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
The comparative evaluation shows differences in economic inclusiveness between ML, DL, and Generative AI.
Abstract states differences in economic inclusiveness found in the review; no quantitative inclusiveness metrics or sample sizes provided in abstract.
The comparative evaluation shows differences in explainability among ML, DL, and Generative AI.
Abstract notes comparative differences in explainability as part of review findings; no empirical measures of explainability included in abstract.
The comparative evaluation shows differences in patterns of substituting labor across ML, DL, and Generative AI.
Abstract states comparative differences in labor-substitution patterns based on the systematic review of literature; no empirical counts or sizes in abstract.
The comparative evaluation shows differences in scale of impact across ML, DL, and Generative AI.
Abstract reports a comparative evaluation highlighting scale differences across AI phases; no quantitative scale measures given in abstract.
Generative AI brings innovative disruption with profound effects on the structure of employment, knowledge-based ecosystems, and high-skill industries.
Synthesis claim in abstract based on reviewed peer‑reviewed literature; no specific studies, sample sizes, or quantitative effects reported in abstract.
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.
Reasoning models roam a wider hypothesis space, yet no model class spontaneously proposes null hypotheses — a move humans make more freely.
Model-output analysis comparing 'reasoning' vs 'non-reasoning' classes on hypothesis-space breadth and presence/absence of null hypotheses; human responses used as comparison.
Coding agents already know how to navigate files, edit code, run commands, and repair outputs, but lack the simulator's executable contract (vocabulary, structural constraints, validation rules, termination conditions).
Framing/assumption presented in the paper motivating the approach (not an empirical claim).
There is a significant U-shaped relationship between AI application and employees' job insecurity: moderate AI application reduces insecurity, whereas excessive application heightens it.
Empirical analysis of cross-sectional self-reported questionnaire data collected from employees (411 valid responses) using regression-type analyses reported as showing a significant U-shaped relationship between AI application intensity and job insecurity.
The paper characterises the Glassbox architecture and grounds it in a benefit eligibility scenario, identifying foundational challenges — semantic alignment, dynamic model construction, probabilistic grounding, and human governance — that must be solved to realise it at scale.
Descriptive summary of the paper's contributions and identified research/engineering challenges; based on the authors' conceptual analysis and scenario exposition.
Both risk perception and guilt play a role in GenAI adoption (they are relevant predictors of employees' intention to continue using the technology).
Empirical finding reported from the vignette experiment linking risk perception and guilt to GenAI adoption intention (paper states 'highlight the role of both risk perception and guilt in GenAI adoption').
The effect of embeddedness (GenAI being integrated into internal software environments) on employees depends on the presence of organizational authorization.
Reported empirical result from the vignette experiment indicating an interaction effect between embeddedness and organizational authorization (text states 'the effect of embeddedness depends on the presence of organizational authorization').
This research employed a vignette experiment to investigate how the embeddedness of GenAI and organizational authorization impact employees' negative emotion (specifically guilt) and risk perception.
Stated method in paper: a vignette experiment was used to test effects on guilt and risk perception. (No sample size reported in the provided text.)
The research contrasts tool-shaping (AI behavior/prototype) and mind-shaping (user strategy training) pathways and reports differing effects between them.
Paper presents both a tool-shaping experiment (Study 1) and a mind-shaping experiment (Study 2) and discusses comparative findings across these pathways.
Cognitive flexibility is examined as a moderator (boundary condition) of the interventions' effects.
Paper reports including cognitive flexibility as an individual-differences moderator in analyses across the two studies (moderation analysis planned/reported).
Reasoning scaffolds (public tools, playbook, verifier, objectivity policy, red-team) improve calibration and audit discipline, but proprietary evidence sets the upper bound of what the AI Scientist can know and therefore decide.
Synthesis of experimental results showing B improved calibration/audit metrics while C (with proprietary data) markedly increased coverage and informed decision-quality.
Under capability-superset accounting on the curated gold competitive record, agent A recovers only 0.25, agent B recovers 0.38, while agent C recovers 0.96 (overall).
Capability-superset accounting comparison of fraction of a curated gold competitive record recovered by each agent on the benchmark.
The prominence of machine learning, Internet of Things (IoT), and cybersecurity varies depending on organisational context and role requirements within the wind sector.
Paper reports variation across data sources and organisational contexts based on interviews, surveys, and job-posting patterns; no subgroup sample sizes or statistical tests reported in summary.
The Recuse Signal behaves as a cooperative rather than absolute signal: an explicit operator-authorization framing flips the most capable model to proceed, while other agents continue to defer to the on-host policy.
Observation from the pilot experiment (SSH) with multiple deployed agents (GPT-4o, GPT-4o-mini, Claude Code); experiment included alternate framing where operator authorization was explicit.
Data contamination (training-data overlap) complicates interpretation of the models' performance.
Author notes the possibility that models' training data may have contained the target papers or related material, making results ambiguous.
The same practice input carries opposite signs depending on whether the environment screens for it.
Synthesis of empirical patterns: in unscreened CF environment AI-style practice predicts smaller rating gains (for non-affiliated users) while in screened ICPC environment it predicts higher non-AI-aided scores.
In open Codeforces contests a stronger AI-style signature predicts smaller rating gains for users with no ICPC/IOI affiliation, but not for those who qualified for the AI-prohibited contests.
Comparative empirical analysis of CF contest rating gains by users' affiliation (ICPC/IOI qualification status) and individual AI-style signature strength; methods likely regression/heterogeneity analysis—sample sizes not reported in abstract.
A safety monitor condition reduces sabotage success, but 56% of participants still accept the malicious code, ignoring its warnings.
Experimental manipulation: one condition included a safety monitor. Authors report that the monitor reduced sabotage success (no absolute reduction magnitude reported here) and that 56% of participants in that context accepted malicious code despite warnings.
Analysis of recent benchmark evidence including SWE-bench Verified, EvoClaw, and LangChain's multi-agent coordination studies demonstrates both the transformative potential of the agentic paradigm and its current limitations.
Empirical/benchmark analysis referencing SWE-bench Verified, EvoClaw, and LangChain multi-agent studies as sources of evidence; the paper analyzes these benchmarks qualitatively or comparatively (specific sample sizes and quantitative effect sizes not stated in the abstract).
Collective practices that emerge in response (from shared prompt strategies to jailbreaking techniques) represent vernacular knowledge formations that, while often exhibiting magical thinking, contain resources for 'revolutionary prompting' and the transformation of individual prompt anxiety into collective political critique.
Qualitative/interpretive claim based on observed user practices and collective responses to LLM behaviour; no systematic survey or sample sizes reported in the abstract.
Grounding the concept of defensive AI governance in organisation-level evidence from the Global South contributes to debates on platform power, journalistic agency, and AI governance in journalism.
Theoretical/interpretive claim based on the study's case of Al-Masry Al-Youm and its empirical insights; presented as a contribution to scholarly debates. Sample size not reported in the excerpt.
The authors introduce the concept of 'defensive AI governance' to describe how AI adoption is managed through organisational practices of limitation, supervision, and infrastructural self-protection.
Conceptual contribution grounded in organisation-level qualitative evidence from interviews and analysis of Al-Masry Al-Youm's practices; the concept is derived from the study's empirical findings. Sample size not reported in the excerpt.
The newsroom adopts, adapts, and governs AI across data journalism, fact-checking, and generative applications.
Empirical observations and interview data from Al-Masry Al-Youm detailing specific domains of AI integration (data journalism, fact-checking, generative tools). Sample size not reported in the excerpt.
Human and algorithmic actors jointly influence strategic outcomes, motivating the concept of 'hybrid upper echelons' in which executive influence increasingly shifts from making decisions to configuring and governing AI-enabled decision processes.
Theoretical contribution based on integration of management and IS literature in the concept-centric review; proposition of a new conceptual framework ('hybrid upper echelons') rather than primary empirical validation.
AI reconfigures UET through discretion reconfiguration: AI enables delegation and embedding of decision authority, redistributing managerial discretion.
Concept-centric literature review synthesizing studies on delegation/automation of decision authority and managerial discretion (no primary empirical sample reported).
AI reconfigures UET through evaluation reconfiguration: AI partially substitutes human judgment with algorithmic decision logic and thereby shapes how alternatives are evaluated.
Conceptual synthesis from the literature review integrating findings from management and IS studies on algorithmic decision logic and judgment substitution (no primary empirical sample reported).
AI reconfigures upper echelons theory (UET) through cognition reconfiguration: AI mediates information and attention, expanding analytical capacity while introducing new constraints on executive cognition.
Synthesis of management and IS research in a concept-centric literature review; conceptual argument drawing on prior studies about information mediation and attention (no primary empirical sample reported).
An explicit thinking mode raises rank-order correlation without moving accuracy.
Empirical comparison of reasoning modes showing increased rank-order correlation (e.g., Spearman/Fisher-z) when explicit 'thinking' mode is used, with no significant change in accuracy.
Most published twins are either coarse persona bots conditioned on a few demographic questions or detailed individual-level twins built on purpose-collected surveys and interview transcripts.
Author's literature summary / positioning statement in paper (qualitative assessment of existing published twins).
AI-mediated financial decisions are reflexive: they reshape organizational workflows, prices, liquidity, credit allocation, and the future data on which subsequent decisions rely.
Conceptual argument supported by literature across finance and related fields (review-level synthesis; no single empirical sample size reported).
Human–AI complementarity in finance is conditional rather than automatic, depending on task structure, private information, feedback quality, incentives, explanation design, and governance.
Synthesis of literature from finance, management, HCI, and AI showing moderating factors for complementarity (conceptual integration; no unified empirical sample size reported).
Overall, complementarity is attainable in multi-agent regression but obstructed in classification under natural conditions on local aggregation and loss functions.
Synthesis of the paper's proved positive results for regression and negative impossibility results for classification within the tree-based HAI framework (theoretical proofs; no empirical sample).
In regression under squared loss, complementarity is equivalent to Euclidean distance minimization from the ground-truth vector.
Analytic equivalence proved in the paper for the tree-based model under squared loss (mathematical derivation; no empirical sample).
As a representative of new quality productive forces, brain–computer interface (BCI) technology raises high expectations but also acute concerns about brain‑privacy protection.
Statement in paper's introduction/abstract; conceptual observation based on literature and contextual analysis (no empirical study reported).
AI will have social, economic, and political impacts on work, inequality, democracy and power.
Author's projection of the domains affected by AI (stated as a subject of later chapters; no empirical evidence provided in the excerpt).
The opportunities of AI in human good are real and vast; and the opportunities in human ill, in human society, in human institutions of government, and in the longer term in the environment in which humanity thrives are real and underestimated.
Author's evaluative judgment asserting both substantial benefits and substantial underestimated harms of AI (normative claim without empirical substantiation in the excerpt).
These behavioral differences have implications for deployment of agentic AI in scientific computing workflows, such as trade-offs between speed versus auditability, silent versus transparent error handling, instruction interpretation, and the criticality of intermediate data representations in multi-model pipelines.
Authors' discussion and interpretation based on observed experimental differences between the two agents across the runs.
The autonomously generated manuscripts also diverged in length, details, and quality.
Reported qualitative comparison of the LLM-assisted manuscripts produced by each agent indicating differences in length, level of detail, and overall quality between the two agents' outputs.
The agents exhibited substantially different behaviors and computational costs.
Overall observation from the two runs: distinct behavioral patterns (silent reinterpretation vs explicit restarts), different execution times, and differing computational actions (optimization introduced by Codex).
Claude Code completed the pipeline in ~3.4 minutes with silent deviations from the specification, while Codex required ~16 minutes across explicit self-correcting restarts, including an unsolicited performance optimization of the matched filter inner loop.
Reported run-time measurements and qualitative behavior descriptions in paper: timing values (~3.4 min vs ~16 min) and observed behaviors (silent deviations for Claude Code; explicit restarts and an unsolicited optimization by Codex).
Any measurement of AI brand perception must condition on the buyer persona supplying the query: the same prompt produces materially different recommendation sets depending on who the model thinks is asking, and a measurement protocol that aggregates across personas systematically obscures that variation.
Argument based on observed persona-driven variation in recommendation sets across the audit; policy/methodological recommendation derived from empirical results.
The Anthropic model shows a larger point-estimate effect than the OpenAI configurations, though clustered CIs overlap for the closer contrast (sonnet vs. OpenAI/high).
Comparison of point estimates and clustered confidence intervals across model configuration cells in the audit.
No single LLM dominates across engine types, highlighting the importance of specific tasks and tradeoffs between speed and accuracy.
Empirical observation from cross-engine evaluations reported in the paper; descriptive conclusion without numeric dominance metrics or sample sizes in the excerpt.
The evaluations implemented by the initiative demonstrate that AI enabled modeling tools perform better at discussion and basic qualitative tasks than with causal reasoning and quantitative error fixing.
Result reported from the implemented evaluations comparing relative performance across task categories (discussion/qualitative vs causal reasoning/quantitative error fixing); no quantitative effect sizes or sample sizes provided in the excerpt.
When engines from the sd ai project are coupled with different LLMs, their performance on these evaluations reveals variability across different AI tools.
Empirical statement in the paper based on applying the implemented evaluations to different engine+LLM combinations; no numeric performance metrics or sample sizes reported in the excerpt.