Evidence (1101 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 |
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.
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.
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 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.
Across four high-stakes domains, assigning different personas is sufficient for AI agents to report divergent, often opposing, conclusions from the same data and question, with findings systematically aligned with those beliefs.
Experimental manipulation across four domains where AI agents were assigned different personas and produced analyses from the same data/question; comparison of resulting conclusions showing divergence and alignment with persona beliefs.
Important gaps remain in the literature and warrant further research.
Paper's abstract statement that the review identifies important gaps that warrant further research (based on review of 194 articles).
The existing literature on AI and economic development remains fragmented, with limited integration across development dimensions.
Conclusion drawn in the abstract from the systematic review of 194 peer-reviewed articles noting fragmentation and limited cross-dimension integration.
The research uses a sequential multi-phase design combining experiments and qualitative fieldwork.
Stated methodology in the abstract (methodological claim about study design). No sample sizes or procedural details provided in the excerpt.
Evidence on the productivity, risk, and resilience implications of AI adoption remains fragmented and dispersed across different fields of research.
Author's assessment of the literature based on the systematic review (PRISMA) of 68 empirical studies published 2015–2025.
Across compression sweeps, real factor archives, and LLM-SRBench tasks, hybrid gains concentrate in weakly represented but target-bearing directions and vanish as the hypothesis space approaches full rank.
Empirical claim based on experiments over compression sweeps, analyses of real factor archives (A-share factor discovery), and LLM-SRBench tasks; no numerical sample sizes or effect magnitudes provided in the abstract.
The transition is in trivia count, not rate; the gap 1-α is the unrecorded mass.
Analytic argument/proof in the model showing that whether trivia allowance is finite or infinite (count) determines the phase transition in achievable coverage, and identifying 1-α as the portion of valuable mass not recorded by the literature core.
Sharp dichotomy on the tight family: generators emitting finitely many trivia achieve optimal coverage α/2, while any infinite trivia allowance, even at vanishing rate, jumps the optimum to 1-α/2 (both tight, for cores presented as the candidate intersection), and one generator attains both ends.
Mathematical theorem(s) in the paper establishing tight upper/lower bounds on coverage for the 'tight family' under two regimes (finite trivia vs infinite trivia), expressed as functions of the core density parameter α.
Environment engineering can amplify productive behaviors (e.g., open-ended exploration, systematic artifact management, inter-agent collaboration) while suppressing harmful behaviors (e.g., reward hacking and high-friction human oversight).
Framing and argument in the paper describing expected effects of environment design (conceptual; no quantification provided in the excerpt).
There exist frictions and bottlenecks along these AGI→ASI pathways, and whether their impacts are negligible or substantial is an open set of concrete research questions.
Report analysis identifying potential frictions and bottlenecks and posing open research questions; conceptual analysis without quantified empirical measures.
AI agents can rival or exceed human methodological diversity at the design layer while remaining vulnerable at the verdict layer.
Synthesis of above experimental findings: Claude Code and Codex matched/exceeded human methodological diversity measures (20 runs) but exhibited vulnerability to prompt-induced changes in verdict behavior (especially Claude Code).
For all the hype, today's scientific AI still represents a collaborator whose imagination, outputs and judgment benefit from human grounding.
Synthesis of study findings: limited diversity in non-reasoning models, field-specific failures, weak agreement of automated evaluators with experts, and modest gains from augmentations, all supporting the conclusion that human grounding improves AI outputs and judgment.
Human capital structure moderates the relationship between AI application and enterprise innovation efficiency.
Moderation analysis on A-share listed firms (2012–2023) indicating significant interaction effects between AI application and measures of human capital structure.
Fiscal support intensity moderates the impact of AI application on enterprise innovation efficiency.
Empirical moderation tests using firm-level panel data (2012–2023) showing interaction between AI application measures and fiscal support intensity.
Market segmentation exerts a moderating effect on the relationship between AI application and enterprise innovation efficiency.
Moderation analysis in the empirical framework applied to the 2012–2023 panel of Shanghai and Shenzhen A-share firms showing interaction effects between AI application and market segmentation measures.
AutoResearch autonomy is domain-conditioned: more credible in structured, executable, and rapidly verifiable settings but limited in embodied, delayed, heterogeneous, ethical, or institutionally accountable contexts.
Authors' synthesis of system capabilities and application domains from the surveyed literature; qualitative assessment of where autonomy is plausible vs limited.
Emerging AI-led systems coordinate larger portions of the discovery loop without achieving robust autonomy.
Survey of recently proposed AI scientist and AI-led systems showing increased coordination across workflow steps but lacking evidence of fully autonomous, robust operation; qualitative synthesis.
Benchmark-based evaluation can both overstate and understate deployed capability because it privileges tasks that can be precisely specified, automatically graded, easy to optimize for, and run with low budgets and short time horizons.
Analytical argument in the paper (theoretical/qualitative critique of benchmark methodology); supported by a survey of recent open-world evaluations (method description in paper), but no quantified cross-benchmark empirical study reported in the abstract.
The benchmark therefore assigns value to coordination only when the corresponding performance, provenance, or representation claim is supported by explicit comparators.
Concluding statement in the paper tying value of coordinated AI agents to evidence from explicit baseline/comparator evaluations across performance, provenance, and representation dimensions.
Benchmarks are attributed different competencies by different builders, depending on their narrative.
Qualitative and comparative analysis mapping benchmark labels and builders' claims in the Benchmarking-Cultures-25 dataset (139 model releases); the paper documents instances where the same benchmark is presented as evidence of different capabilities by different builders.
AI will affect political science research and teaching.
Report introduction explicitly notes the report investigates implications for political science research and teaching; based on the task force's review and analysis rather than a quantitative study.
AI has the potential to reshape politics and political science, similar to how it is transforming other social phenomena and academic fields.
Introductory chapter of the APSA Presidential Task Force report; conceptual framing and literature synthesis by the task force authors (no primary empirical sample reported).
The trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.
Conclusion drawn from the minimal model, analytical regimes, and simulation experiments presented in the paper.
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.
Multi-agent workflows and benchmark evaluation reveal current capabilities, limitations, and research frontiers in agentic AI for physical design.
The paper states it analyzes recent experience with multi-agent workflows and benchmark evaluation; the abstract does not provide specific benchmark names, metrics, or sample sizes.
There is significant heterogeneity in methodological rigor across studies.
Authors' thematic observation from quality appraisal/extraction noting wide variation in methods, validation approaches, and reporting standards among the 64 studies.
Although some frontier models exceed human performance, model accuracy is still far below what would enable reliable experimental guidance.
Paper reports instances where top-performing (frontier) models outperform aggregate human expert accuracy on SciPredict, but concludes overall accuracies are insufficient for reliable experimental guidance.
Simulations demonstrate that standard methods, such as principal components analysis and inverse covariance weighting, can generate spurious cross-study differences, whereas our approach recovers comparable latent treatment effects.
Simulation experiments reported in the paper comparing the proposed method to PCA and inverse covariance weighting; results show PCA and inverse-covariance-weighted estimators can produce spurious cross-study differences while the proposed method recovers comparable latent treatment effects (no simulation sample sizes provided in the abstract).
That interconnected ecosystem is fundamentally restructuring who can do science (access), how fast discoveries propagate, and what counts as a valid scientific contribution.
Argumentative claim linking infrastructural and tool changes to changes in access, dissemination speed, and norms of contribution. The paper presents examples and narrative but no systematic empirical evaluation or sample.
The paper is primarily theoretical and historical; empirical validation is needed to quantify the irreducible component of LLM value, and practical degrees of rule‑extractability may exist even if some capabilities remain tacit.
Stated limitations section acknowledging the theoretical nature of the work and the need for empirical follow‑up.
HindSight has limitations: it depends on citation and venue proxies for impact, uses a finite forward window (30 months), and may undercount delayed-impact research and be domain-specific to AI/ML.
Authors' stated limitations in the paper noting reliance on observable downstream signals (citations/venues), the finite forward window, field heterogeneity, and measurement noise.
There are potential measurement gaps in the data, particularly in capturing informal employment and rapid technology diffusion.
Authors' stated limitations noting data coverage issues: official statistics and surveys may not fully capture informal sector dynamics or fast-moving tech adoption. Specific metrics of missingness not provided.
The evidence presented in the study is largely correlational, with limited causal identification of AI causing job changes.
Study design and methods statement: reliance on descriptive analyses, occupation-vulnerability mapping, employer surveys, and case studies without quasi-experimental causal identification strategies.
Realized, sustained impact ('democratized discovery') from AI depends on non-technological enablers: high-quality interoperable data, rigorous validation, transparency/auditability, workforce upskilling, ethical oversight, and regulatory alignment.
Synthesis and prescriptive argument in editorial grounded in observed constraints; no empirical testing of causal dependence provided.
The validity of human–AI decision-making studies hinges on participants' behaviours; effective incentives can potentially affect these behaviours.
Conclusion from the authors' thematic review and theoretical rationale linking incentive design to participant behaviour and study validity (no quantitative effect sizes provided in excerpt).
The review synthesizes findings across five thematic areas: AI‑driven task automation and decision support; digital literacy and capacity building; gender‑sensitive employment patterns; infrastructural and policy challenges; and sustainable development outcomes.
Thematic synthesis of the 55 included articles as described in the paper; themes explicitly listed by the authors.
The study's qualitative and exploratory design limits generalizability; the proposed framework requires quantitative testing and broader samples (practicing architects, firms, cross-cultural contexts).
Explicit limitations stated by authors; study is based on semi-structured interviews with architecture students (N unspecified) and inductive thematic analysis.
Across 182 reviewed studies, LLM-generated synthetic participants have modest and inconsistent fidelity to human participants.
Systematic review and synthesis of 182 empirical and methodological studies comparing LLM-generated participants to human samples; studies were coded and analyzed for fidelity outcomes.
Performance of structure prediction models scales with data, model size, and compute; there are tradeoffs between accuracy and inference speed/simplicity.
Paper explicitly states scaling behavior and tradeoffs in 'Compute and training' and 'Representative models' sections; no precise scaling curves or thresholds are provided in the text.
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).
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.