Evidence (14922 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
9047 claims
Filter claims →
Productivity
8066 claims
Filter claims →
Governance
7278 claims
Filter claims →
Human-AI Collaboration
6912 claims
Filter claims →
Org Design
4439 claims
Filter claims →
Innovation
4359 claims
Filter claims →
Labor Markets
3652 claims
Filter claims →
Skills & Training
3018 claims
Filter claims →
Inequality
2160 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 | 795 | 210 | 105 | 955 | 2131 |
| Governance & Regulation | 886 | 414 | 197 | 126 | 1654 |
| Organizational Efficiency | 826 | 204 | 129 | 87 | 1257 |
| Technology Adoption Rate | 681 | 259 | 128 | 110 | 1189 |
| Research Productivity | 464 | 138 | 65 | 349 | 1028 |
| Output Quality | 503 | 196 | 61 | 53 | 813 |
| Decision Quality | 351 | 180 | 84 | 51 | 673 |
| AI Safety & Ethics | 238 | 288 | 71 | 34 | 637 |
| Firm Productivity | 455 | 58 | 92 | 20 | 631 |
| Market Structure | 186 | 172 | 123 | 25 | 511 |
| Task Allocation | 222 | 70 | 76 | 34 | 407 |
| Innovation Output | 238 | 28 | 48 | 18 | 334 |
| Skill Acquisition | 177 | 62 | 62 | 17 | 318 |
| Employment Level | 107 | 57 | 108 | 13 | 287 |
| Fiscal & Macroeconomic | 135 | 72 | 44 | 26 | 284 |
| Firm Revenue | 172 | 50 | 28 | 5 | 256 |
| Consumer Welfare | 121 | 68 | 45 | 12 | 246 |
| Task Completion Time | 183 | 33 | 10 | 13 | 240 |
| Inequality Measures | 45 | 126 | 50 | 6 | 227 |
| Worker Satisfaction | 95 | 74 | 23 | 12 | 204 |
| Error Rate | 77 | 98 | 11 | 4 | 190 |
| Regulatory Compliance | 84 | 73 | 17 | 7 | 181 |
| Automation Exposure | 61 | 61 | 27 | 14 | 166 |
| Training Effectiveness | 98 | 21 | 14 | 19 | 154 |
| Wages & Compensation | 78 | 37 | 25 | 6 | 146 |
| Developer Productivity | 105 | 18 | 14 | 6 | 144 |
| Team Performance | 87 | 17 | 28 | 10 | 143 |
| Job Displacement | 12 | 83 | 23 | 1 | 119 |
| Hiring & Recruitment | 53 | 8 | 8 | 3 | 72 |
| Social Protection | 39 | 17 | 8 | 2 | 66 |
| Creative Output | 32 | 20 | 8 | 3 | 64 |
| Skill Obsolescence | 5 | 50 | 6 | 1 | 62 |
| Labor Share of Income | 17 | 20 | 17 | — | 54 |
| Worker Turnover | 15 | 15 | — | 3 | 33 |
| Industry | — | — | — | 1 | 1 |
RARRL improves task success rates compared with fixed or heuristic reasoning strategies in embodied robotic tasks (evaluated using ALFRED-derived latency profiles).
Empirical experiments reported in the paper compare RARRL to baselines (fixed strategies and heuristic triggers) using an embodied task suite based on ALFRED and empirical LLM latency profiles; results claimed to show higher task success across extensive experiments.
Policy instruments that can support shorter workweeks include tax incentives for firms that maintain pay while reducing hours, regulatory transition frameworks, and conditionality on AI subsidies or public procurement tied to job-preservation or reduced hours.
Policy-analytic argument drawing on standard policy toolkits and selected prior examples; no new policy pilot results presented.
Shorter workweeks help sustain consumer purchasing power by reducing aggregate labor supply and thereby distributing automation gains more equitably.
Theoretical labour-supply reasoning plus historical case studies of work-time reductions; argumentual and normative rather than demonstrated with new macroeconomic empirical tests in AI-rich settings.
A gradual, policy-driven reduction in the standard workweek can absorb labor displaced by automation, help maintain employment levels, and preserve wages per hour.
Synthesis of prior empirical findings on work-hour reductions and historical precedents (e.g., six-day to five-day transition); no new randomized or large-scale contemporary trials presented.
Firms use layoffs strategically to signal efficiency and boost short-term stock prices, even when automation is not fully substitutive.
Organizational- and finance-literature synthesis on signaling and market reactions to cost-cutting; historical/case examples referenced rather than new econometric estimates.
Employers are increasingly demanding digital literacy, basic data competencies, and stronger communication and interpersonal skills.
Employer survey analysis tracking changes in required skills; descriptive summary of survey frequencies and employer-reported skill priorities. Survey sample size and representativeness not specified in summary.
Some occupations experience efficiency and productivity gains where AI complements tasks, implying complementarity effects for those jobs.
Qualitative case studies of firms and employer survey reports documenting productivity/efficiency improvements in certain roles following AI adoption; descriptive analysis of sectoral/occupational outcomes. Quantitative magnitude not specified.
Policymakers should prioritize retraining programs, strengthened social protection, and redistributive policies to mitigate automation-induced unemployment and inequality.
Policy recommendation based on the author's synthesis of risks and expert judgment; not based on an empirical intervention study in the paper.
There has been progress in software import substitution, contributing to partial technological sovereignty in Russia.
Use of statistics on software import substitution (authors reference national statistics but do not report detailed numbers or methodology).
Digitalization enables management optimization (improved management processes and decision-making) in Russian enterprises and public administration.
Qualitative analysis of policy documents and expert assessment by the author; no empirical evaluation or quantified effect sizes provided.
Digitalization has produced measurable labor productivity growth in segments of the Russian economy.
Author's interpretation drawing on national statistics and strategic documents; statistical details (period, sectors, sample sizes) not specified in the paper.
Policy implication: prioritize large-scale, targeted reskilling and lifelong learning programs to enable workforce adaptability and capture AI complementarity gains.
Policy recommendations derived from the paper's findings (association between AI adoption and skill shifts, heterogeneous sectoral impacts) and the literature synthesis that links reskilling interventions to better labor outcomes; recommendation is prescriptive rather than empirically tested within the study.
The paper provides empirical support for the complementarity hypothesis: AI tends to reconfigure jobs and create hybrid roles rather than eliminate employment wholesale.
Convergence of simulated sectoral employment patterns (some sectors showing net gains and hybrid-role growth), the strong correlation between AI adoption and skill shifts (r = 0.71), and corroborating studies from the literature synthesis emphasizing augmentation and hybridization mechanisms.
Institutional reskilling programs and governance frameworks markedly moderate labor-market outcomes: better frameworks correlate with more complementarities and lower net job loss.
Integration of literature-derived mechanisms with simulated empirical patterns; paper reports correlations/moderation-style comparisons across simulated sector-year cases incorporating policy/institutional variables (described in methods), supported by studies in the systematic review linking policy interventions to labor outcomes.
Healthcare and IT Services experienced net employment gains consistent with AI complementarity (augmented tasks and creation of new hybrid roles).
Simulated sectoral employment trends and net-change metrics for Healthcare and IT Services (2020–2024) presented in the paper, supported by literature synthesis examples showing human–AI complementarities in these sectors.
The largest rises in hybrid jobs occurred in IT Services and Healthcare.
Sectoral decomposition of hybrid job share trends in the simulated dataset across the seven industries (2020–2024) and supporting qualitative/quantitative findings from the literature synthesis focused on IT Services and Healthcare.
Hybrid human–AI jobs increased substantially across all seven analyzed sectors between 2020 and 2024.
Descriptive trend analysis of the simulated dataset's hybrid job share metric (fraction of roles reclassified as human–AI hybrid) for the seven industries over 2020–2024, combined with corroborating examples from the literature synthesis (selected ACM/IEEE/Springer studies 2020–2024).
A matching/ranking algorithm that scores candidate-job pairs by skill fit and predicted remuneration (and proximity) improves the alignment of workers to short-term gigs.
System incorporates a ranking algorithm combining inferred-skill fit, predicted wages, and proximity constraints; pilot comparison reported improved matches, but quantitative algorithmic performance metrics are not provided in the summary.
ML models can continuously derive available gigs and demand signals from marketplace activity, producing up-to-date opportunity lists and predicted wages.
Implemented ML models ingest real-time market activity/platform signals in the pilot to generate opportunity lists and wage predictions; no reported out-of-sample accuracy or prediction error metrics in the summary.
Skills can be inferred from multiple nontraditional inputs—self-reported information, short-term work histories, and community recommendations—creating richer profiles beyond formal work experience.
System design uses NLP to normalize and extract skills from profiles, short-term work records, and community recommendations; claim is supported by the implemented data integration approach rather than by quantified external validation in the summary.
The pilot implementation produced higher reported wages for youth matched through the system relative to baseline informal methods.
Pilot comparison reported higher reported wages for matched youth; summary lacks sample size, measurement protocol, and statistical inference.
The pilot implementation led to higher correct matches compared to existing informal search methods.
Pilot deployment compared matching accuracy versus baseline informal job-search approaches; the paper summary reports a 'marked increase' but provides no numerical details, sample size, or significance levels.
AI-driven NLP and ML can substantially reduce search frictions in Nairobi’s informal and gig economies by dynamically deriving individual skills and real-time market opportunities, then algorithmically matching youth to short-term work.
Pilot implementation of an end-to-end system combining NLP, ML and a matching algorithm deployed in Nairobi and compared qualitatively/aggregately against baseline informal search methods; paper summary does not report sample size, statistical tests, or numerical effect sizes.
Firms should pair strong-performing ensemble/deep models with explainability tools (e.g., feature-importance, SHAP) and fairness audits, and prefer pilot human-in-the-loop implementations to validate economic impacts and reduce operational risks.
Authors' practical recommendations based on empirical model performance, interpretability analyses, and noted limitations; presented as guidance rather than empirically validated interventions.
Variable-contribution analyses (feature importance / model explanation techniques) clarified which inputs drive predictions, making results actionable for HR decision-making.
The paper reports use of feature-importance and model-explanation methods to quantify variable contributions and interpretable outputs intended for HR practitioners.
Employee engagement/participation levels, learning agility (pace of acquiring new skills), tenure in current role, and perceived workload/manageability are consistently among the most important predictors of job performance in the datasets examined.
Feature-importance and model-explanation analyses (e.g., feature importance, SHAP-style approaches) applied across multiple publicly available workforce datasets produced consistently high importance scores for these variables.
The models' superior performance hinges on their ability to capture complex, non-linear patterns in features (e.g., engagement, learning agility, tenure, workload perception).
Inference from comparative model performance: non-linear models (ensembles, DNNs) outperform linear baselines; feature engineering captured engagement dynamics and learning trends; variable-contribution analyses highlighted these feature types as influential.
These predictive gains persist when models are applied to different company datasets, indicating better generalization of AI methods.
Cross-company tests described in the paper: models trained/tuned on one dataset and evaluated on others (holdout across organizations) with reported performance metrics demonstrating persistent improvements for AI methods.
Responsible implementation requires legal/liability clarity, continuous monitoring for performance drift and distributional shifts, usable explanations, baseline AI literacy for clinicians, and co-design with frontline radiology teams.
Synthesis of governance literature, implementation best-practice reports, and recommendations from usability and deployment studies.
Triage and automation can shorten time-to-diagnosis, increase throughput, and reduce time spent on repetitive tasks.
Observational deployment reports and simulation studies that measured time-to-report or throughput improvements in pilot settings (evidence heterogeneous and context-dependent).
Integration points for AI across the imaging pathway include acquisition (image quality/protocol selection), triage (prioritization), interpretation/reporting (detection, quantification, report pre-population), and post-interpretation (teaching, QA, model improvement loops).
Descriptive synthesis of reported implementations and proposed use cases in the literature and deployment reports across multiple institutions.
Human-AI collaboration can produce synergistic gains (diagnostic complementarity) when errors are uncorrelated and tasks are allocated to leverage comparative strengths.
Theoretical/analytical models of error complementarity and empirical reader studies showing instances where combined readings outperform either agent alone (evidence drawn from multiple small-to-moderate reader studies and simulations).
AI in radiology has clear potential to improve diagnostic performance and workflow efficiency.
Narrative synthesis of laboratory evaluation studies, reader/comparison studies, and a limited number of observational deployment reports showing improved algorithm accuracy and some improvements in measured throughput or time-to-review in pilots (study sizes and settings heterogeneous; few large-scale RCTs).
Cognitive Shadow supports real-time model updates based on immediate user feedback, enabling iterative improvement and continuous alignment with human decision patterns.
Described human-in-the-loop interaction loop where CS captures human decisions, provides recommendations, receives immediate feedback, and updates models dynamically in the simulation environment (implementation detail).
HACL/CS reduces omission rates (missed detections) in the simulated scenarios.
Omission/error rates were tracked and compared between conditions in the simulated testbed; summary claims reduction in omissions with HACL assistance but does not report numeric effect sizes or significance.
HACL/CS reduces time-to-decision in the simulated maritime surveillance tasks.
Measured time-to-classify in simulation under human-alone vs HACL-assisted conditions; summary indicates reductions in time-to-decision but lacks detailed statistics in the provided description.
In the simulated Canadian Arctic maritime surveillance domain, HACL/CS shows promise for improving classification accuracy.
Performance comparison between human-alone and HACL-assisted conditions in the maritime surveillance simulation measuring classification accuracy; summary reports improvement but does not provide sample size or significance levels.
Adjustable autonomy via self-confidence thresholds enables the system to act autonomously on high-certainty predictions and defer to humans on low-certainty cases.
System design feature of Cognitive Shadow implemented in simulation: autonomy decision rule based on meta-model confidence thresholds; behavior demonstrated in human-in-the-loop scenarios.
The Cognitive Shadow toolkit quantifies AI reliability with an empirical (0–1) confidence metric produced by a recursive meta-model.
Design and implementation detail: primary supervised models are paired with a recursive meta-model that predicts the primary model's reliability per situation and outputs a 0–1 empirical confidence score; applied in the simulated testbed.
Implementing an adaptive command-and-control process augmented by AI metacognition (the Cognitive Shadow toolkit) aligns AI judgments with expert human decision patterns.
Cognitive Shadow (CS) implemented as supervised ML models trained to mimic expert human decisions in the simulated maritime scenarios; alignment assessed by comparing model outputs to human expert decisions during human-in-the-loop interaction (implementation validated in simulation).
Human-AI co-learning (HACL) improves human-autonomy teaming (HAT) effectiveness.
Evaluated in a simulated Canadian Arctic maritime surveillance testbed using human-in-the-loop experiments comparing human-alone vs HACL-assisted conditions; exact participant sample size and statistical details not provided in the summary.
Self-directed autonomous agents (those that autonomously generated prompts and selected tools) bypassed human prompting failures and outperformed most human teams on the challenge set.
Comparative analysis of the four autonomous agents' trajectories, tool use, and success rates versus the 41 human participants/teams on the same fresh challenges; observed correlation between autonomous self-direction and higher success relative to most teams.
Clinical-interface validation with real physicians on mobile devices confirmed the practical viability and usability of the HADT system and interface.
Paper reports an interface test with real doctors using the mobile interface as part of validation (number of physicians, tasks, and quantitative usability metrics not specified in the summary).
On public datasets HADT achieves superior accuracy/human-effort trade-offs compared to baselines (fully human, fully automated, and simpler assignment strategies).
Comparative evaluations reported in the paper on public medical-consultation datasets (baselines listed broadly; exact baseline implementations, dataset names, and quantitative comparisons not included in the provided summary).
The execution machine uses masked hierarchical reinforcement learning with bottom-up training to ask informative symptom questions from a large symptom space.
Methodological description and reported training procedure in the paper (bottom-up training applied to the execution module; claimed to improve question selection; experimental validation referenced but details not provided in the summary).
A two-layer hierarchical reinforcement learning system—an assignment 'master' and an execution 'machine' (plus human doctors)—effectively balances accuracy and human cost.
Architectural design and experimental evaluation in the paper demonstrating trade-offs between diagnostic accuracy and human involvement using the hierarchical RL setup (experiments run on public datasets; exact sample sizes not given).
The Human-AI Diagnostic Team (HADT) framework can deliver near-expert-level online symptom inquiry and diagnosis while using very little human labor.
Performance reported on public datasets and clinical-interface tests with real physicians; described comparisons to expert-level performance in the paper's main finding (specific comparative statistics to experts not provided in the summary).
HADT reached up to 89.4% diagnostic accuracy while requiring only 10.9% human effort.
Empirical results reported in the paper from experiments on public online medical-consultation datasets and clinical-interface validation with real doctors (dataset names and sample sizes not specified in the provided summary).
HR and organizational implication: Firms should consider emotional dynamics when designing hybrid teams; training, monitoring, and pairing strategies (human–human, human–AI) matter for short-term task performance.
Interpretation by authors based on experimental findings that emotion and perceived service empathy alter collaboration proficiency across partner types in temporary virtual tasks (n = 861).
Design implication: Investing in AI features that convey empathy or supportive social cues could increase collaboration proficiency when emotion matters.
Authors' inference grounded in mediation (service empathy → collaboration proficiency) and the partner-type moderation of that effect in the experiment (n = 861), suggesting empathy-signaling features could alter outcomes.