Evidence (8807 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
Filtered →
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 |
Productivity
Remove filter
Baseline (non-structured) interactions had 16 of 50 accepted on first pass.
Reported counts in the paper for the baseline group (16 accepted of 50 baseline interactions).
In an observational study of documented interactions across four AI tools (Claude, ChatGPT, Cowork, Codex), incomplete context was associated with 72% of iteration cycles.
Observational study reported in the paper covering interactions across four AI tools; the paper reports the 72% figure.
Job insecurity emerges as a critical mediating factor influencing employee attitudes and behavioural responses to generative AI, including upskilling intentions and resistance to technological change.
Review-level synthesis identifying job insecurity reported in included studies as mediating relationships between AI adoption and employee attitudes/behaviours (e.g., upskilling, resistance).
Employees express concerns about role displacement (job loss or role changes) associated with generative AI adoption.
Reported across multiple studies included in the review; the review summarises these concerns as part of mixed employee perceptions.
These positive perceptions coexist with employee concerns about skill obsolescence related to generative AI.
Synthesis of studies included in the review documenting worker concerns about skills becoming obsolete due to AI-driven changes.
Income inequality, measured by the Gini index, rises moderately in every scenario we examine due to the polarising effect of job losses and wage and capital income increases on the income distribution.
Calculation of Gini index across multiple simulated scenarios using the SWITCH-linked distributional analysis; reported in the report.
The largest average losses are experienced by middle and higher income households, for whom job displacement outweighs any wage or capital income gains. Lower income households also lose, but by much less.
Distributional results from microsimulation (SWITCH) applying scenarioled job displacement, wage and capital effects across income groups; reported in the report.
When these effects are combined, we find an average decline in household disposable income as a result of AI adoption.
Combined scenario simulations incorporating job displacement, wage effects and capital income effects linked to the Irish tax-benefit system using SWITCH; result reported in the report's main findings.
These wage gains are not large enough to counterbalance the average fall in income due to job displacement.
Combined simulation results (displacement + wage effects) using scenario assumptions and microsimulation (SWITCH), reported in the report's distributional analysis.
Those most likely to experience this disruption are found in higher income households, where the share of workers transitioning into unemployment is substantially larger than in lower income families.
Microsimulation (SWITCH) linking simulated job displacement scenarios to household income groups; results reported in the report.
In our central scenario — drawn from credible international estimates — around 7 per cent of current jobs could be displaced in the short–medium run.
Scenario simulation based on international estimates of AI exposure/adoption; central scenario reported in the report (linked to SWITCH microsimulation for distributional analysis).
AI tends to place higher earning and highly educated workers at greater risk of disruption, because the occupations most exposed to AI are predominantly in these groups.
Synthesis of international research on occupational exposure to AI and the report's analysis linking exposure to worker characteristics (education and earnings); presented as descriptive finding in the report.
Result 2: When managers are short-termist or worker skill has external value, the decision-maker's optimal policy can produce the augmentation trap, leaving the worker worse off than if AI had never been adopted.
Analytical result from the dynamic model comparing planner/objective variations (short-termist manager or externalities) and showing an outcome labeled the 'augmentation trap'.
Result 1: Even a decision-maker who fully anticipates skill erosion rationally adopts AI when front-loaded productivity gains outweigh long-run skill costs, producing steady-state loss: the worker ends up less productive than before adoption.
Analytical result from the dynamic model showing optimal adoption choice can lead to a steady-state where worker productivity is lower than pre-adoption (model-based comparative statics).
Experimental evidence shows that sustained use of AI tools can erode the expertise on which productivity gains depend (deskilling).
Statement in paper referencing experimental studies (no specific study, method, or sample size reported in the excerpt).
Claude Sonnet 4.6 achieves only 33.3% (completion rate) on ClawBench.
Paper gives a concrete example performance result for Claude Sonnet 4.6 (reported completion percentage on the benchmark).
The authors evaluated 7 frontier models on ClawBench and found that both proprietary and open-source models can complete only a small portion of these tasks.
Paper reports evaluations of 7 models on the ClawBench tasks (empirical evaluation across the benchmark).
Aggressive compression increased total session cost by 67% despite reducing input tokens by 17%, because it shifted interpretive burden to the model's reasoning phase.
Result reported from the controlled experiment comparing log-format conditions; four conditions described but specific number of sessions/replications not provided in the abstract.
Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall.
Paper reports an evaluation across 17 models and states the maximum overall score observed was below 66%.
Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval.
Statement in paper introduction contrasting prior benchmarks' focus on explicit recall with a claimed gap in evaluating implicit (non-declarative) memory; no systematic literature review or quantitative survey reported in the excerpt.
OpenAI o3 achieves only 17% of optimal collective performance.
Experimental measurement of collective performance for OpenAI o3 in the paper's multi-agent setup (value reported in abstract; no sample size provided there).
The study observed errors and limitations in both phases (test generation and refactoring), and manual intervention was necessary at times.
Case study observations reported in the paper describing observed model errors/limitations and instances requiring manual developer intervention.
Current AI coding assistants, such as GitHub Copilot and Amazon CodeWhisperer, emphasize developer speed and convenience, with energy impact not yet a primary focus.
Stated as an observation in the paper; no specific empirical comparison or quantification provided in this excerpt.
Frontend code, replicated across millions of page views, consumes significant energy and contributes directly to digital emissions.
Asserted in paper's introduction; no specific empirical data or sample reported in this excerpt.
We posit that persistence is reduced because AI conditions people to expect immediate answers, denying them the experience of working through challenges on their own.
Authors' proposed psychological mechanism / explanation inferred from observed behavior; presented as a hypothesis rather than directly proven causal mediator.
These negative effects (reduced persistence and impaired unassisted performance) emerge after only brief interactions with AI (approximately 10 minutes).
Experimental manipulation / exposure in RCTs where participants interacted with AI for about 10 minutes and subsequent outcomes were measured.
People are more likely to give up after interacting with AI (increased likelihood of quitting tasks unassisted).
Randomized controlled trials (N = 1,222) measuring rates of task abandonment/giving-up after AI interaction vs. control.
AI assistance impairs unassisted performance: although AI improves short-term performance, people perform significantly worse without AI after interacting with it.
Randomized controlled trials (N = 1,222) comparing performance with and without AI assistance across tasks; causal inference from randomized assignment.
Through a series of randomized controlled trials on human-AI interactions (N = 1,222), we provide causal evidence that AI assistance reduces persistence.
Randomized controlled trials (RCTs) on human-AI interactions with total sample size N = 1,222; persistence measured after AI interaction across tasks.
AI-assisted evaluation reduces variance in research quality.
SEM and regression analyses on OECD panel data report a decrease in variance of research quality measures associated with higher AIRC.
Current research has largely focused on short-horizon tasks over a limited set of software with limited economic value (e.g., basic e-commerce and OS-configuration tasks).
Narrative literature/field observation reported in paper introduction (no numeric study reported in excerpt).
There is a fundamental gap in current agent capabilities: functional correctness alone is insufficient for design-aware issue resolution, motivating design-aware evaluation beyond functional correctness.
Synthesis of experimental findings: low design-satisfaction despite functional correctness, prevalence of design violations, and only partial improvement from guidance support the conclusion.
Design violations are widespread in agent-produced patches.
Empirical results from experiments on the benchmark showing many patches violate validated design constraints; backed by counts/percentages in evaluation (as summarized in abstract).
Test-based correctness substantially overestimates patch quality: fewer than half of resolved issues are fully design-satisfying.
Experimental evaluation with state-of-the-art LLM-based agents on the benchmark (reported in paper). Sample implicit: benchmark issues (495) used to evaluate agents; comparison between test pass rates and design-satisfaction measured by verifier.
Despite growing investment in data analytics, the decision-making and coordination layers of these workflows remain predominantly manual, reactive, and fragmented across outlets, distribution centers, and supplier networks.
Stated as an observation in the paper (abstract); no quantitative evidence, metrics, or comparative analysis provided in the excerpt.
Retail supply chain operations in supermarket chains involve continuous, high-volume manual workflows spanning demand forecasting, procurement, supplier coordination, and inventory replenishment.
Descriptive claim stated in the paper's introduction/abstract; no empirical data, sample, or methods reported to substantiate this characterization within the text provided.
The two margins interact through a self-undermining feedback that can generate low-archive traps (multiple equilibria with low accumulated public archive).
Dynamic equilibrium analysis in the theoretical model showing interacting feedbacks and possible trap equilibria (model-derived result).
Resolution margin: the probability that posted queries are resolved declines because AI raises contributors' outside options, thinning the contributor pool and creating congestion on the platform.
Mechanism and comparative-static implication produced by the paper's theoretical model; no empirical sample provided in the excerpt.
Flow margin: the posted volume of knowledge-enhancing queries declines as AI resolves more problems privately before they reach the platform.
Mechanism derived in the theoretical model; stated as the flow-margin channel (no empirical quantification in the provided text).
AI reduces archive creation through two distinct margins: a flow margin and a resolution margin.
Analytical decomposition derived within the paper's theoretical model (mechanism claimed by the model).
Generative AI resolves user problems without leaving a public trace, so fewer discussions and solutions reach public platforms.
Stated as an empirical motivation in the paper; no empirical sample or quantified measurement reported in the provided text.
Green AI research has largely measured the footprint of models rather than the downstream workflows in which GenAI is a tool.
Literature review / mapping of recent Green AI literature reported in the paper; descriptive claim about the focus of the field (no sample size or numerical counts reported in the abstract).
Existing benchmarks differ from real usage in programming language distribution, prompt style and codebase structure.
Paper asserts mismatch between existing benchmarks and production usage as motivation for producing a production-derived benchmark (stated differences: language distribution, prompt style, codebase structure).
Replacing deterministic components with probabilistic workflows changes the failure mode: LLM pipelines may generate plausible but incorrect outputs that pass superficial checks and propagate into irreversible actions such as DOI minting and public release.
Conceptual argument supported by the paper's incident descriptions (e.g., a detected coordinate transformation error); the statement is presented as a general risk rationale.
Occupations whose AI-exposed steps are more dispersed across the production workflow (higher fragmentation) exhibit a substantially lower share of their steps actually executed by AI, conditional on AI exposure share.
Empirical regression analysis controlling for share of AI-exposed steps; uses dataset linking O*NET tasks, human AI exposure assessments, Anthropic Economic Index execution outcomes, and GPT-generated workflow orderings (details in Sections 5.1 and 7).
Treated firms' demand for external capital investment falls by just over $220,000 relative to the control group.
RCT with 515 firms; reported dollar-change in external investment demand between treated and control firms.
Despite faster growth, treated firms do not scale inputs proportionally: their demand for external capital investment falls by 39.5% relative to the control group.
RCT with 515 firms; firms reported external capital demand/investment requests; comparison of investment demand between treatment and control groups.
For the private business sector, if the set of automated tasks were frozen in 1950, 87% of TFP growth between 1950 and 2023 would have been eliminated.
Counterfactual growth-accounting exercise that freezes the set of automated tasks at 1950 while allowing capital, labor, and other productivity growth to follow historical rates (simulation based on calibrated accounting).
The sum of "other" TFP growth and average labor productivity growth (ˆZt + ˆψℓt) is small — for example equal to -0.1% per year for the private business sector since 1950.
Growth-accounting decomposition for the private business sector since 1950 using BEA/BLS data in the task-based framework.
Under the rapid scenario, economists forecast the share of wealth held by the wealthiest 10% of households rising to 80.0% by 2050.
Conditional forecasts in Key Findings for the economist respondent group under the rapid AI scenario (2050 horizon).