Evidence (16496 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
Filter claims →
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 |
Adjusting for the realized audience is biased because audience is a post-treatment mediator.
Causal inference argument in paper explaining why conditioning on realized audience induces bias (audience as post-treatment mediator).
Every two-arm test conflates the creative's effect with the algorithm's targeting response.
Theoretical/causal argument presented in the paper about confounding in standard two-arm experiments when algorithmic delivery is endogenous.
Simultaneously, there is a structural shortage of qualified personnel and a gap between the education system and the needs of the economy in Uzbekistan.
Synthesis of statistical data, industry reviews, and regulatory/legal document analysis presented in the paper (no primary survey/sample size reported).
As these systems scale, the bottleneck shifts away from raw model capability toward coordination.
Analytical/argumentative claim in the paper framing a shift in primary constraint; no empirical study or quantified benchmark reported.
More persuasive narratives may have had a detrimental effect on the ability to discriminate between a correct and incorrect AI prediction.
Exploratory analyses in the paper reporting reduced discrimination between correct and incorrect AI predictions when narratives were more persuasive.
More persuasive narratives may have had a detrimental effect on decision response times.
Exploratory analyses reported in the paper indicating persuasive narratives were associated with longer decision response times.
Higher benchmark performance does not reliably show that a system can carry out knowledge work in real-world deployment settings.
Argument based on review of current knowledge-work evaluation and benchmark design literature; paper motivates with conceptual analysis and references to empirical work showing mismatch between benchmark tasks and deployed work settings.
AI systems intended to simulate companionship or emotional responsiveness raise risks such as emotional manipulation, addictive interaction patterns, and potential impact of prolonged AI interaction on users’ mental well-being, particularly for vulnerable users.
Asserted risk statement in policy recommendations; no empirical study, prevalence data, or sample provided in the text.
Current systems still struggle with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure.
Survey-identified recurring failure modes and limitations reported in literature and system descriptions; qualitative synthesis.
Current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight.
Survey of existing systems and categorization across the listed dimensions; descriptive synthesis rather than an empirical meta-analysis.
AI power demand is growing at an unprecedented rate while power grids are often ailing and struggle to keep up.
Statement in paper's motivation/background; no empirical method or sample size reported in the abstract.
The potential widening of the gender wage gap would operate through existing patterns of gender-based occupational sorting (i.e., because women are concentrated in occupations more exposed to generative AI).
Mechanistic interpretation supported by the combination of descriptive occupational sorting evidence from Swedish administrative data and results from the partial-equilibrium simulations incorporating predicted AI exposure and task complementarity.
Mechanical partial-equilibrium simulations indicate that generative AI may widen the gender wage gap.
Counterfactual simulations (mechanical partial-equilibrium) based on hypothesized deviations from the 2021 occupational and wage distribution, incorporating predicted AI exposure and task complementarity; applied to Swedish context.
Women are overrepresented in occupations predicted to be more affected by generative AI (using pre-ChatGPT occupational sorting).
Descriptive analysis of Swedish administrative data characterizing occupational gender composition before the release of ChatGPT and mapping occupations to predicted exposure to generative AI.
A reported limitation is that at this privacy level the released valuations remain noise-dominated; the system's utility derives primarily from public index routing and adaptive scheduling driven by low-sensitivity statistics.
Authors' limitation/analysis section and experimental observations.
Static temporal knowledge-graph data marketplace designs suffer three coupled failures: (i) stale hybrid index shortcuts reduce recall as edges evolve, (ii) stationary Shapley pricing misattributes value after distribution shifts, and (iii) uncoordinated agents over-consume a shared differential-privacy budget.
Authors' problem statement / conceptual diagnosis presented in the paper (no numeric sample size reported).
Monotonic baselines collapse when extrapolating beyond the training regime (e.g., predicting a 12B model up to 307B tokens) whereas the Shannon Scaling Law remains predictive.
Empirical comparison on the held-out 12B extrapolation: authors report collapse/failure of monotonic baseline scaling laws in that regime contrasted with Shannon law's successful prediction (pooled R^2 reported).
This Shannon perspective reveals a fundamental Shannon capacity for LLMs: scaling model size or data without preserving a sufficient signal-to-noise ratio (SNR) inevitably amplifies noise, inducing a transition from monotonic improvement to U-shaped performance degradation.
Theoretical argument derived from the Shannon-Hartley based formulation plus supporting empirical examples claimed in the paper showing non-monotonic (U-shaped) loss/accuracy behavior when SNR is insufficient.
Existing scaling laws for Large Language Models (LLMs), predominantly monotonic power laws, fail to explain emerging non-monotonic phenomena such as catastrophic overtraining and quantization-induced degradation, where performance deteriorates despite increased compute.
Author assertion based on literature/contextual observation and motivating examples (catastrophic overtraining, quantization-induced degradation) referenced in the paper; no specific numeric sample provided in the excerpt.
Commercial or dual-use AI models and semiconductors do not meet the security exception criteria under GATT Article XXI(b), so security interests should be interpreted restrainedly.
Legal argument and interpretive analysis in the paper contending that the GATT Article XXI(b) security exception does not encompass routine commercial or dual-use AI models and semiconductors; doctrinal legal reasoning rather than empirical measurement.
Overusing export controls can complicate dispute resolution and hinder AI progress.
Normative and legal-political argument in the paper: overuse raises legal disputes (e.g., WTO litigation) and may slow cross-border AI development and diffusion (qualitative reasoning).
Overly strict or arbitrary controls may violate WTO obligations.
Legal analysis in the paper arguing that some export controls could conflict with WTO law (GATT) depending on scope and justification; interpretive legal reasoning cited.
The long-term effectiveness of export controls is questionable.
Paper's argumentative assessment drawing on historical examples and theoretical considerations (qualitative reasoning rather than quantitative causal inference).
China responded with export curbs on critical minerals and filed a WTO complaint against the U.S. under GATT.
Factual claim citing China's counter-measures (export curbs) and legal action (WTO complaint under GATT) as described in the paper.
Even SOTA coding agents (Codex with GPT-5.4 and Claude Code with Opus 4.6) succeed on only 2/7 distributed key-value-store specifications.
Empirical evaluation reported in the paper comparing two SOTA coding agents on a suite of 7 distributed key-value-store specifications; success counted as meeting the specification.
Large retrieval models based on Small Language Models (SLMs) such as Qwen3-Embedding-4B/8B set strong upper bounds on public benchmarks but their deployment in high-throughput, latency-sensitive environments remains impractical.
Statement about model performance on public benchmarks (upper bounds) and practical deployment constraints (throughput and latency), asserted by authors; no numerical deployment analysis provided in excerpt.
Rule debt is a governance burden that accrues when organizational decision rules migrate from formal information systems into ungoverned agentic execution environments.
Conceptual construct introduced and defined in the paper; supported by illustrative examples, no empirical measurement reported.
AI-enabled capabilities whose outputs require evidence, review, signoff, or assignable responsibility may retain integrated accountability boundaries even when their technical interfaces become modular.
Theoretical claim supported by conceptual analysis and domain illustrations; no empirical sample or formal measurement reported.
A complementary Oaxaca–Blinder decomposition shows that shifts in occupational composition account for about 90% of the exposure change attributable to observable job characteristics.
Oaxaca–Blinder decomposition reported in the paper attributing ~90% of exposure change (among the portion explained by observable job characteristics) to occupational composition shifts.
Within-job redesign accounts for 39.5% of the aggregate decline in generative-AI exposure and becomes increasingly important over time.
Same decomposition as above reported in the paper (result: within-job redesign = 39.5% of aggregate decline; authors note its increasing importance).
Hiring reallocation explains the largest share of the aggregate decline in generative-AI exposure, accounting for 52% on average.
Decomposition of changes in aggregate exposure into two margins (reallocation across jobs and within-job redesign) reported in the paper (result: hiring reallocation = 52% of aggregate decline).
We argue that regions are unlikely to maximize all three [Progress, Sustainability, Equity] simultaneously under current technological, institutional, and resource conditions.
Argument based on synthesis of prior literature on limits of AI development and illustrative evidence (regional cases and stakeholder comment analysis); explicitly stated in the abstract.
The rapid expansion of artificial intelligence infrastructure, including data centers and the energy, land, water, and labor systems that support them, presents regional policymakers with trade-offs that are poorly captured by the prevailing "innovation versus regulation" frame.
Conceptual argument drawing on prior literature and illustrative regional examples presented in the paper; stated explicitly in the abstract.
Two of the top three leaderboard models (gpt-5 and claude) are noticeably more locally volatile than the third (gemini-3.1-pro), despite being close in overall strength.
Comparison of jaggedness/local volatility measures and overall scores from the tournament (top-three leaderboard).
Existing strategic-reasoning benchmarks evaluate models on fixed canonical games and may saturate as the frontier improves and fail to generalize to varied real-world strategic environments.
Conceptual critique stated in the paper's motivation/background; no empirical test reported in abstract.
Evaluating state-of-the-art kernel agents on FastKernels, the strongest agent achieves only 0.94× aggregate speedup over production baselines, with weaker agents at 0.78× and 0.53×.
Empirical evaluation of multiple state-of-the-art kernel-generation agents on the FastKernels benchmark; aggregate speedup factors reported in abstract. The number of benchmark tasks is likely the FastKernels task set (46), though the abstract does not explicitly state the evaluation sample size for this measurement.
Existing benchmarks are poorly aligned with production inference frameworks: they evaluate kernels on a single GPU with synthetic inputs, ignore the surrounding compilation stack, and reward replicating known optimizations rather than discovering new ones.
Stated as motivating observation in the paper (conceptual/empirical critique of existing benchmark design and incentives). No numerical sample size given in the abstract.
Other changes are more nuanced and put the typical career growth opportunities, like receiving feedback from professional networks and promoting leadership and mentorship, at risk.
Qualitative reports from interview participants (n=24) expressing concerns that AI-driven changes may reduce feedback, leadership development, and mentoring opportunities.
Integrations of AI that neglect human factors are associated with increased anxiety, burnout, and disengagement among users.
Aggregate findings from the systematic review reporting associations in the literature between non-human-centered AI integration and negative psychological/work outcomes.
Notable challenges to AI implementation include concerns about algorithmic bias, privacy, transparency, job displacement, organizational culture, and issues related to ethical and legal oversight.
Synthesis of reported challenges across the 29 empirical studies included in the scoping review.
Fragmented, uncoordinated approaches in the absence of national strategy constitute a structural barrier to technological development in Georgia.
Method: logical inference and country assessment presented in the paper documenting fragmentation across policy and institutional actors; qualitative evidence rather than quantitative causal estimation.
In Georgia, the total absence of a national AI strategy and legal definition produces fragmented approaches, creating a structural barrier to technological development.
Method: country-level assessment of policy and legal framework for AI in Georgia; descriptive analysis identifying lack of a national strategy and definition. (No sample size reported.)
This transition proceeds without tools to forecast how individual employees will respond psychologically and behaviorally.
Asserted by the authors as a gap/need; no empirical inventory or systematic review presented in the excerpt to substantiate completeness of tool absence.
Workforce transformations are difficult to forecast and costly to mismanage.
Stated as a general assertion in the paper's introduction; no empirical data, sample, or formal analysis reported in the excerpt.
Student-designed tasks reveal hidden failures in current deep research systems: fluent, source-backed answers can still miss the right query, source, term, or evidence standard.
Qualitative analysis of failure modes from student-designed tasks and system evaluations reported in the paper (examples and discussion of how answers can be fluent and sourced yet incorrect on key criteria).
Evaluation on QuestBench shows that student-designed tasks reveal hidden failures in current deep research systems: across thirteen evaluated systems, the mean question-level pass rate is only 16.85%.
Empirical evaluation reported in the paper: 13 systems evaluated on QuestBench; aggregated mean question-level pass rate reported as 16.85%.
Zero-shot evaluation shows the best positive-query mask success rate at IoU@0.75 remains below 0.17.
Empirical evaluation reported in the paper: zero-shot tests across 26 model configurations with reported mask success rate at IoU@0.75.
Zero-shot evaluation of 26 model configurations spanning closed-source MLLMs, open-source VLMs, and specialized grounding systems reveals persistent gaps: the best multi-target Set-F1 reaches only 0.35.
Empirical evaluation reported in the paper: zero-shot tests across 26 model configurations with reported Set-F1 metric.
Reliable evaluation of agricultural visual grounding remains challenging because agricultural targets are often small, repetitive, occluded, or irregularly shaped, and instructions may refer to one, many, or no objects in an image.
Problem characterization / motivation described in the paper (qualitative reasoning about dataset and task properties).
Technical bottlenecks (cross-border data compliance, algorithm interpretability) and ethical challenges (algorithmic bias, privacy infringement, cultural conflicts) are intertwined impediments to intelligent international marketing.
Synthesis of challenges identified across the reviewed literature (systematic review and content analysis, 2010–2025) as reported in the paper.