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
Early evidence indicates AI is reducing the productivity difference between beginner and expert employees.
Reported 'early evidence' from the paper's empirical analysis (difference-in-differences on freelance platforms) indicating convergence in productivity between novices and experts; no numeric effect estimates given in the abstract.
A strategic AI sender may withhold evidence or garble information in order to steer the human's decision.
Theoretical reasoning and examples within the Bayesian persuasion framework showing that sender-optimal signaling need not fully reveal the state and can be manipulative; supported by model analysis rather than empirical data.
These results demonstrate how people's decision-making processes can be insufficient for overseeing AI in high-stakes domains.
Synthesis/interpretation of experimental findings (longer viewing when no AI, small increases in selection probability with more time for non-recommended candidates, IAT effects) to argue that human decision processes may not adequately supervise biased AI in high-stakes settings. This is an interpretive/concluding claim based on the experiment; not a direct empirical measure. Sample size not stated in the excerpt.
Pooled across five AI coding agents, pull requests (PRs) with a human Co-Authored-By trailer merge less often than purely-autonomous ones (53.8% vs. 79.8%).
Aggregate analysis of PR merge rates across five AI coding agents in the AIDev dataset; pooled sample of PRs (33,596 PRs referenced elsewhere in the paragraph).
The paper identifies two distinct gaps that have widened as GPTs exposure scores traveled from their time and place of production: (1) a structural gap between what static exposure scores measure and what policy questions require, and (2) a coordination gap between researchers and policymakers.
Explicit framing and thesis presented in the paper summarizing the central arguments.
Policy-relevant work that asks who is harmed or benefits, how, and when continues to reference static GPTs exposure scores without engaging with methodological updates needed to answer these questions more reliably.
Critical literature review and observed citation practices reported by the authors; claim based on review of how policy analyses cite/ use the scores.
These temporal, geographic, and ontological limitations compound when exposure scores are used in policy-facing analyses.
Conceptual argument and case-study approach in the paper showing how limitations interact and worsen policy analysis outcomes.
The GPTs exposure scores have temporal, geographic, and ontological limitations that do not always travel with the scores as they are reused.
Authors' methodological critique discussing the limits named by Eloundou et al. (2023) and how those limits are often ignored when scores are repurposed.
Under individual selection, self-interested prompts dominate, causing populations to collapse into collective defection.
Simulation experiments with individual-level selection/transmission showing emergence and dominance of self-interested prompts and subsequent decline into collective defection.
As frontier training shifts toward individual rewards for verifiable tasks (e.g., mathematics and coding), this outcome-based focus may further undermine cooperation in multi-agent settings.
Argumentative/prognostic claim in the paper's motivation; not an empirical result from the study but framed as a risk informed by the literature and authors' reasoning.
Current approaches to instill prosociality in LLM agents often rely on humans specifying desired behaviors at the individual level, which does not guarantee cooperation within LLM populations.
Background statement in paper; conceptual critique of human-specified, individual-level reward/behavior specification as commonly used in LLM alignment and fine-tuning literature (no new empirical test reported in this study).
Existing frameworks address AI-assisted development maturity or the productivity-reliability tension but offer no mechanism for calibrating human oversight intensity to regulatory impact.
Comparative framework analysis and literature review reported in the paper (claims about gaps in existing frameworks).
The adoption of agentic AI coding systems -- where autonomous agents generate, review, test, and deploy code with minimal human intervention -- creates a governance challenge in regulated industries.
Argumentation in the paper framing the problem; conceptual analysis of agentic AI capabilities and regulatory constraints (literature/contextual reasoning rather than empirical data).
Neither the task design nor the retrieval approach of Finance Agent v2 addresses the distinct challenges of IPO due diligence.
Author argument comparing periodic reporting tasks to IPO due-diligence requirements, noting Finance Agent v2's task and retrieval design do not address IPO-specific complexities.
The Finance Agent v2 agentic harness relies on naive, unenriched chunk retrieval.
Author statement describing the retrieval approach used by Finance Agent v2 as naive chunk retrieval.
The essay introduces the concept of a 'vouching gap' to describe a growing divide between students who graduate with credible advocates willing to stake their reputations on their behalf and those who do not.
Conceptual contribution defined in the essay and motivated by social capital theory and mentoring research; no empirical quantification or sample provided.
Automation of student work and candidate screening will widen existing inequalities between students.
Theoretical claim in the essay linking AI-driven automation to differential outcomes across students, motivated by social capital and mentoring literature; no empirical data or sample reported.
This automation threatens to hollow out the value of a university degree.
Argument presented in the essay, grounded in social capital theory and mentoring research; no empirical test or sample size reported.
Manual preparation of engineering designs for thousands of wells constitutes an enormous administrative burden and is prone to inconsistencies.
Introductory/background statement in the paper describing the pre-existing manual workflow burden; no numerical study reported for this specific statement.
The demand premium enjoyed by workers with strong human capital declines in more AI-exposed categories.
Heterogeneity analysis within the Upwork dataset: workers characterized by stronger human-capital signals (via profile embeddings) show a reduced demand premium in job categories more exposed to AI following ChatGPT; identified using difference-in-differences around ChatGPT release. (Sample size not reported in abstract.)
In more AI-exposed job categories, the importance of human capital information in predicting labor demand declines.
Empirical analysis of Upwork platform data using high-dimensional text embeddings to represent worker profiles; the paper computes the predictive importance of human-capital-related profile information and uses a difference-in-differences design around the release of ChatGPT to estimate changes by AI exposure of job categories. (Sample size not reported in abstract.)
Adding relevant collaborators can lower performance when teams lack structure to coordinate their contributions.
Empirical comparisons across experimental sessions in the Collaborative Gym / DiscoveryBench setup; result reported across the study (1,482 sessions).
A wide range of empirical evidence shows that humans avoid complexity, delegate judgement, and prefer simplified social worlds.
Asserted as empirical background; paper references a broad empirical literature but does not report primary data, sample sizes, or specific studies in the provided text.
Most organizations (59%) approach AI implementation through a technology-first lens, layering intelligent systems onto legacy processes rather than intentionally redesigning how humans and machines collaborate.
Reported descriptive statistic from Deloitte's 2026 Global Human Capital Trends survey of over 3,000 business leaders across 15 countries (paper cites 59% figure).
Only 14% of organizational leaders report proficiency in designing effective human-machine interactions.
Reported descriptive statistic from the same Deloitte 2026 Global Human Capital Trends survey of over 3,000 business leaders across 15 countries.
Current machine learning models commonly require large and well-annotated datasets, and the annotation process often becomes a bottleneck with increased complexity leading to higher chances of human errors.
Background statement in the paper summarizing common knowledge and prior literature about dataset requirements and annotation challenges.
At the macro level, values-driven withdrawal from AI use has the potential to narrow the diversity of visible applications, amplifying risk-focused narratives and reinforcing perceptions of harm in public discourse.
Theoretical extension of the guarded engagement loop to societal/public discourse dynamics; based on synthesis of social amplification of risk literature rather than empirical measurement in the abstract.
These constrained (guarded) interactions can lower output quality and increase the likelihood of visible errors, which may further erode trust and reinforce cautious engagement.
Theoretical causal chain posited by the authors within their conceptual framework; supported by literature-based argumentation rather than reported empirical results in the abstract.
At the micro level, elevated risk salience related to privacy, safety, or ethical concerns may lead users to adopt guarded interaction strategies characterized by reduced contextual disclosure and limited iteration.
Theoretical proposition within the paper's guarded engagement loop framework, drawing on prior research in privacy calculus and algorithm aversion; no specific empirical data reported in the abstract.
Generative AI adoption is often framed primarily as a question of learning technical skills, and this perspective overlooks a defining feature of large language models (LLMs): their output quality depends heavily on how users engage with them.
Conceptual argument presented in the paper's introduction/abstract; literature synthesis framing adoption debates (no empirical sample or experimental method reported in the abstract).
Expertise moderated the effect of LLM guidance: novices exhibited passive AI reliance.
Stratified analyses by participant expertise level using behavioral and eye-tracking measures indicating novices shifted attention to the AI/chat and exhibited more passive acceptance of guidance.
AI augmentation breaks the accounting link between labor time and productive contribution, yet firms continue to evaluate talent through time-based overhead bundles.
Theoretical argument and conceptual framing presented in the paper (no empirical sample reported for this specific proposition).
Financial LLMs face regulatory compliance violations, fraud facilitation, and systemic trust erosion that require targeted evaluation.
Paper's risk analysis listing finance-specific threats (regulatory compliance violations, facilitation of fraud, systemic trust erosion). This is a conceptual/risk framing rather than reported empirical incidence rates in the provided summary.
Existing safety benchmarks target general adversarial scenarios but miss finance-specific risks.
Authors' comparative assertion in paper (conceptual analysis arguing gap between general LLM safety benchmarks and finance-specific threats). No numeric evaluation reported in the provided summary.
Investment is being directed toward AI deployment when achieving productivity gains requires prior development of convergence capacity (C), leading to a misallocation of investment.
Theoretical reasoning within the paper: conceptual argument that deployment-focused spending misses prerequisite cognitive capacity (C).
Prevailing production-function frameworks encounter a structural boundary because they treat AI as a separable factor of production without modeling the cognitive mediation through which AI generates productive value.
Theoretical / conceptual argument presented in the paper (derivation and critique of existing production-function approaches).
Massive AI investment has failed to generate commensurate productivity gains (the "AI productivity paradox").
Stated as the motivating empirical paradox in the paper; presented as an observed phenomenon motivating the theoretical argument (no specific dataset or numeric evidence provided in the abstract).
The translation of AI's potential into operational capability within government audit contexts requires navigating complex technical, institutional, legal, and ethical challenges that differ substantially from private sector environments.
Paper's conceptual analysis and comparative argument (paper contrasts government audit contexts with private sector origins of many AI tools); no quantitative empirical evidence or sample size reported.
There are barriers and challenges that the labor force faces in meeting new skill requirements.
Review conclusion noting barriers and challenges reported in the empirical literature (types of barriers not enumerated in the excerpt; no measures or prevalence reported).
The root causes of these problems include the disruption of labor relations boundaries by the transformation of the means of production, the exclusion of implicit data labor from distribution rules, the concentration of capital driven by high industry barriers, and social structural constraints on technological dissemination.
Synthesis and causal argumentation grounded in Marx's theory of reproduction; conceptual reasoning rather than empirical testing.
In the consumption phase, high costs lead to service stratification, making it difficult for technological dividends to benefit the general public.
Theoretical/qualitative argument about cost barriers and unequal access to AI-enabled services; no empirical evidence or sample sizes reported.
In the exchange phase, high barriers to entry for technology and capital foster market monopolies.
Analytical claim based on structural characteristics of AI/embodied intelligence industries; no empirical sample or quantitative measures provided in the paper.
In the distribution phase, behavioral data unconsciously generated by workers drives algorithmic iteration yet remains excluded from the distribution system, resulting in hidden data exploitation.
Theoretical argument that worker-generated behavioral data fuels algorithmic development but is not accounted for in value distribution; no empirical data or sample reported.
In the production stage, workers are alienated into becoming data producers.
Conceptual claim based on Marxian analysis of labor and data extraction; no empirical sample or quantitative evidence presented.
In the production stage, workers are disciplined by algorithms.
Theoretical/qualitative argument in the paper describing algorithmic management and control; no empirical measures or sample provided.
In the production stage, workers lose decision-making power.
Theoretical analysis of production relations using Marxist reproduction framework; qualitative claim without reported empirical data.
The canonical manifestation of this failure pattern is called 'Phantom Legislation' (internally consistent symbolic outputs disconnected from real business semantics).
Terminology and descriptive example provided by the authors based on their analysis of observed failure cases in the Bang-v3 project.
Test file counts substantially overestimate verification strength.
Conclusion drawn from the high prevalence (80.2%) of test patches with weak/no oracle signals compared to mere presence-of-test-file counts.
Raw merge rates are lower for strong-oracle PRs.
Unadjusted (raw) comparisons of merge rates between PRs classified by oracle strength in the study dataset.
Applied at scale, 80.2% of test patches contain weak or no explicit oracle signals.
Automated/syntactic classification of oracle-signal categories applied to the full dataset of test-file patches (as described in methods).