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
Generative AI serves as an effective 'wingman' for employment lawyers, capable of replacing substantial junior associate work while requiring continued human expertise for client counseling, supervision, and final legal advice preparation.
Authors' synthesis of experimental results showing AI-produced substantive analysis plus discussion about remaining limitations (e.g., citation errors) and required human oversight; qualitative assertion about substitutability for junior associate tasks.
PPS gains are task-dependent: gains are large in high-ambiguity business analysis tasks but reverse in low-ambiguity travel planning tasks.
Task-level analysis across the three domains (business, technical, travel) within the controlled study (60 tasks total); authors report differential performance patterns by domain/ambiguity.
AI usage has dual effects on employees: it can both enhance innovative behavior and predict disengagement, as revealed by a dual-path (SOR-based) model.
Interpretation/synthesis from the four-stage longitudinal study of 285 finance professionals using a dual-path model based on SOR theory (combining the mediation and moderation results).
We evaluate 14 LLMs under zero-shot prompting and retrieval-augmented settings and witness a clear performance gap.
Experimental evaluation reported in the paper: authors state they ran experiments on 14 different large language models, under zero-shot and retrieval-augmented configurations, and observed differing performance across models.
Artificial intelligence embedded in human decision-making can either enhance human reasoning or induce excessive cognitive dependence.
Stated as a conceptual claim in the paper's introduction/abstract; supported by the paper's conceptual framing (theoretical argument), no empirical sample or experimental data reported here.
These productivity gains are most pronounced for lower-skilled workers, producing a pattern the authors call “skill compression.”
Cross-study pattern reported in the literature review: comparative evidence across worker-skill strata in multiple empirical papers showing larger relative gains for lower-skilled/junior workers; specific underlying studies and sample sizes are not enumerated in the brief.
Study 1 quantifies confirmation bias through controlled experiments on 250 CVE vulnerability/patch pairs evaluated across four state-of-the-art models under five framing conditions for the review prompt.
Controlled experiment described in the paper: 250 CVE vulnerability/patch pairs evaluated across four state-of-the-art LLMs under five prompt framing conditions.
These findings challenge the narrative of complete automation by AI and underscore the enduring importance of human expertise in data science.
Interpretation based on competition results where AI-only baselines underperformed relative to many participant teams and top solutions used human-AI collaboration.
These findings indicate a misalignment between the perceived benefit of AI writing and an implicit, consistent effect on the semantics of human writing, with potential implications for cultural and scientific institutions.
Synthesis and interpretation of the paper's empirical results (user study, essay revision experiments, and peer-review analysis); presented as the paper's broader conclusion.
The paper formalizes the distinction using a signal-aggregation model in which an organization maintains an anchor belief and achieves agreement through two exclusion channels: (1) report shrinkage toward the anchor and (2) a tolerance rule that discards reports deviating beyond a threshold.
Analytical formal model presented in the paper specifying an anchor belief and two exclusion mechanisms; model assumptions and mechanisms are explicit in the theoretical development. No empirical sample.
Organizational cohesion is observationally ambiguous: it can arise either from genuine information integration (debate and synthesis of heterogeneous inputs) or from exclusionary processes (conformity pressure, gatekeeping, intolerance of dissent).
Conceptual argument and formal definition in the paper framing; supported by the analytic distinction introduced in the paper between integration and exclusion as alternative generative mechanisms for observed agreement. No empirical sample—argument is theoretical and illustrated by model construction.
The authors identify ten evaluation practices that teams use, ranging from lightweight interpretive checks to formal organizational processes (examples: qualitative user reviews, red-team testing, A/B experiments, telemetry/log analysis, structured annotation, governance/meta-evaluation).
Thematic coding of 19 interview transcripts produced a taxonomy enumerating ten practices (paper reports the taxonomy as an outcome).
The net educational value of AI-generated feedback depends on alignment with pedagogical goals, quality evaluation, integration with human teaching, and governance to manage equity, privacy, and incentives.
Synthesis statement from the meeting report produced by 50 interdisciplinary scholars; conceptual judgment rather than empirical proof.
Convergence after exemplar exposure occurred by both tightening of estimates within a measure family and by agents switching measure families.
Agent-level tracking across stages showed two patterns following exemplar exposure: (1) reduced within-family dispersion (tighter estimates) and (2) categorical switches in measure selection by some agents, as recorded across the 150-agent sample.
LLMs excel at extracting and generating arguments from unstructured text but are opaque and hard to evaluate or trust.
Synthesis of recent LLM literature and observed properties (generation capability vs. opacity); no empirical evaluation within this paper.
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.
If an LLM's full capability were reducible to an explicit rule set, that rule set would be an expert system; because expert systems are empirically and historically weaker than LLMs, this leads to a contradiction (supporting non‑rule‑encodability).
Logical proof‑by‑contradiction presented in the paper, supported by conceptual mapping between rule sets and expert systems and qualitative historical comparisons.
Teamwork partner type moderates the effect of service empathy on collaboration proficiency (i.e., the impact of service empathy on proficiency differs by human vs AI partner).
Reported interaction/moderated-mediation analyses from the online experiment (n = 861) indicating a significant partner-type × service-empathy interaction predicting collaboration proficiency.
Employees' emotional state significantly moderates the relationship between partner type (human vs AI) and collaboration proficiency.
Moderation analyses reported from the same online experimental dataset (n = 861), testing interaction terms between partner type and measured employee emotion on collaboration proficiency; authors report a significant moderating effect.
AI adoption has an inverted U-shaped effect on employee-related corporate social responsibility (ECSR).
Panel regression with quadratic specification (AI and AI^2) showing statistically significant positive coefficient on AI and statistically significant negative coefficient on AI^2; sample of 2,575 Chinese listed firms observed 2013–2023; controls, firm and/or year fixed effects and robustness checks reported.
Token overhead varies from modest savings to a 451% increase while pass rates remain unchanged.
Measured token usage for agent runs with and without skills, reporting a range from modest token savings up to a 451% token increase with no corresponding change in pass rates.
The research methodology combines systemic analysis, comparative assessment of international practices, and analytical generalization of organizational learning models, enabling capture of both structural trends and concrete institutional responses to technological changes.
Methodological statement from the paper describing its approach; this is a factual claim about methods used rather than an empirical finding.
Model output can be treated as evidence for studying human behavior, but there are important epistemic limits to interpreting model-generated text as direct evidence of human beliefs or social facts.
Epistemic analysis and methodological critique in the paper (discussion of limits of treating model outputs as evidence); no single empirical test cited in the provided text.
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 study's counterfactual analytical model links HR indicators (training intensity, absenteeism, labor productivity, turnover rates, workforce allocation) to organizational performance outcomes using regression-based simulations and predictive estimation.
Methodological claim explicitly stated: model construction from an industrial firm dataset using regression-based simulations and predictive techniques. (Specific sample size, variable operationalizations, and time frame not reported in the description.)
Helicoid dynamics is a specific failure regime: a system engages competently, drifts into error, accurately names what went wrong, then reproduces the same pattern at a higher level of sophistication, recognizing it is looping and continuing nonetheless.
Definition introduced in the paper and illustrated by the reported case series; the claim is conceptual/phenomenological rather than a statistical result.
A minimal linear specification (linearized model) demonstrates how coupling strength, persistence, and dissipation determine local stability and oscillatory regimes through spectral conditions on the Jacobian.
Analytic linear model and local stability analysis in the paper: computation of Jacobian, derivation of spectral conditions (eigenvalue locations) that separate stable/oscillatory regimes; illustrative examples within the paper (no empirical data).
Distinct AI features (recommendation engines, chatbots, and comparison tools) influence consumer outcomes when modeled as latent constructs.
Methodological claim: the study modeled three AI features as latent constructs and analyzed their relationships with dependent variables using SEM (quantitative questionnaire data).
Both time constraints and LLM use significantly alter the characteristics of decision-makers' mental representations.
Results from the 2 × 2 experiment (N = 348) comparing representation-related measures across manipulated conditions; reported statistically significant differences associated with time constraints and with LLM use.
We develop a theoretical framework - the productivity funnel - that traces how technological potential narrows through successive stages, from access and digital infrastructure, through organizational absorption and human capital adaptation, to ultimate value capture.
Conceptual/theoretical development presented in the paper; no empirical sample needed (framework-building).
Effects of curated Skills are highly heterogeneous across domains (e.g., +4.5 pp in Software Engineering vs. +51.9 pp in Healthcare).
Per-domain pass-rate deltas reported in the paper (SkillsBench per-domain analysis). The example domain deltas (+4.5 pp and +51.9 pp) are taken from the reported per-domain results.
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.
XChronos reframes transhumanist technology evaluation in experiential terms, creating both market opportunities and measurement/regulatory challenges for AI economics.
Synthesis and concluding argument in the paper summarizing proposed implications; conceptual reasoning without empirical tests.
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.
Participant targeting: 44% of programs targeted doctors and 44% targeted medical students (with possible overlap), and 56% targeted entry‑to‑practice career stages.
Participant audience and career-stage data extracted from the 27 included programs; proportions reported in the review.
Most programs were delivered in academic settings: 56% of evaluated programs reported an academic setting.
Setting information extracted from the 27 included programs, with 56% reported as delivered in academic settings.
A plurality of programs were short in duration: 44% of programs were categorized as short courses.
Extraction of program length from the 27 included studies; 44% were classified as short courses per the review's categorization.
Most programs were introductory in content: 67% of included programs taught introductory AI concepts rather than advanced/technical AI skills.
Program content extraction across the 27 included studies yielded that 67% were classified as teaching introductory AI.
The methodological landscape of the evidence base is heterogeneous, consisting of cross-sectional surveys, case studies, quasi-experimental designs, and a limited number of longitudinal analyses.
Study design information was extracted from the 145 included studies revealing a mix of designs and relatively few longitudinal or experimental studies.
Human factors (training, trust calibration, workflows) determine whether clinicians accept, override, or ignore GenAI suggestions.
Qualitative and quantitative human-AI interaction studies and pilot deployments discussed in the paper; specific sample sizes and effect sizes are not reported in the paper.
Safety and net benefit of GenAI CDS hinge on deployment details: user interface, real-time feedback, uncertainty quantification, calibration, and how recommendations are presented (strong vs. suggestive).
Human factors and implementation studies referenced; early A/B tests and human-AI interaction research suggest interface and presentation affect acceptance and error rates; no large-scale standardized implementation trial data cited.
Reimbursement models (fee-for-service vs. capitation) will influence whether cost savings from GenAI are realized or offset by increased service volume.
Economic incentive framework and prior health-economics literature cited; the paper does not provide direct empirical tests but references plausible incentive channels.
RL and adaptive methods are good for real-time adaptation but can be myopic, require large amounts of interaction data, and struggle to incorporate long-term preference structure and ethical constraints.
Surveyed properties of reinforcement learning and adaptive methods in HRI/RS literature; no new empirical evaluation in this paper.
The community knowledge functions both as practical how-to guidance and as collective experimentation with platform rules and revenue mechanisms.
Observed dual nature in the 377-video corpus: instructional workflows alongside demonstrations/testing of platform-tailored monetization tactics and workarounds.
Typical practices emphasized by creators include rapid mass production of content, productizing prompt engineering, repurposing existing material via synthesis/localization, and packaging AI outputs as sellable creative services or assets.
Recurring practices surfaced through qualitative coding of workflows, tools, and pipelines described in the 377 videos.
Across the 377 videos, creators converge on a set of repeatable use cases and platform‑tailored monetization tactics.
Thematic coding of 377 videos produced a catalog of recurring use cases and tactics; the paper reports convergence across that sample.
YouTube creators have collectively constructed and circulated a practical knowledge repository about how to monetize GenAI-driven creative work.
Systematic qualitative content analysis (thematic coding) of 377 publicly available YouTube videos in which creators promote GenAI workflows and monetization strategies.
Choice of scaffold materially affects outcomes: an open-source scaffold outperformed vendor-provided scaffolds by up to approximately 5 percentage points.
Comparative experiments across three scaffolding approaches (vendor scaffolds and at least one open-source scaffold) showing up to ~5 percentage point differences in measured outcomes.
Adoption of NFD approaches in regulated domains will depend on standards for validation, auditability, and update procedures.
Implications and governance discussion emphasizing regulatory constraints (finance, healthcare) and the need for validation/audit standards; logical/ normative claim rather than empirical finding.
Limitations include generalizability beyond Chatbot Arena data, calibration of priors on novel tasks, audit costs/latency, user comprehension/cognitive load, and strategic manipulation.
Authors' stated limitations and open questions; these are candid acknowledgements rather than empirical findings.