Evidence (7560 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
Filtered →
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 Collab
Remove filter
Using a stylised inpatient capacity signalling example and minimal game-theoretic reasoning, task optimisation alone is unlikely to change system outcomes when incentives are unchanged.
Theoretical analysis using a stylised inpatient capacity signalling example and game-theoretic reasoning presented in the paper (no empirical data/sample reported in the abstract).
Deployment of AI systems carries significant costs including ongoing costs of monitoring and it is unclear whether optimism of a deus ex machina solution is well-placed.
Conceptual/argumentative claim made by the authors in the paper (no empirical study or sample size reported in the abstract).
Cross-equipment generalization is poor, with 42.7% performance on held-out datasets.
Paper reports held-out dataset evaluation showing 42.7% (presumably accuracy or task completion) for cross-equipment generalization.
Multi-asset reasoning causes a 14.9 percentage point degradation in performance.
Paper reports a 14.9 percentage point performance degradation attributed to multi-asset reasoning in comparative analyses.
There are systematic failures in tool orchestration, with 23% incorrect sequencing.
Paper reports a measured incorrect sequencing rate of 23% during evaluation of agent tool orchestration across scenarios.
Even top-performing configurations achieve only 68% task completion.
Reported aggregated performance result from the benchmark evaluation across the tested frameworks and LLMs (paper statement). The benchmark contains 75 scenarios (used as evaluation instances).
Enterprise adoption of LLMs is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level.
Framed as the motivating problem in the paper's introduction/abstract (conceptual claim; no empirical test reported here).
Specific occupations such as credit analysts, judges, and sustainability specialists reach ATE scores of 0.43-0.47 by 2030.
Reported model outputs / ATE score estimates for individual occupations within the paper's 2025-2030 regional application.
Applying the ATE framework across five major US technology regions (Seattle-Tacoma, San Francisco Bay Area, Austin, New York, and Boston) over a 2025-2030 horizon, 93.2% of the 236 analyzed occupations across six information-intensive SOC groups cross the moderate-risk threshold (ATE >= 0.35) in Tier 1 regions by 2030.
Modeling/application of the ATE score to O*NET-derived tasks for 236 occupations in six SOC groups across five named US regions with forecasts for 2025-2030; explicit numeric result reported (93.2%).
Agentic AI systems execute end-to-end workflows (multi-step reasoning, tool invocation, autonomous decision-making) and substantially expand occupational displacement risk beyond what existing task-level analyses capture.
Theoretical extension of the Acemoglu-Restrepo task exposure framework described in the paper; conceptual argument contrasting prior automation (subtask substitution) with agentic AI (end-to-end workflow automation). No empirical sample size reported for this conceptual claim.
Agent contributions are associated with more churn over time compared to human-authored code.
Longitudinal comparison between agent-generated and human-authored contributions reported in the paper (churn/survival estimates described; association between agent contributions and higher churn asserted).
Practitioners identified specific functional deficiencies in AI: inability to maintain sustained partnerships.
Theme from semi-structured interviews with 10 practitioners; cited as an example of the functional gap.
Practitioners identified specific functional deficiencies in AI: inability to adapt contextually.
Theme from semi-structured interviews with 10 practitioners; cited as an example of the functional gap.
Practitioners identified specific functional deficiencies in AI: inability to negotiate responsibilities.
Theme from semi-structured interviews with 10 practitioners; cited as an example of the functional gap.
Practitioners currently view AI models as intellectual teammates rather than social partners and expect fewer SEI attributes from them than from human teammates.
Qualitative findings from semi-structured interviews with 10 software practitioners reported in the study.
Current AI systems lack SEI capabilities that humans bring to teamwork, creating a potential gap in collaborative dynamics.
Framed as background/context in the paper; asserted rather than empirically tested in this study.
Informal workers cannot capture augmentation rents: the estimated coefficient for H^A in informal sector is negative (beta_2 = -0.044).
Subsample or interaction estimate from the augmented Mincer regression using the same merged dataset (N = 105,517); reported coefficient beta_2 = -0.044 for informal workers.
New mechanisms of surplus value distribution operate in platform-based business models and through AI-mediated processes.
Analytical/theoretical argumentation and literature synthesis in the conceptual paper (no empirical validation reported).
Task orchestration is the most under-researched dimension among the five workplace-design components.
Finding from the PRISMA-guided systematic review of 120 papers, which mapped coverage across the five dimensions and identified task orchestration as having the least research attention.
Decision authority allocation emerges as the binding constraint for Society 5.0 transitions.
Result synthesized from the systematic review and theoretical analysis mapping the five workplace-design dimensions; stated as the binding constraint in the paper's findings.
The literature shows persistent gaps in empirical validation, standardized evaluation methods, and sector-specific comparative analyses of agentic AI in financial services.
Review-level assessment noting limited empirical studies, heterogeneous evaluation metrics, and few direct cross-sector comparisons up to mid-2024.
Significant implementation barriers persist, notably workforce transformation challenges, legacy system integration difficulties, and trust deficits.
Thematic synthesis across empirical and conceptual papers in the review reporting implementation barriers and change management issues.
Ethical concerns—including bias, lack of transparency, and regulatory compliance risks—remain critical for agentic AI in financial services and necessitate layered governance and human-AI collaboration.
Collation of ethical, legal, and governance issues reported across the reviewed multidisciplinary studies and normative discussions.
Insurance is comparatively underrepresented in the literature and in reported agentic AI deployments compared with banking and investment.
Review finding (counts/themes across included studies indicating fewer studies/applications in insurance relative to banking and investment).
When predictions from the two sources conflict, the AI agent aligns more frequently with the prompt, despite its lower accuracy.
Analysis of cases where prompt-based and revealed-data-based AI predictions differed; reported frequency with which the AI's action matched the prompt versus the revealed-preference prediction.
Task complexity shapes substitution: low-complexity tasks see high substitution, while high-complexity tasks favor limited partial automation.
Calibration of the model to O*NET tasks + expert survey + GPT-4o decompositions; implementation results reported for computer vision showing substitution varies with task complexity.
AI systems exhibit predictable but diminishing returns to data, compute, and model size (scaling-law experiments), implying the cost of higher accuracy is convex: good performance may be inexpensive, but near-perfect accuracy is disproportionately costly.
Scaling-law experiments estimating performance as a function of data, compute, and model size; described experimental estimation of production function.
Under low emotional intelligence, the model predicts higher risks of over-reliance on AI, emotionally detached communication, and weaker delegation quality.
Theoretical predictions derived from the EI-moderated human–AI model presented in the paper.
The common claim that generative AI simply amplifies the Dunning–Kruger effect is too coarse to capture the available evidence.
Paper's synthesis of heterogenous empirical findings from human–AI interaction, learning research, and model evaluation used to critique the uniform-amplification interpretation; no single empirical countertest reported.
LLM use degrades metacognitive accuracy and flattens the classic competence–confidence gradient across skill groups (i.e., reduces calibration and narrows differences in self-assessed confidence by skill level).
Synthesis of studies from human–AI interaction and learning research reported in the paper that document worsened calibration and a reduction in the competence–confidence gradient when users rely on LLM outputs; the paper does not report a single combined sample size.
The authors introduce the concept of 'cascading bounded rationality' to describe how failures compound across bounded principals, agents, and evaluators.
Paper explicitly coins and defines the concept 'cascading bounded rationality' as part of its theoretical contribution.
Open-weight models cluster a full tier below the frontier models (Cohen's d larger than 1.4).
Between-group comparison reported in the paper showing a large standardized effect (Cohen's d > 1.4) separating frontier models from open-weight Meta models across the semantic closeness metric.
Azar et al. (2023) show that monopsonistic employers have stronger incentives to automate and document that US commuting zones with higher labor market concentration experienced more robot adoption.
Citation reported in the paper summarizing Azar et al. (2023); empirical analysis across US commuting zones (no sample size provided here).
Acemoglu and Restrepo (2022) attribute 50–70% of the increase in US wage inequality between 1980 and 2016 to displacement of workers from tasks by automation.
Citation reported in the paper summarizing Acemoglu and Restrepo (2022)'s attribution of the rise in wage inequality to automation-driven task displacement.
Dechezleprêtre et al. (2025) exploit Germany's Hartz reforms to estimate an elasticity of automation innovation to low-skill wages of 2–5 at the firm level.
Citation reported in the paper summarizing Dechezleprêtre et al. (2025)'s empirical estimate (elasticity 2–5); the paper states this was estimated at the firm level.
Eloundou et al. (2024) predict that half of US jobs are significantly exposed to recent advances in generative AI.
Citation reported in the paper summarizing Eloundou et al. (2024)'s prediction; no sample size provided in the excerpt.
When employers have monopsony power, they choose technologies that expand this power beyond what a social planner would consider optimal.
Model results on monopsonistic employer incentives and their technological choices; discussion supported by citations.
Profit-maximizing firms pursue innovations that erode workers' market power by making them more easily replaceable, even at the expense of production efficiency; a social planner who values worker welfare would employ technologies that preserve workers' market power.
Theoretical analysis of interactions between technological choice and market power; supported by cited empirical evidence (e.g., Azar et al. 2023) in the paper.
A welfare-maximizing planner would choose to automate fewer tasks than production efficiency would dictate when workers' welfare is heavily weighted.
Model analysis of welfare-maximizing automation level compared to production-efficient automation; analytical result in the automation application.
Observed declines in browsing time due to ChatGPT adoption are concentrated in website categories such as search and news, which are highly exposed to substitution by generative AI.
Category-level browsing time changes across website classification; concentration of declines in categories identified as highly overlap-exposed to chatbot capabilities using web-scraping and LLM site-level overlap classification.
High-income and younger households adopt generative AI substantially faster than low-income and older counterparts, and this gap is widening over time ('generative AI divide').
Descriptive heterogeneity analysis using Comscore household demographics (income and age bins) and observed adoption trajectories across 2021–2024; authors report widening gap rather than convergence.
Most of today's agents remain isolated tools or closed-ecosystem orchestrators rather than socially integrated participants in open networks.
Author claim/assessment presented as current-state analysis; no empirical breakdown or study sample provided in the text.
Prominent studies predict substantial job displacement due to automation.
Paper asserts this as background, referencing the existence of prominent studies in the literature (no specific citations or sample sizes provided in the abstract).
For organizations of n humans with AI agents, the optimal team size decreases with agent capability.
Derived implication from the stylized model's analysis of multi-human organizations interacting with AI agents.
There is no smooth sublinear regime for human effort; it transitions sharply from O(E) to O(1) with no intermediate scaling class.
Mathematical derivation from a stylized model of human-AI collaboration that assumes tasks decompose into atomic decisions, a fraction ν are novel, and specification/verification/error correction scale with task size.
So far the maintenance and migration work was done largely manually by human experts.
Background assertion in the paper's introduction/abstract; no empirical backing provided in abstract.
Performance degradation persists even when context is provided via structured semantic layers including AST-extracted function context and import graph resolution.
Experiments comparing unstructured versus structured context provision; structured semantic layers (AST context, import graph resolution) were evaluated and models still degraded with more context.
Models' performance degrades monotonically from diff-only (config_A) to diff+file content (config_B) to full context (config_C) across all 8 models.
Systematic ablation across three frozen context configurations (config_A, config_B, config_C) reported; all 8 evaluated models show monotonic performance decline as more context is provided.
Eight frontier models detect only 15–31% of human-flagged issues on the diff-only configuration (config_A).
Empirical evaluation across 8 models on SWE-PRBench (350 PRs) under the diff-only configuration; reported detection rates of 15–31% relative to human-flagged issues.
There is a growing gap between rapid experimentation with AI tools and limited organizational capability to institutionalize them in everyday workflows.
Argument supported by targeted literature synthesis and review of recent scholarly and institutional sources; no primary empirical sample reported in this paper.