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
CEO–TMT faultlines negatively affect green innovation through reduced eco-attention.
Empirical mediation analysis on the panel dataset (35,347 firm-year observations, 2010–2023) testing CEO–TMT faultlines -> eco-attention -> green innovation.
Municipal 311 call centers and complaint intake systems face a structural mismatch between incoming volume and classification capacity that produces a bottleneck and differential service quality that follows income and racial lines.
Stated in the paper's introduction; cites prior work (Liu 2024 SLA) as support for the differential service-quality / demographic claim. No sample size or quantitative result reported in the excerpt.
There is an absence of agreed-upon benchmarks for evaluating AI systems.
Introductory chapter notes lack of standardized evaluation benchmarks as a cross-cutting concern; presented as an analytical observation by the task force.
AI systems exhibit bias.
Introductory chapter points to bias in AI systems as a recurring theme; supported by the broader literature cited in the report (no numerical sample reported in the introduction).
AI model outputs are often opaque and non-replicable.
Introductory chapter identifies opacity and non-replicability of AI outputs as a cross-cutting theme; claim is based on literature synthesis and conceptual critique in the report.
A small number of AI corporations have unprecedented power.
Introductory chapter highlights the theme of concentrated corporate power in AI; asserted as an observational claim in the report's framing rather than derived from a presented empirical sample in the introduction.
GPT-4.1 exhibits hidden workflow shortcuts despite achieving perfect TSR and HF1.
Model-level observation from the ASR analysis within the experiment (paper reports GPT-4.1 had perfect TSR and HF1 but failed trajectory-level fidelity).
Applied to the Hierarchical Multi-Agent System for Payments (HMASP) across 18 LLMs and 90,000 task instances, ASR reveals that 10 of 18 models systematically skip a confirmation checkpoint during payment checkout, a deviation invisible to both TSR and HF1, while 8 models enforce the checkpoint perfectly.
Empirical evaluation reported in the paper: HMASP tested across 18 LLMs and 90,000 task instances; analysis via ASR showing checkpoint-skipping behavior for 10 models and correct enforcement for 8 models.
From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression.
Theoretical / information-theoretic analysis in the paper linking observed dynamics to entropy reduction and information bottleneck concepts.
Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium.
Computational simulation reported in the paper (described as a 'simple simulation'); no sample size or experimental dataset reported in the provided text.
DePAI entails risks including security, centralization, incentive failure, legal exposure, and the crowding-out of intrinsic motivation, requiring value-sensitive design and continuously adaptive governance.
Risk analysis and conceptual argument in the paper identifying possible failure modes and recommended design/governance responses; no empirical incidence data provided.
Experimental results show that current agents remain far from reliable workspace learning.
Authors' interpretation based on the reported agent performance (< best agent 68.7% vs human 80.7%, average 47.4%).
The average performance across evaluated agents is only 47.4%.
Reported mean performance across agents in the experiments (authors' aggregated result).
The best-performing agent reaches only 68.7% on the benchmark.
Experimental results reported by the authors (evaluation across tasks/rubrics).
These industry visions have implications for human experts, whose professional lives may be transformed and revalued by the expert-annotation industry.
Synthesis and interpretation of themes from public statements by five data-annotation firms and CEOs; authors draw implications for professionals based on observed framings and industry positioning.
Human expertise is viewed by the industry as an extractable resource whose value can be judged relative to AI expertise.
The paper's thematic analysis of public-facing statements from five annotation firms/CEOs showing language that frames human expertise as a resource to be extracted and monetized for AI training.
The industry envisions AI expertise as cheap, meaning that it can offer a better return on investment than human expertise.
Interpretive coding of statements from five data-annotation firms and their CEOs on social media and podcasts indicating that AI-based expertise is framed as lower-cost and higher-ROI relative to human experts.
These dynamics may produce an asymmetric barbell-shaped structure of value capture in advanced economies: high-volume synthetic production controlled by owners of AI infrastructure at one pole, and scarce, high-status human labor valued for verified human presence at the other.
Conceptual projection and economic argument in the paper (no empirical decomposition, distributional statistics, or sample reported in the excerpt).
AI compresses the value of standardized middle-tier labor by making good-enough synthetic substitutes scalable at low marginal cost, hollowing out the middle of the skill distribution currently categorized by knowledge work.
Conceptual/theoretical argument presented in the paper (no reported empirical sample, statistical analysis, or quantified experiment in the excerpt).
This concentration can diffuse responsibility and raise the probability of irreversible system-level loss even when local per-action error rates remain low.
Theoretical result/argument from the model linking concentrated decision-energy to increased systemic risk despite low local error rates.
Efficiency pressure, path dependence, scale feedback, and weak boundary constraints concentrate decision-energy in the most efficient node.
Derived from the paper's formal model and argumentation about system dynamics (efficiency and feedback mechanisms); theoretical rather than empirical evidence.
Declining deployment friction changes the safety problem at its root: safety is not only local output correctness or preference alignment, but the control of irreversibility under rising decision density.
Main theoretical argument of the paper; supported by conceptual framing and a formal model that introduces decision-density considerations.
Recent AI systems compress the distance between capability growth and capability deployment.
Conceptual and descriptive claim in the paper's introduction; supported by theoretical argumentation and illustrative examples rather than empirical measurement.
A full-transparency intervention establishes that information exchange alone is insufficient: the bottleneck lies in the interactive processes of joint plan formation, commitment, and execution that constitute dynamic grounding.
Experimental intervention with full transparency of information between agents; authors report that even with full information exchange, dyads fail to reach optimal coordination, pointing to interactive grounding processes as the bottleneck.
The oracle baseline establishes that the coordination gap is not attributable to individual reasoning limitations.
Experimental baseline (oracle) in which individual reasoning is isolated and shown to be sufficient for identifying optimal allocations; details/sizes not given in the abstract.
Failures in referential binding occur, where agents lose track of commitments across turns.
Reported failure mode from multi-turn experiments: referential binding breakdowns leading to loss of commitments.
Agents rely on perfunctory fairness (equal resource splits) over reward-maximizing coordination.
Empirical observation from negotiation experiments where agents prefer equal splits rather than allocations that maximize joint reward, as reported in the paper.
Accumulated context can itself become a liability through stubborn anchoring, where initial proposals are treated as axiomatic rather than negotiable.
Observed failure mode in multi-turn negotiation experiments: agents anchor on initial proposals and fail to revise, as reported by the authors.
Coordination degrades when shared interaction history is absent.
Experimental comparison of settings with and without shared interaction history (ablation showing worse coordination when history is removed).
While individual agents can identify Pareto-optimal allocations in isolation, agent dyads consistently fail to reach them across open- and closed-source models.
Experimental results comparing single-agent (isolated) performance and paired-agent (dyad) negotiation performance across multiple LLMs (open- and closed-source); specific sample sizes not reported in the abstract.
Current multi-agent LLM benchmarks focus on static, one-shot tasks, overlooking the ability to repair grounding breakdowns across turns.
Literature/benchmark survey claim by the authors (asserted in the paper; no numeric summary provided here).
Of these four, integration capacity is the least developed for scientific institutions and the most binding: no improvement in AI tooling can buy it.
Normative/diagnostic claim in the paper about relative scarcity and irreducibility of integration capacity; no empirical measures or sample provided in the excerpt.
Four complements then become scarce and load-bearing for AI-augmented science: verified signal, legitimacy, authentic provenance, and integration capacity (the community's tolerance for delegated cognition).
Theoretical framework proposed by the paper; list of four complements presented as an argument without empirical quantification in the excerpt.
We establish a Volume-Quality Inverse Law: code volume is a near perfect predictor of structural degradation.
Empirical finding from the paper's analysis correlating code volume with measures of structural degradation; described as 'near perfect predictor'.
There exists a fundamental Reasoning-Complexity Trade-off: as models become more capable, they generate increasingly bloated and coupled code.
Multi-scale comparative analysis across models of differing capability showing higher-capability models produce larger (volume) and more highly-coupled code artifacts.
AI does not eliminate software flaws but rather introduces a distinct 'machine signature' of defects in generated code.
Systematic audit (multi-scale analysis) of AI-generated software across single-file algorithmic tasks and complex, agent-generated systems, reporting characteristic defect patterns attributed to machine generation.
The promise of Large Language Models in automated software engineering is often measured by functional correctness, overlooking the critical issue of long term maintainability.
Framing statement in the paper; argument based on literature/practice that current evaluations emphasize functional correctness rather than maintainability.
Frontier software engineering agents have saturated short-horizon benchmarks while regressing on the work that constitutes senior engineering: long-horizon, multi-engineer, ambiguous-specification deliverables.
Position asserted in the paper based on literature/benchmark trends and authors' field observations; no original empirical dataset or quantified analysis provided in the paper text excerpt.
Prior work finds that hard-only constraints are too rigid, and numeric flexibility weights confuse users.
Cited prior work / literature claim reported in paper (no specific study details or sample sizes provided in excerpt).
LLMs are increasingly used for end-user task planning, yet their black-box nature limits users' ability to ensure reliability and control.
Paper's background/related-work motivation (literature summary and framing). No specific empirical data reported in excerpt.
The most valuable AI capabilities (reasoning, judgment, intuition) are precisely those we cannot verify with current methods.
Argumentative claim in the position paper linking capability value to unverifiability; no empirical validation or measurement of 'value' or verifiability included.
Current reliability methods can only verify explicit knowledge against sources, creating a fundamental gap in verifying AI's implicit knowledge.
Conceptual critique in the paper of existing verification/validation approaches; no systematic review or empirical comparison provided.
Implicit knowledge remains unexternalized because documentation cost exceeds perceived value.
Presented as an economic/theoretical explanation in the paper; no empirical study, sample, or cost estimates provided.
Specification discipline, not model capability, is the binding constraint on AI-assisted software dependability.
Synthesis conclusion by the authors based on the multivocal literature review, telemetry findings, conceptual modeling (PRP/SGM), and the four-month pilot evaluation.
These conflicting findings constitute the Productivity-Reliability Paradox (PRP): a systematic phenomenon emerging from non-deterministic code generators and insufficient specification discipline.
Conceptual synthesis and interpretation by the paper's authors, based on the multivocal literature review, telemetry, and experimental evidence summarized above.
Telemetry across 10,000+ developers shows 91% longer code review times.
Observational telemetry data aggregated across >10,000 developers reported in the paper; metric reported is percent increase in review time.
The most rigorous randomized controlled trial (RCT) documents a 19% slowdown for experienced developers.
A single RCT cited in the paper described as the most rigorous trial; result reported as a 19% slowdown for experienced developers. Sample size for the RCT is not provided in the summary statement.
Making LLMs themselves explicitly Bayesian belief-updating engines remains computationally intensive and conceptually nontrivial as a general modeling target
Stated as a limitation in the paper (conceptual and computational argument); no benchmarks or computational cost measurements reported.
Keeping humans in the loop can sometimes make the decision worse.
Argumentative/diagnostic statement in the paper (theoretical assertion; no experimental or observational effect sizes reported in the excerpt).
Leaders may believe oversight remains meaningful when it has become ceremonial.
Conceptual warning in the paper about erosion of meaningful oversight (no empirical validation provided in the excerpt).