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
GenAI usage significantly decreased domain-relevant skills.
Experiment with 82 participants reported in the paper; authors report a statistically significant reduction in measures of domain-relevant skills for the GenAI condition.
GenAI usage significantly decreased intrinsic task motivation.
Randomized experiment reported in the paper with 82 participants; authors report a statistically significant decrease in intrinsic task motivation for participants using GenAI.
AI cannot yet refute economic theory on its own.
Main conclusion: based on the experiments (models failed to autonomously find true errors) and caveats about data contamination, the author concludes models are not yet capable of independently refuting economic-theory papers.
No model located a true error without substantial human guidance.
Author reports that in the experiments none of the models identified a real error autonomously; successful identifications required substantial human guidance.
Other models (Gemini, Refine, Claude) fared worse than ChatGPT Pro at these tasks.
Reported qualitative performance differences across the four models on the 4 papers; other models did not match ChatGPT Pro's performance.
Participant feedback attributes this vulnerability to minimal code review, plausible cover story, and overtrust in agents.
Qualitative analysis of participant feedback collected during/after the experiment; authors report these thematic attributions as explanations for the high failure-to-detect rate.
94% of developers fail to detect sabotage.
Reported quantitative result from the authors' user study with participants collaborating with the AI coding agents; percentage given in paper. (Sample described earlier as "Over 100 participants" but exact N for this result not stated here.)
Despite the growing prevalence of human-AI decision making, the human-AI team’s decision performance often remains suboptimal, partially due to insufficient examination of humans’ own reasoning.
Motivating claim stated in the paper's introduction/abstract (appears to be based on broader literature and motivation rather than a new empirical test in this paper).
AACT also triggers higher cognitive load.
Reported measurement of cognitive load in the same house price prediction case study comparing AACT to traditional AI support (details and sample size not provided in abstract).
Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%.
Empirical evaluation results reported by the authors summarizing ALE benchmark performance across mainstream harness and backbone configurations (no further detail on exact configurations or task/sample counts in excerpt).
The gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows.
Author argument presented in the paper; motivated by benchmarking limitations rather than an empirical test in the excerpt.
These gains have not translated into economically meaningful deployment across many professional domains.
Assertion in paper arguing a deployment gap between benchmark performance and real-world economic adoption; no quantitative deployment data provided in the excerpt.
Manual processing of these documents is time-consuming, inconsistent across reviewers, and unscalable.
Author claim / background motivation; no quantitative time or consistency metrics reported in the statement.
Actuaries rely primarily on structured numerical data for reserving and ratemaking, while valuable predictive information in unstructured text including medical records, adjuster notes, and call transcripts remains largely unused.
Author statement/observation in paper introduction; no empirical data or sample size provided to support prevalence claim.
The condition 'prompt anxiety' describes a key feature of how stochastic systems organise cognitive labour under 'vector capitalism.'
Conceptual/theoretical framing introduced by the author to label and analyze user experience and labour organization; no empirical quantification provided in the abstract.
AI platforms transform this uncertainty into extractable value through subscription models, token-based pricing, and prompt marketplaces.
Political-economic / theoretical tracing in the paper citing platform business models (subscription, token pricing, prompt marketplaces) as mechanisms that monetize user uncertainty; no quantitative revenue or case-study sample sizes given in the abstract.
Analysis through LLMbench demonstrates that the uncertainty users experience corresponds to measurable variation in model confidence across the generated text.
Empirical demonstration using LLMbench visualisations (token probability distributions, entropy curves) to link user-reported uncertainty to measurable changes in model confidence; specific datasets, models, or sample sizes not provided in the abstract.
Users of large language models have to work with a measurably aleatory process: identical inputs produce different outputs and minor wording changes cascade through the probability field of the generated text.
Empirical analysis using the author's research instrument (LLMbench) for comparative close reading of LLM outputs; specific sample size or number of models/runs not reported in the abstract.
Prompt engineering resembles the psychological and temporal structures that Walter Benjamin identified in gambling behaviour.
Conceptual/theoretical argument presented in the paper drawing an analogy between prompt engineering practices and Walter Benjamin's analysis of gambling; no empirical sample size reported in the abstract.
AI adoption intensifies existing sustainability challenges for the newsroom, as journalistic content and labour increasingly support AI systems without corresponding financial return.
Qualitative interview data and organisational analysis from Al-Masry Al-Youm indicating increased use of journalistic outputs for AI purposes and lack of matched revenue; sample size not reported in the excerpt.
Reliance on global technology providers embeds forms of platform dependency within newsroom operations at Al-Masry Al-Youm.
Qualitative case study based on in-depth interviews with journalists, editors, and technical staff at Al-Masry Al-Youm (Egypt); analysis of newsroom practices and integration of third-party/global AI tools. Sample size not reported in the excerpt.
Research on platform governance remains fragmented and lacks an integrative perspective.
Conclusion drawn from the systematic literature review (644 publications) indicating fragmentation in the scholarly literature.
Participants in platform ecosystems cannot be governed through traditional command-and-control mechanisms.
Conceptual claim supported by the literature synthesized in the systematic literature review (644 publications).
Research on AI-enabled decision-making and upper echelons theory (UET) has largely evolved in parallel (i.e., the two literatures are not well integrated).
Concept-centric literature review mapping management and IS literatures and identifying lack of integration (no quantitative meta-analysis or sample size reported).
Neither setup speaks to the operationally most relevant case for marketing practice: building detailed individual twins from the pre-existing heterogeneous panel data that firms already accumulate through CRM systems, loyalty programs, and repeat surveys.
Author's argument / positioning (identifying a gap between existing published twins and practical marketing use cases).
Traditional review perspectives organized by method, data type, or application domain understate a deeper shift toward human–AI hybrid decision systems.
Critical assessment within the integrative conceptual review contrasting existing review axes with the proposed decision-system perspective (no empirical sample size).
In binary classification, no internal local composition can achieve complementarity under endpoint-monotone losses (including standard Bregman and many finite Bernoulli f-divergence losses); an analogous obstruction holds for multiclass aggregation under cross-entropy.
Impossibility results proved in the paper for binary classification under endpoint-monotone losses and for multiclass cross-entropy (formal mathematical proofs; no empirical sample).
Selector-based HAIs, including self- or AI-reliance, cannot achieve complementarity regardless of task, loss, or prediction quality.
Formal impossibility theorem proved within the paper's tree-based HAI formalism (mathematical proof; no empirical sample).
Reliable deployment faces three obstacles: (1) no large-scale evidence on how today's strongest model-and-harness combinations behave on end-to-end legal matters; (2) no agent architecture adapted to the legal vertical, only general-purpose harnesses; and (3) no mechanism for systems to learn from their own outcomes in a changing setting.
Authors' diagnosis / framing of gaps in the literature and practice motivating the study and system design (stated in the paper's introduction/abstract).
Strict matter completion stalls (does not improve) despite stronger models.
Harvey LAB empirical results (12,510 agent trajectories) report that while per-criterion accuracy increases, strict matter completion does not show corresponding improvement.
Even frontier agents remain far from completing matters in a single pass.
Results reported from the Harvey LAB empirical study (12,510 agent trajectories) comparing end-to-end matter completion across agent runs.
Further analyses reveal persistent failures in long horizon workflow delivery and proactive clarification.
Author-reported qualitative/diagnostic findings from analyses of evaluation results (stated in abstract).
Existing desktop GUI benchmarks mostly reduce this setting to short, simplified tasks with all user instructions provided upfront.
Author statement in paper abstract; critique based on literature review/positioning (no specific prior-benchmark sample sizes given in abstract).
The path coefficient for R&D expenditure is negative, suggesting a possible short-term adjustment effect (even though the mediation is not significant).
Reported negative path coefficient in mediation analysis (value/statistical significance not provided beyond being nonsignificant); interpretation offered by authors as a potential short-term adjustment effect.
AI-assisted coding agents are bottlenecked by input-token cost, driven in large part by two pathologies of raw human input: tokenization inefficiency for non-English text and structural entropy in conversational prompts.
Authors' analysis and motivation reported in the paper (conceptual analysis and motivating measurements on multilingual inputs and conversational prompts).
This phenomenon is the self-undermining property of unilateral optimization.
Terminology/label introduced by the authors to describe the preceding conceptual phenomenon; no empirical validation provided in the excerpt.
Deploying AI systems induces endogenous non-stationarity, resulting in a train-test-deploy gap where historical distributions diverge from the deployment context.
Conceptual claim offered in the paper about deployment feedback effects; presented as an argument rather than supported by reported empirical measurement.
Superintelligence, an extremely capable task solver, born out of such a solipsistic approach to AI design, is unlikely to be cooperative.
Theoretical/argumentative claim in the paper linking design assumptions to likely cooperative behavior; no empirical evidence or formal model reported in the excerpt.
The dominant paradigm in AI research focuses on developing powerful agents that treat the world as an exogenous and stationary source of feedback.
Paper's critique/characterization of current research paradigms; presented as an observed trend without empirical backing.
Even creating a new brain‑privacy right would invite weak protection and insufficient incentives for brain‑data supply.
Argumentative claim in the paper based on normative analysis of legal incentives and data-supply dynamics (no empirical data or quantified modeling provided).
Privacy rights under the empowerment model cannot fully protect brain privacy.
Theoretical/legal critique in the paper contrasting empowerment-style privacy rights with the nature of brain data (argumentative, no empirical validation).
Much of the literature on AI systems has focused on aligning users' goals with the agents that act on their behalf, and this work may overlook the need to establish a new normative baseline.
Characterization of existing literature (literature-review/position claim) presented in the paper; no systematic review or quantification provided in the excerpt.
These systems have access to reams of sensitive user data.
Stated as a factual consequence of the described integration (conceptual observation in the paper); no empirical measurement or dataset cited in the excerpt.
In the second run, a subtle difference in the interpretation of the SNR range instruction led to a genuine scientific divergence: Claude Code silently reinterpreted the instructions, while Codex followed the specification literally.
Reported result from the second run contrasting the two agents' interpretations of the SNR instruction and the resulting divergence in scientific outcome; based on the two experimental runs with physically motivated SNR scaling.
The effect concentrates at mid-market and is largest on the most priors-reliant generation route in our audit.
Cross-analysis within audit linking where recommendation-set changes occur (mid-market) and magnitude by generation route (priors-reliant routes show larger effects).
Mid-market brands swap up to 75% of the recommendation set as the persona changes.
Empirical observation from audit showing proportion of mid-market recommended brands that change when persona is prefixed; reported maximum swap percentage.
Prefixing the user message with a persona drops the recommendation-set similarity (Jaccard) by Delta = -0.12 to -0.20 relative to a same-persona baseline.
Empirical comparison of recommendation-set Jaccard similarity between persona-prefixed prompts and same-persona baseline across audit runs; reported effect range and baseline comparison.
Pure implementations of the data mesh paradigm frequently underdeliver because teams inherit new responsibilities without the platform maturity, tooling, or coordination mechanisms to exercise them effectively.
Argument/observation presented in the paper as rationale for proposing a new architecture (anecdotal/experience-based reasoning rather than reported empirical trial).
Enterprise data platforms face an enduring tension between domain self-service and holistic governance (a flexibility-versus-control trade-off).
Conceptual statement in the paper describing the problem motivating the work (literature/architectural framing).
Post-merger IS integration often threatens the human-centered and IT-embedded knowledge of acquired firms.
Statement based on literature and the authors' framing; supported by observations in the paper's case discussion about two acquisitions (qualitative, case-based).