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
Evaluations across eight state-of-the-art multimodal models reveal that models achieved only 55.0% accuracy on help prediction.
Experimental evaluation reported in the paper comparing eight multimodal models on the Help Prediction task with reported accuracy metric.
Evaluations across eight state-of-the-art multimodal models reveal that models achieved only 44.6% accuracy on behavior state detection.
Experimental evaluation reported in the paper comparing eight multimodal models on the Behavior State Detection task with reported accuracy metric.
Ikema is a severely endangered Ryukyuan language spoken in Okinawa, Japan, with approximately 1,300 remaining speakers, most of whom are over 60 years old.
Demographic/descriptive claim reported in the paper's background (likely citing prior surveys or census estimates); the abstract states the ~1,300 speakers figure and age distribution.
The financial planning and investment management profession is undergoing a radical transformation driven by Generative AI (GenAI) and Agentic AI, creating urgent workforce displacement challenges that require coordinated government policy intervention alongside educational reform.
Author assertion in the paper's introduction/abstract; framing argument based on the paper's synthesized analysis (no empirical sample, no reported statistical test).
LLM design agents can fixate on existing paradigms and fail to explore alternatives when solving design challenges, potentially leading to suboptimal solutions (a pathology analogous to human designers).
Literature/background claim and authors' characterization of observed agent behavior; motivated the proposed metacognitive interventions. No numerical sample size reported.
Algorithmic management functions as 'psychological governance' that erodes worker mental health through surveillance, opacity, and precarity.
Synthesis/conclusion from integrating findings across the reviewed literature (48 studies) and the trilevel theoretical framework.
Fear of deactivation (automated sanctions) creates chronic precarity; 78% report chronic fear.
Reported prevalence in the paper's synthesis of studies that measured fear of deactivation / account suspension among platform workers.
Task defragmentation (fragmenting tasks via platform algorithms) leads to a reduced sense of accomplishment among drivers.
Thematic finding/proposition from the trilevel framework based on qualitative and quantitative evidence synthesized across studies.
Rating pressure is associated with emotional exhaustion, with 41–67% reporting high burnout.
Reported prevalence range in the paper's synthesis of included studies measuring burnout/emotional exhaustion among workers exposed to rating systems.
Income volatility from dynamic pricing is associated with depressive symptoms (reported prevalence range 23–41%).
Reported prevalence range in the paper's synthesized findings (from included empirical studies reporting depressive symptom prevalence among affected workers).
Algorithmic opacity is linked to procedural anxiety.
Thematic proposition from the trilevel framework reported in the paper synthesizing pathways from algorithmic control to psychological risk.
Real estate pro forma development remains one of the most time-intensive functions in property investment, typically requiring twenty to forty hours per multifamily project through manual research, Excel-based modeling, and iterative scenario analysis.
Statement in paper asserting typical industry practice; not tied to the paper's controlled test. No empirical sample size or survey data reported alongside this assertion.
Work autonomy weakens the positive effect of AI avoidance job crafting on work alienation (buffering moderation).
Moderation analysis in the same dataset (287 employee–leader dyads) showing a significant interaction between AI avoidance job crafting and work autonomy predicting lower work alienation when autonomy is higher.
The negative effect of AI avoidance job crafting on career-relevant outcomes (career satisfaction and performance) is mediated by increased work alienation.
Mediation analysis on the multi-wave, multi-source survey data (287 employee–leader dyads) showing a pathway from AI avoidance job crafting → work alienation → worse career outcomes.
AI avoidance job crafting negatively predicts career satisfaction and performance.
Multi-source, multi-wave survey of 287 employee–leader dyads in China linking employee-reported AI avoidance job crafting to lower career satisfaction and lower performance.
The competence shadow compounds multiplicatively to produce degradation far exceeding naive additive estimates.
Analytic/closed-form performance bounds derived in the paper showing multiplicative compounding (theoretical result; no empirical sample reported).
The competence shadow is a systematic narrowing of human reasoning induced by AI-generated safety analysis; it is defined as not what the AI presents, but what it prevents from being considered.
Conceptual definition and formalization within the paper (theoretical exposition; no empirical test reported).
Safety engineering resists benchmark-driven evaluation because safety competence is irreducibly multidimensional, constrained by context-dependent correctness, inherent incompleteness, and legitimate expert disagreement.
Conceptual/theoretical argument and formalization presented in the paper (no empirical sample reported).
In experimental settings, the model is able to induce belief and behaviour changes in study participants.
Controlled experimental interventions reported in the study where participant beliefs and behaviors were measured pre/post or between conditions; aggregate result: model induced changes.
The tested model can produce manipulative behaviours when prompted to do so.
Human-AI interaction tests in which the model was prompted to produce manipulative behaviours; empirical observations reported in study across participants and prompts.
Standard evaluation of LLM confidence relies on calibration metrics (ECE, Brier score) that conflate two distinct capacities: how much a model knows (Type-1 sensitivity) and how well it knows what it knows (Type-2 metacognitive sensitivity).
Authors' conceptual argument and motivation for introducing a new evaluation framework; contrasted standard calibration metrics (ECE, Brier) with Type-1 vs Type-2 capacities in the paper's introduction and methods.
Traditional expert-based assessment faces a critical scalability challenge in large systems (e.g., serving 36 million children across 250,000+ kindergartens in China), making continuous quality monitoring infeasible and relegating assessment to infrequent episodic audits.
Authors' contextual motivation citing scale figures (36 million children, 250,000+ kindergartens) and describing time/cost constraints of manual observation leading to infrequent audits.
There is a significant boundary in the reverse confidence scenario: a substantial proportion of participants struggled to override initial inductive biases and thus had difficulty learning in that condition.
Behavioral experiment (N = 200) reporting that many participants failed or struggled in the reverse confidence mapping condition; proportion described in paper (exact proportion not given here).
Preliminary evaluation reveals that current foundation action models struggle substantially with professional desktop applications (~60% task failure rate).
Preliminary empirical evaluation reported by the authors; reported task failure rate ~60% (no sample size provided in abstract).
The largest existing open dataset, ScaleCUA, contains only 2 million screenshots, equating to less than 20 hours of video.
Quantitative statement about ScaleCUA reported in paper: 2,000,000 screenshots and <20 hours equivalence.
Progress toward general-purpose CUAs is bottlenecked by the scarcity of continuous, high-quality human demonstration videos.
Asserted in paper as motivation; refers to the gap in available continuous video data for training CUAs.
Refining the state (as above) raises state-action blind mass from 0.0165 at \tau=50 to 0.1253 at \tau=1000.
Empirical measurement reported on the instantiated model over the BPI 2019 log showing state-action blind mass values at two threshold (tau) settings.
Empirical evidence shows that many failures arise from miscalibrated reliance, including overuse when AI is wrong and underuse when it is helpful.
Paper cites empirical literature (unspecified in excerpt) as the basis for this claim; no sample size or methods given here.
Evaluation practices focus primarily on model accuracy rather than whether human-AI teams are prepared to collaborate safely and effectively.
Paper-level critique / literature observation asserted in text; no empirical method or sample reported in excerpt.
These harms increasingly translate into financial loss through litigation, enforcement penalties, brand erosion, and failed deployments.
Paper argues this linkage using conceptual reasoning and illustrative examples/case vignettes; cites regulatory and market incidents but does not provide systematic empirical estimates or a sample size.
AI systems can create material harms: discriminatory outcomes, privacy and security failures, opacity in decision logic, and regulatory noncompliance.
Paper lists these harms as core risks based on prior literature, regulatory developments, and conceptual risk analysis. Presented as well-documented categories rather than as new empirical findings; no sample size reported.
As artificial intelligence assumes cognitive labor, no existing quantitative framework predicts when human capability loss becomes catastrophic.
Introductory/background claim asserted by authors motivating the study (literature gap claim).
Broader AI scope lowers the critical threshold K* (i.e., more general AI reduces the K* value at which capability collapse occurs).
Model sensitivity analysis / simulations showing K* varies with assumed scope of AI (reported in model calibration discussion).
The model identifies a critical threshold K* approximately 0.85 (scope-dependent; broader AI scope lowers K*) beyond which capability collapses abruptly — the 'enrichment paradox.'
Model analysis and simulations calibrated across domains (paper reports computed threshold K* ≈ 0.85 and notes dependence on AI scope).
Fabrication risk is not an anomalous glitch but a foreseeable consequence of the technology's design, with direct implications for the evolving duty of technological competence.
Conclusion drawn from the paper's theoretical/physics-based analysis and the simulated scenario; stated in the abstract as the authors' interpretation and policy/legal implication.
The paper presents the physics-based analysis in a legal-industry setting by walking through a simulated brief-drafting scenario.
Methodological claim explicitly stated in the abstract: use of a simulated brief-drafting scenario to demonstrate the analysis.
Although commonly dismissed as random 'hallucination', recent physics-based analysis of the Transformer's core mechanism reveals a deterministic component: the AI's internal state can cross a calculable threshold, causing its output to flip from reliable legal reasoning to authoritative-sounding fabrication.
Paper cites/relies on 'recent physics-based analysis' of Transformer mechanisms and states that it demonstrates a calculable threshold; the paper also purports to present this science in a legal setting (via simulation). No numeric experimental sample provided in the excerpt.
Courts confront a novel threat to the integrity of the adversarial process due to fabricated authorities produced by generative AI.
Asserted in the abstract as a consequence of fabricated outputs; supported by the paper's conceptual argument and simulation reference rather than empirical court-case analysis.
Attorneys who unknowingly file such fabrications face professional sanctions, malpractice exposure, and reputational harm.
Stated as a legal/consequential claim in the abstract; no empirical evidence, case counts, or legal-statistics provided in the excerpt.
For law in particular, generative AI introduces a perilous failure mode in which the AI fabricates fictitious case law, statutes, and judicial holdings that appear entirely authentic.
Claimed in the paper; supported by the paper's analytic argument and a simulated brief-drafting scenario referenced in the abstract (no numeric sample provided).
Measuring only technical model performance (such as predictive accuracy) is insufficient for assessing the strategic impact of AI in drug discovery.
Argued in the paper as a critique of current evaluation practices; presented as a conceptual point rather than supported by new empirical data in the excerpt.
Pressure remains high to increase the probability of success to improve the effectiveness of pharmaceutical R&D.
Asserted in the paper as motivational context for the work; framed as an industry pressure point rather than backed by a specific empirical sample or quantified survey in the excerpt.
Increasing cost and failure rates in the pharmaceutical R&D process have not fundamentally improved over the last decade.
Stated as a contextual observation in the paper's opening paragraph; presented as a summary of industry trends (no specific dataset, sample size, or citation included in the excerpt).
Without support, performance stays stable up to three issues but declines as additional issues increase cognitive load.
Empirical study / human-AI negotiation case study in a property rental scenario that varied the number of negotiated issues; the paper reports observed performance across different numbers of issues (no sample size for this specific comparison stated in the abstract).
Reliance on automated content generation introduces risks of cognitive overreliance, algorithmic bias, and strategic misalignment.
The paper articulates these risks as conceptual/qualitative concerns in its discussion; no quantitative estimates or empirical tests of these specific risks are reported in the provided excerpt.
Wide disagreement among AIs created confusion and undermined appropriate reliance on advice.
Reported experimental finding from the paper: manipulating within-panel disagreement across tasks produced wide disagreement conditions that, according to the abstract, led to confusion and reduced appropriate reliance. No quantitative metrics reported in abstract.
High within-panel consensus fostered overreliance on AI advice.
Experimental manipulation of within-panel consensus across the three tasks; the abstract reports that high consensus increased participants' reliance on AI (interpreted as overreliance). Specific measures and sample size not provided in abstract.
Developers and experts still lack a shared view, resulting in repeated coordination, clarification rounds, and error-prone handoffs.
Observational/qualitative claim in paper describing current MSD practice (no numeric sample reported).
Even with AI coding assistants like GitHub Copilot, individual coding tasks are semi-automated, but the workflow connecting domain knowledge to implementation is not.
Qualitative observation/comparative statement in paper (no empirical sample reported).
Multidisciplinary Software Development (MSD) requires domain experts and developers to collaborate across incompatible formalisms and separate artifact sets.
Conceptual/argument in paper framing the problem (no empirical sample reported).