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 |
Human-AI interaction factors influence people’s reliance on AI advice.
Synthesis claim from the analytical review indicating that interaction design and other human-AI interaction variables affect reliance (specific factors not enumerated in the abstract).
The AAITCF treats context as constitutive of intervention effectiveness and highlights underexplored causal pathways from AI deployment to long-term institutional change, taxpayer trust, and equitable fiscal governance.
Framework description and identification of research gaps in the paper based on the literature synthesis using the CIMO framework.
The study finds that prior reviews tended to focus narrowly (e.g., on detection metrics, behavioral dynamics, or ethical deficits) without integrating institutional boundary conditions, governance capacity, or an overarching theoretical framework.
Critical comparison and gap analysis of existing review literature as reported in the paper's introduction and synthesis sections.
Effectiveness of AI in tax compliance is contingent on data quality, governance capacity, and organizational readiness.
CIMO-structured synthesis of contextual factors across the 68 reviewed articles highlighting data, governance, and organizational readiness as moderators of AI effectiveness.
Policy implications derived from the literature include interventions spanning labor transition (reskilling/transition support), competition regulation, and digital governance.
Narrative synthesis of policy recommendations across the 78 studies and institutional reports included in the SLR.
Firm-level productivity gains from AI are contingent on complementary organizational investment.
Synthesis finding from the SLR: multiple studies report that complementary investments (e.g., organizational change, worker training, data infrastructure) are necessary for realizing productivity benefits.
Although AI working autonomously achieved a 37% reproduction rate, it could be useful for automated screening when human review is cost-prohibitive.
Interpretation in paper: authors note 37% autonomous reproduction rate as potentially useful for large-scale screening where human review is infeasible; based on empirical results of the experiment.
AI policies' carbon outcomes depend on regional economic structures, implying the need for spatially differentiated governance.
Interpretation/implication drawn from heterogeneous and spatial analyses showing region-specific effects; result is policy recommendation based on study findings (supporting analyses referenced but not detailed in abstract).
Heterogeneous effects: emissions decreased in the Pearl River Delta and increased in the Chengdu–Chongqing region and in resource-based cities (these heterogeneous findings are statistically marginal).
Subgroup/regional heterogeneity analysis comparing policy effects across regions (Pearl River Delta, Chengdu–Chongqing, resource-based cities); statistical significance described as marginal in the paper (no sample sizes or exact p-values provided in abstract).
The remaining clicks from ChatGPT are not a scaled-down Google stream: they skew toward specialized destinations and away from ad-supported sites.
Categorical analysis of destinations clicked from ChatGPT versus Google using URL-level Comscore U.S. desktop clickstream; comparison of destination types (specialized sites vs. ad-supported sites).
Although SMEs anchor employment and output across Sub‑Saharan Africa, their uptake of AI lags global benchmarks, and prevailing explanations emphasize capital, infrastructure, and institutional voids while overlooking leadership competencies.
Background/introductory claim made by the authors to motivate the study (presented as context rather than an empirical finding from this study).
The task-based adaptive collaboration model hypothesizes that trust, explainability, and task difficulty moderate the effect of human–AI collaboration on performance.
Statement of hypothesized relationships within the model developed in the paper (theoretical hypotheses rather than reported experimental estimates).
Essay quality changes little while students have AI access but improves in style and relevance one week later when students write unaided.
Open-ended essay assessments (higher-order skills) collected immediately (with AI access for treatment group) and one week later (unaided) in the randomized experiment; quality measured on dimensions including style and relevance.
Long-run asymmetric response to renewable energy shocks is statistically confirmed (Wald χ² = 5.42, p = 0.020).
Long-run Wald test for asymmetry from CS-PMG-NARDL on the 18-country panel (2000–2023); reported χ² and p-value.
Short-run asymmetric response to renewable energy shocks is statistically confirmed (Wald χ² = 4.102, p = 0.043).
Short-run Wald test for asymmetry from CS-PMG-NARDL on the 18-country panel (2000–2023); reported χ² and p-value.
The research is limited by the current state of AI technology and the available proxies; therefore the validity of the present optimistic findings must be continually re-evaluated.
Authors' stated limitations in the abstract noting rapid AI advancement and proxy measurement constraints.
These positive results are not supported in all contexts (i.e., the positive effects are not universally found across all specifications/contexts).
Abstract statement noting heterogeneity/robustness: preferred results hold but are 'not supported in all contexts.' Implies some specifications or subsets do not show the effects.
Unprecedented AI capital expenditure coexists with persistent operating losses, speculative valuations, and fragile revenue models.
Empirical characterization asserted in the paper (references implied); the provided excerpt does not state specific datasets, firms counted, dates, or sample size.
Employment effects follow the same timing (i.e., emerge in 2021) but diverge by exposure type.
Paper reports employment effects with temporal alignment to output effects (emerging in 2021) and heterogeneity by type of AI exposure.
Firm profitability shows a "J-curve" as firms move from no adoption to deep adoption.
Reported relationship between adoption intensity and firm-level profitability (authors' empirical comparison/regression of profitability across adoption categories).
Adoption is slowly accelerating among non-technology firms but very aggressive adoption in the technology sector which accounts for two-thirds of deeply integrated enterprise adoption.
Reported sectoral breakdown and temporal trend in adoption (authors' sector analysis of SEC 10-K–based adoption measure; statement that tech sector comprises two-thirds of deep adopters).
Empirical claims across the reviewed literature vary in methodological rigor and should be viewed with caution before standardized replication.
Meta-level assessment presented in the review of peer‑reviewed literature (2020–2025); no formal quality-assessment statistics provided in the excerpt.
Approach motivation (BAS Drive) moderates whether interactive partnership benefits originality.
Moderation analysis reported from the pilot (N = 62) showing interaction between BAS Drive (a measured personality/motivation scale) and the effect of interactive partnership on originality.
The Twin Transition is macro-feasible, but its adjustment costs fall unevenly on the manufacturing workforce.
Distributional outcomes and sectoral labor adjustment results from the CGE model (S4) showing heterogeneous effects across manufacturing sectors and implied labor reallocation costs.
Under S3 alone the Electricity sector expands only +0.10% by 2030 despite a 14.87% per year IT investment surge, indicating binding generation capacity that the Green AI productivity shock relaxes in S4.
CGE simulation of S3 (exogenous IT investment surge at 14.87% per year) and comparison with S4 results in the 23-sector model calibrated to 2019 I-O table.
Green AI’s export surge causes real exchange rate appreciation that displaces output in Textiles by 5.5% and in Leather and Footwear by 16.1%, while Heavy Manufacturing expands by 12.9% and IT Hardware by 10.9%.
Sectoral output changes reported from the CGE model under scenario S2 (Green AI) calibrated to Vietnam 2019 I-O table.
Brown AI (S3) is macro-neutral in GDP terms but imposes a consumption cost of 0.42% by 2030 as infrastructure investment crowds out household expenditure.
Model simulation of scenario S3 (Brown AI) in the 23-sector recursive dynamic CGE calibrated to Vietnam 2019 I-O table; S3 modelled as an exogenous IT hardware and services investment surge.
Traditional jobs based on manual work are transforming into collaborative management and exception-handling roles that demand new cognitive and ethical skills from employees.
Secondary data literature review of peer-reviewed research and industry evidence published 2022–2026 (method: secondary data review / synthesis). No specific sample size reported.
The model yields propositions on threshold effects, productivity J-curve dynamics, distributional stress, and policy sequencing.
Model-derived propositions and theoretical implications presented in the paper (analytical derivations and theory-building).
The DIAC model identifies three regimes of AI adoption and absorption: adoption without absorption, constrained complementarity, and adaptive complementarity.
Taxonomy and regime definitions derived in the paper's theoretical model (analytical/theory-building).
The same AI shock can produce divergent outcomes in small open economies.
Core theoretical claim derived from the Dynamic Institutional Absorptive Capacity (DIAC) model developed in the paper (analytical/theory-building).
Artificial intelligence is widely expected to raise productivity, yet its macroeconomic gains remain uncertain, uneven, and institutionally mediated.
Statement and literature-motivated framing in the paper's introduction; supported by analytical theory-building (DIAC model) rather than empirical data.
AI performs best in routine, data-rich situations but falls short when decisions require lived experience and contextual understanding.
Synthesis of cross-domain empirical studies and theoretical arguments showing differential AI performance by task type (routine/data-rich vs. experience-dependent/contextual).
A distinctive feature of the taxonomy is a dedicated category for self-evaluation: every improvement loop is a claim that some signal can substitute for human judgment.
Authors' taxonomy and conceptual argument emphasizing self-evaluation as a separate category across surveyed works.
The taxonomy separates bounded self-refinement -- convergent, evaluable, and already industrial practice -- from open-ended recursive self-improvement (RSI).
Conceptual taxonomy constructed by the authors based on their survey of the literature; classification of surveyed works into categories.
The literature's vocabulary ("self-refine," "self-reward," "self-play," "self-evolve") conflates fundamentally different ambitions.
Qualitative analysis of terminology across the surveyed arXiv papers (2024-2026) reported in the paper's survey and taxonomy section.
The organizing claim of the theory is that review is the control point through which a coding agent's effect on software is decided, and that AI does not fix the sign of that effect: the team sets it, through the expertise its humans bring and how it structures the review process.
Synthesis of practitioner discourse coded into a causal model derived from the LLM-assisted analysis of 3,100 sampled documents; presented as the central theoretical claim.
Practitioners sharply disagree about how coding agents change code review: whether review becomes the bottleneck, whether human review remains necessary, and whether agents erode the understanding that review once built.
Synthesis of practitioner discourse at scale via collected grey-literature (engineering blogs and Reddit threads) and a coded sample; claim summarizes observed disagreement in practitioner sources.
The direction of these observed trends (review frequency, merge speed, discussion) flips under different but equally defensible analysis choices.
Authors' sensitivity/robustness checks on the observational GitHub analysis indicating that trend direction depends on analysis choices; reported in abstract without numeric detail.
The dominant paradigm has shifted from 'substitution' (machines replacing workers) to 'augmentation' (AI augmenting human work).
Interpretive conclusion in the paper drawn from secondary literature (WEF, ILO, McKinsey, PwC) and observed policy/industry trends.
A backdating exercise on the synthetic difference-in-differences yields larger absolute estimates than the actual treatment date across most age bands.
Robustness/check: synthetic DiD backdating experiment reported in the paper produced larger absolute estimates when using earlier (backdated) treatment dates.
We propose 'contextuality' — the degree to which an AI system autonomously accesses a user's accumulated knowledge capital — as a dimension of AI-mediated inequality that complements, but is not reducible to, the Sharp et al. framework.
Conceptual proposal and definitional contribution in the paper presenting contextuality as a new analytic dimension.
The paper analyzes the technical basis of the Context Access Divide in Model Context Protocol (MCP) and retrieval-augmented generation (RAG) architectures.
Technical analysis and architectural examination reported in the paper (discussion of MCP and RAG as implementation-relevant architectures).
The CAD is formalized with a probabilistic model grounded in the fan effect literature in cognitive psychology.
Paper reports a formal probabilistic model drawing on the fan effect literature; model described as the formalization of CAD.
Sharp et al. (2025) introduce "agentic inequality" as a framework for analyzing disparities in access to AI agents across three dimensions: availability, quality, and quantity.
Statement and citation in the paper (reference to Sharp et al. 2025); descriptive synthesis of prior work.
The paper identifies four systemic tensions generated by embodied AI adoption: openness versus control; scaling versus local fit; automation ambition versus reliability constraints; and monetization versus trust.
Explicit listing of four tensions in the abstract as theoretical findings (conceptual analysis).
Data generated through physical use of embodied AI travels beyond the adopting firm (i.e., data flows cross firm boundaries).
Explicit conceptual claim in the abstract about data movement across ecosystems (theoretical observation).
Embodied AI implies a double learning loop: a closed learning loop inside the adopting firm (transforming situated use into operational feedback and workflow changes) and an external learning loop across the ecosystem of technology providers, component suppliers, software firms, platform orchestrators, and users.
Conceptual model/argument presented in the abstract describing intra-firm and inter-organizational learning loops (theoretical development).
Because AI externalities differ in nuanced ways, tax policy must be carefully designed and matched to the specific harms and policy objectives.
Author conclusion/recommendation based on the paper's analysis of heterogeneous AI externalities and tax instrument trade-offs; normative claim in text (no empirical test in excerpt).
The benefits and pitfalls of these instruments include feasibility, measurement problems, incidence, leakage, and innovation costs.
Author assessment summarized in paper identifying common advantages and disadvantages of proposed tax instruments; descriptive/theoretical evaluation rather than empirical evidence in the excerpt.