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 productivity gap is attributable to organizational and behavioral factors.
Theoretical analysis, generalization, and synthesis of corporate cases and empirical reports linking observed micro-behaviors to macro-level productivity outcomes.
There is a gap between anticipated macroeconomic efficiency gains (aggregate labor productivity) and observed micro-level outcomes following AI adoption.
Comparison of aggregate productivity trends (BLS series/AAPC calculations) with micro-level evidence drawn from corporate case studies and empirical reports documenting localized impacts of AI.
The CAD has implications for knowledge-work stratification and AI platform governance.
Argumentative/policy discussion in the paper linking the CAD to potential stratification among knowledge workers and governance considerations for AI platforms.
The probabilistic model demonstrates that manual context attachment leads to a combinatorial collapse in task-success probability as corpus size and task conjunctivity grow.
Results from the paper's probabilistic model (analytic/theoretical demonstration based on fan effect reasoning); no empirical sample reported.
For knowledge-intensive workers whose intellectual capital spans tens of thousands of files, the CAD constitutes a qualitative threshold in AI usefulness: below it, the cognitive burden of context curation falls on the human, reproducing the inefficiencies AI is meant to eliminate.
Theoretical argument grounded in the paper's conceptual discussion about large personal/organizational corpora (stated scale: tens of thousands of files) and the user burden of manual context attachment.
There exists a finer-grained divide at the level of individual interaction — the Context Access Divide (CAD) — whereby two users with nominally equivalent agent access may experience qualitatively different AI utility depending on whether the system can autonomously retrieve context (Dynamic Context Retrieval) or requires manual document attachment (Manual Attachment).
Conceptual argument and definitional framing in the paper introducing the CAD as a novel dimension of inequality; comparison of two interaction modalities (Dynamic Context Retrieval vs Manual Attachment).
Free overrides also cut sales by 1.19%.
Randomized field experiment comparing free-overrides arm to control; effect reported as 1.19% reduction in sales.
Free overrides reduce inventory by 1.95%.
Randomized field experiment comparing free-overrides arm to control; effect reported as 1.95% reduction in inventory.
The testing tool raised cost by 42 to 68 percent without improving functional score or reliability, even on interface visible criteria.
Comparison between runs with and without the testing tool showing reported cost increase (42–68%) and no improvement in functional score or reliability.
Container deployment was the dominant defect, failing first try in 44 percent of runs.
Criterion-level analysis of failure modes across the 90 runs reporting first-try failure frequency for container deployment.
Aligning AI systems with human teams remains a major challenge to realizing AI's full potential in organizations.
Authors' statement in abstract framing the motivation for the study; supported by literature cited in full paper (abstract asserts this as a core challenge).
A substantial and persistent gap below expert level reliability therefore remains.
Inference in abstract based on reported model accuracies (~82.7%) compared with the informal expert reference (~95%).
Against an informal expert reference of around 95%, obtained from a low sample quiz of aviation professionals at a conference, even the strongest model evaluated (released in 2026) reaches 82.7%.
Reported comparison in abstract between informal expert reference (~95%) and top model accuracy (82.7%) on the Pre-Flight benchmark; expert reference described as a 'low sample quiz' (sample size not reported). Model accuracy implicitly measured across benchmark questions (300).
The high stakes, regulated nature of the aviation domain makes the gap (in domain-specific evaluation) consequential.
Argumentative statement in abstract emphasizing domain characteristics; no quantification or external citation provided in the abstract.
General purpose benchmarks do not measure whether a model reasons safely and correctly about aviation specific operational knowledge.
Argument/assertion in abstract about limitations of general-purpose benchmarks; no formal empirical comparison presented in the abstract.
Unconstrained agents introduce security risks, erode codebase scalability, and make human review increasingly costly.
Authors' argumentative claim supported by the controlled experiment showing lower recall (more missed backdoors) in unconstrained condition and discussion of costs and scalability.
Computational theorising, synthetic task simulations, real LLM agent traces, and robustness analyses show that human-imitation forms often underperform when they add lossy handoffs, correlated deliberation, and verification burdens.
Empirical and simulation-based methods listed in the paper (computational theorising, synthetic task simulations, analysis of real LLM agent traces, robustness checks). The excerpt does not report sample sizes, numeric effect sizes, or statistical tests.
Applied to the human immigration study, 13.5% of reported human analyses fell in the most extreme 5% of the analysis space (m<0.05).
Application of Agentic Bootstrap / m-value estimation to the dataset analyzed by 42 human research teams; result that 13.5% of those reported analyses had m-values below 0.05.
AI agents may amplify this longstanding problem by making such exploration inexpensive and scalable.
Argument based on observed ability of AI agents to generate many plausible analysis paths and the lower cost/effort of producing such analyses using AI.
These findings suggest that the central challenge is often not flawed analyses, but selective exploration and reporting from a large space of methodologically defensible analyses.
Interpretation based on empirical results showing that divergent/opposing conclusions can be produced without clear methodological flaws and that many analyses pass review despite divergence.
Public discourse still focuses heavily on job losses while paying less attention to the opportunities that AI creates.
Author's observation/argument in the paper (qualitative commentary comparing public discourse emphasis).
The report identifies 'AI washing,' a practice in which companies mention AI as justification for what are really financially motivated layoffs.
Identification/term introduced in the paper based on examples or synthesis of corporate reporting and layoff cases (as described).
Roughly 92 million jobs might face displacement by 2030.
Projection synthesized from cited external reports (WEF/PwC/MGI/Gartner/IMF) as reported in the paper.
The review identifies attention bias as a focal mechanism, particularly salient for retail investors.
Synthesis from the literature review highlighting attention bias as a recurring mechanism with emphasis on retail investor susceptibility (no quantitative sample size reported).
In information-intensive and high-noise stock markets, investors often face information overload and rely on heuristics and selective attention, which can lead to cognitive biases and reduced decision quality.
Argument and synthesis from the paper's literature review describing investor behavior in noisy, information-intensive markets (no specific sample size reported).
Many participants used the model to rubber-stamp a prior guess and, as a result, performed worse than the model alone.
Pilot analysis comparing hybrid forecasts to model-alone forecasts and observing a subset whose forecasting error exceeded the model's error; described qualitatively in the paper.
Awards from bots with the lottery rationale can in fact reduce user activity and downstream impact.
Reported negative effect found in the Reddit field experiment: awards administered by apparent bot accounts with a lottery rationale were associated with reduced subsequent user activity and downstream impact. Sample size not reported in the abstract.
The identified gap is the price of non-credible oversight communication.
Theoretical argument and formal analysis linking the existence of the slab (gap) to non-credible communication/signaling constraints in the oversight interface; the paper interprets the gap as arising from inability to credibly communicate oversight-relevant information.
There is a slab (region) of avoidable harm: cases where the AI privately knows the proposed action is harmful and shutdown would help, yet a myopic human, trusting her prior, declines to oversee.
Analytic characterization of the gap between the team-optimal policy and the myopic human rule in the one-shot model; the paper identifies parameter-region (the 'slab') where the myopic rule fails to oversee while oversight would reduce harm.
The relational value of workplace AI companions remains underexplored.
Claim motivated by the authors' systematic interdisciplinary literature review (paper states relational value is underexplored); no numeric count of studies provided in excerpt.
This asymmetry highlights dual-use risks of AI systems designed to influence group behavior in collective action settings.
Interpretation and implication drawn from experimental findings showing that antisocial personalization can more strongly and persistently undermine cooperation.
There is an asymmetry between prosocial and antisocial persuasion: antisocial (selfish) persuasion produces larger and more persistent reductions in cooperation than prosocial persuasion produces increases.
Direct experimental comparison of prosocial versus antisocial AI persuasion treatments in the iterated Collective Risk Game showing differential magnitudes and temporal persistence of effects (reported results from N = 1,283).
The larger and more persistent negative effects of antisocial AI persuasion were particularly pronounced for personalized interventions.
Subgroup or interaction analysis in the experiment indicating that personalization (targeting by Social Value Orientation) amplified the persistence and magnitude of antisocial framing effects (reported within the N = 1,283 sample).
When AI treatments were reconfigured to promote selfish behavior through exculpatory framing, the negative effects on contributions and group success were larger and substantially more persistent.
Experimental comparison between prosocial and antisocial (exculpatory/selfish) AI treatments in the iterated Collective Risk Game showing larger and longer-lasting reductions in contributions and lower group success rates under antisocial framing (reported across N = 1,283).
In open-ended collaboration and bargaining, the same manipulation substantially degrades performance.
Experimental manipulation of agreeableness in LLMs on open-ended research collaboration and competitive bargaining tasks; authors report substantial performance degradation in these domains. Abstract lacks numeric metrics, sample sizes, and statistical significance details.
In the same repositories, agent-authored contributions concentrate repository-level friction roughly twice as much as human ones (intraclass correlation 0.30 versus 0.16); this gap holds after controlling for codebase size, age, task shape, process maturity, and merge path.
Comparison of intraclass correlations (ICC) between agent-authored and human-authored pull requests using multilevel models with controls for codebase characteristics and process variables. Dataset includes >930,000 agent-authored PRs (human sample size not specified in excerpt).
The paper reframes humans not as passive users, but as core system components whose competencies, limitations, and adaptive capacities constrain the performance envelope of optimized AI systems.
Framing/interpretive claim derived from the paper's perspective and literature synthesis (conceptual; no empirical support provided in text).
Organizational structures, bias susceptibility, retraining constraints, and interface design co-determine system stability, error propagation, and optimization ceilings.
Conceptual claim based on synthesis of literature across organizational adoption and ML lifecycle management (no empirical tests or sample sizes reported).
Human interfaces define throughput limits in areas such as prompt engineering, data-stream curation, adjudication of model outputs, and the orchestration of hybrid automation workflows including robotics, scraping, and digitization.
Theoretical assertion supported by the paper's systems-oriented analysis and literature synthesis (no empirical measurement or sample size provided).
Despite accelerating advances in AI capabilities, human capital remains the enduring and dominant system constraint.
Argument and synthesis of emerging research across human-AI interaction, ML lifecycle management, organizational adoption, and adult learning theory (conceptual synthesis; no empirical sample size reported).
Longer system responses and more information-providing turns negatively affect user satisfaction.
Statistical modeling of user satisfaction using features of multi-turn interactions (response length, number of information-providing turns) derived from the 49 participant sessions; models show negative associations reported in the paper.
Accuracy of developer+LLM assessments against expert ground truth is low.
Comparison of participant/LLM assessment outcomes to expert-annotated ground truth for the 148 NFRs; reported low accuracy in the paper.
We surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents (e.g., shortcuts).
Empirical analysis of failures and shortcuts observed when evaluating more capable agents on CORE-Bench Hard (case-study observations).
When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version; this accuracy-centric approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance (construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration).
Argument and framing in the paper supported by conceptual analysis and the CORE-Bench Hard case study (qualitative reasoning and empirical examples).
Diagnostic heuristic: if letting AI in makes the task feel effortless, it is in the wrong place.
Authors' heuristic for educators (conceptual guidance; no empirical test reported in the excerpt).
An unguarded AI helper left high-school students about 17% worse on an unaided exam than peers with no tool at all.
Described as the 'strongest causal evidence' in the paper; empirical study of high-school students measuring unaided exam performance. (Study design details and sample size not provided in the excerpt.)
Used poorly, AI replaces the cognitive work that learning requires and leaves an illusion of learning: a confident sense of mastery that collapses on the unaided task.
Authors' conceptual claim supported in the paper by reference to causal evidence (see following empirical claims); no sample size given in the excerpt.
Barriers limiting full AI adoption in auditing include resistance to change, algorithm aversion, heuristics and biases, lack of transparency, expertise and training gaps, and technological complexity.
Systematic Literature Review (SLR) of 43 studies synthesizing reported inhibitors and challenges to AI uptake in auditing from empirical and conceptual papers.
GenAI adoption carries risks including overreliance on models, misalignment between model outputs and human needs, and uneven performance across tasks and contexts.
Reported adverse effects and risks identified in the reviewed literature (task-level experiments and applied studies summarized by the paper).
This convergence has the potential to lower wages on entry-level thinking jobs.
Theoretical/empirical implication drawn from observed reduction in productivity differences; presented as a potential consequence rather than an established empirical result in the abstract.