
Result of the paper's behavioral model showing an individual-level incentive to use genAI versus a collective cost from adoption (theoretical/model-based; no empirical sample reported in abstract).
Token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens.
Observed run-to-run variability in total token counts for identical tasks across the collected agentic trajectories from eight frontier LLMs on SWE-bench Verified.
ASC (adaptive stopping criterion) halts harmful refinement but incurs a 3.8 pp confidence-elicitation cost.
Reported experiment with ASC showing that it prevents harmful iterative refinement yet causes a measured cost described as 3.8 percentage points due to confidence elicitation.
Only o3-mini (+3.4 pp, EIR = 0%), Claude Opus 4.6 (+0.6 pp, EIR ~ 0.2%), and o4-mini (+/-0 pp) remain non-degrading under self-correction; GPT-5 degrades by -1.8 pp.
Reported measured changes in accuracy (percentage-point changes) and measured EIR values for the named models after applying iterative self-correction across the experiment suite.
Across 7 models and 3 datasets (GSM8K, MATH, StrategyQA), we find a sharp near-zero EIR threshold (<= 0.5%) separating beneficial from harmful self-correction.
Empirical experiments reported across 7 LLMs and 3 benchmark datasets (GSM8K, MATH, StrategyQA) comparing outcomes of iterative self-correction as a function of measured EIR.
These efficiency gains are offset by a growing 'Efficiency-Legitimacy Paradox' (i.e., improvements in efficiency come with worsening legitimacy concerns).
Conceptual synthesis from the systematic review (2018-2026) identifying a recurring trade-off across reviewed studies; specific empirical quantification not provided in abstract.