Evidence (8974 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
10085 claims
Filter claims →
Productivity
8974 claims
Filtered →
Governance
8062 claims
Filter claims →
Human-AI Collaboration
7749 claims
Filter claims →
Org Design
5057 claims
Filter claims →
Innovation
4896 claims
Filter claims →
Labor Markets
4088 claims
Filter claims →
Skills & Training
3372 claims
Filter claims →
Inequality
2377 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 | 882 | 244 | 117 | 1097 | 2424 |
| Governance & Regulation | 1010 | 469 | 229 | 135 | 1875 |
| Organizational Efficiency | 977 | 235 | 149 | 90 | 1462 |
| Technology Adoption Rate | 781 | 299 | 143 | 128 | 1362 |
| Research Productivity | 506 | 155 | 74 | 363 | 1110 |
| Output Quality | 555 | 219 | 71 | 70 | 915 |
| Decision Quality | 395 | 200 | 95 | 54 | 751 |
| Firm Productivity | 523 | 67 | 101 | 27 | 724 |
| AI Safety & Ethics | 262 | 309 | 75 | 36 | 688 |
| Market Structure | 195 | 201 | 135 | 30 | 566 |
| Task Allocation | 248 | 77 | 96 | 38 | 464 |
| Innovation Output | 300 | 34 | 55 | 20 | 411 |
| Skill Acquisition | 207 | 75 | 65 | 21 | 368 |
| Employment Level | 138 | 67 | 119 | 24 | 350 |
| Fiscal & Macroeconomic | 156 | 80 | 53 | 33 | 329 |
| Task Completion Time | 211 | 38 | 13 | 16 | 280 |
| Firm Revenue | 183 | 52 | 29 | 5 | 270 |
| Consumer Welfare | 131 | 77 | 48 | 13 | 269 |
| Inequality Measures | 50 | 141 | 54 | 9 | 254 |
| Worker Satisfaction | 104 | 85 | 25 | 13 | 227 |
| Error Rate | 87 | 112 | 11 | 5 | 215 |
| Automation Exposure | 69 | 69 | 37 | 20 | 198 |
| Wages & Compensation | 102 | 49 | 31 | 11 | 193 |
| Team Performance | 115 | 30 | 30 | 11 | 187 |
| Regulatory Compliance | 88 | 74 | 17 | 7 | 186 |
| Training Effectiveness | 109 | 22 | 14 | 21 | 168 |
| Developer Productivity | 116 | 21 | 15 | 8 | 161 |
| Job Displacement | 12 | 92 | 26 | 1 | 131 |
| Hiring & Recruitment | 57 | 12 | 9 | 5 | 83 |
| Skill Obsolescence | 6 | 59 | 10 | 2 | 77 |
| Social Protection | 43 | 17 | 8 | 2 | 70 |
| Creative Output | 35 | 21 | 9 | 4 | 70 |
| Labor Share of Income | 18 | 23 | 17 | 1 | 59 |
| Worker Turnover | 15 | 16 | — | 4 | 35 |
| Industry | — | — | — | 1 | 1 |
Productivity
Remove filter
Evaluation demonstrates that Flowr significantly reduces manual coordination overhead.
Empirical claim reported in the paper's evaluation section; the excerpt notes an evaluation and collaboration with a large supermarket chain but provides no sample size figures or quantitative effect sizes.
Central to the framework is a human-in-the-loop orchestration model in which supply chain managers supervise and intervene across workflow stages via a Model Context Protocol (MCP)-enabled interface, preserving accountability and organizational control.
Design/organizational claim describing human-in-the-loop orchestration and MCP interface; asserted in the paper without empirical measures of accountability or control in the excerpt.
To ensure task accuracy and adherence to responsible AI principles, the framework employs a consortium of fine-tuned, domain-specialized large language models coordinated by a central reasoning LLM.
Technical/design claim in the paper describing model architecture and approach; no evaluation metrics or tests of accuracy/responsibility provided in the excerpt.
Flowr systematically decomposes manual supply chain operations into specialized AI agents, each responsible for a clearly defined cognitive role, enabling automation of processes previously dependent on continuous human coordination.
Architectural claim — asserted mechanism of the framework in the paper; presented as part of the framework design, no quantitative evaluation details in the excerpt.
This paper introduces Flowr, a novel agentic AI framework for automating end-to-end retail supply chain workflows in large-scale supermarket operations.
Design and system-proposal claim in the paper; supported by framework description rather than empirical testing in the provided text.
Generative AI helps users solve problems more efficiently.
Motivating empirical observation stated in the paper (no sample or empirical analysis reported in the provided text); assumption used to motivate the theoretical model.
Prompts can be treated as decision policies that allocate discretion between researcher and system, governing what is executed and when iteration stops.
Methodological framing advanced by the authors describing prompts as decision policies; conceptual claim based on the paper's analytic framework rather than empirical measurement.
Operational constraints and decision rule prompts deliver large and stable footprint reductions while preserving decision equivalent topic outputs.
Experimental comparisons of prompt strategies in the benchmarked workflow showing reductions in runtime/CO2e and evaluated topic outputs' decision-equivalence (asserted in abstract; no numeric reductions or sample sizes provided).
We benchmark a modern economic survey workflow, an LDA-based literature mapping implemented with GenAI assisted coding and executed in a fixed cloud notebook, measuring runtime and estimated CO2e with CodeCarbon.
Experimental benchmark described in the paper: single implemented workflow (LDA-based literature mapping) executed in a fixed cloud notebook with runtime and CO2e measured using CodeCarbon (methodological claim).
Training footprint is the largest cluster in the mapped Green AI literature.
Result from the paper's literature mapping / clustering (statement in abstract; no numeric cluster sizes given).
We map the recent Green AI literature into seven themes: training footprint is the largest cluster, while inference efficiency and system level optimisation are growing rapidly, alongside measurement protocols, green algorithms, governance, and security and efficiency trade-offs.
Bibliometric / thematic mapping of recent Green AI literature described in the paper (method: literature mapping; exact number of papers or mapping procedure not specified in abstract).
We share our methodology and lessons learned to enable other organizations to construct similar production-derived benchmarks.
Paper states intention and contribution: releasing methodology and lessons to allow replication by other organizations.
We detail data collection and curation practices including LLM-based task classification, test relevance validation, and multi-run stability checks to address challenges in constructing reliable evaluation signals from monorepo environments.
Methodological description in paper listing specific practices (LLM-based classification, test relevance validation, multi-run stability checks) aimed at producing reliable evaluation signals in monorepos.
Models making greater use of work validation tools, such as executing tests and invoking static analysis, achieve higher solve rates.
Reported relationship from paper's analysis correlating models' use of verification tools (test execution, static analysis) with higher solve rates across evaluated models.
Systematic analysis of four foundation models yields solve rates from 53.2% to 72.2%.
Empirical evaluation reported in paper: four foundation models were evaluated on the ProdCodeBench benchmark producing reported solve-rate range.
Each curated sample consists of a verbatim prompt, a committed code change and fail-to-pass tests spanning seven programming languages.
Descriptive dataset claim in paper specifying components of each sample and that samples cover seven programming languages.
We present ProdCodeBench, a benchmark built from real sessions with a production AI coding assistant.
Paper describes methodology and introduces ProdCodeBench explicitly as constructed from real production assistant sessions.
Benchmarks that reflect production workloads are better for evaluating AI coding agents in industrial settings.
Argument presented in paper motivating creation of production-derived benchmark; no specific empirical comparison to alternative benchmarks reported in the abstract.
A representative incident (ISS-004) demonstrated boundary-based containment with 10-minute detection latency, zero user exposure, and 80-minute resolution.
Incident ISS-004 report in the paper giving specific timings for detection latency (10 minutes), user exposure (zero), and resolution (80 minutes).
The multi-agent approach improved reliability: audited handoffs detected and blocked a coordinate transformation error affecting all 2,452 stations before publication.
Incident detection reported in the SF2Bench deployment where audited handoffs prevented publication of a coordinate transformation error that would have affected all 2,452 stations.
The multi-agent approach improved efficiency — the SF2Bench deployment was completed by a single operator in two days with repeated artifact reuse across deployments.
Operational report from the production deployment: single operator completion time of two days and reuse of artifacts across deployments as stated in the paper.
SF2Bench, a compound flooding benchmark comprising 2,452 monitoring stations and 8,557 published files spanning 39 years, validates the multi-agent workflow.
Reported dataset composition and use in the paper: SF2Bench with stated counts and temporal span used to validate the multi-agent workflow.
EnviSmart treats reliability as an architectural property through two mechanisms: (1) a three-track knowledge architecture that externalizes behaviors (governance constraints), domain knowledge (retrievable context), and skills (tool-using procedures) as persistent, interlocking artifacts; and (2) a role-separated multi-agent design where deterministic validators and audited handoffs restore fail-stop semantics at trust boundaries before irreversible steps.
System architecture and design description in the paper; presented as the core reliability mechanisms implemented in EnviSmart.
We introduce EnviSmart, a production data management system deployed on campus-wide storage infrastructure for environmental research.
System description and statement of deployment in the paper; presented as a production deployment (no randomized evaluation reported).
Embedding LLM-driven agents into environmental FAIR data management can externalize operational knowledge and scale curation across heterogeneous data and evolving conventions.
Conceptual / argumentative claim made in the paper as a motivation for the system; no quantitative experiment tied to this statement in the excerpt.
In the short run, with fixed human capital, wages, and job boundaries, AI raises productivity by reducing the time required to perform steps.
Model distinction between short-run (fixed job design and skills) and long-run horizons; short-run optimization shows AI reduces expected execution times for steps, thereby raising productivity.
Aggregating heterogeneous firms that deploy a commonly available AI technology yields an aggregate production function that admits a constant elasticity of substitution (CES) representation with three inputs: aggregate manual labor, aggregate AI-assisted labor, and aggregate capital.
Theoretical aggregation argument drawing on Houthakker (1955) and Levhari (1968), deriving a macro-level CES representation from a microfounded algorithmic cost function defined by firms' joint optimization over AI deployment and job design.
Improvements in AI quality generate non-linear effects on labor demand and wages because firms' cost-minimizing AI deployment and job designs change discretely at particular AI quality thresholds (microfoundation for the productivity J-curve).
Theoretical analysis of discrete switches in the cost-minimizing arrangement as AI success probability and execution times change; characterization of threshold effects and discussion linking to the J-curve phenomenon (model results and comparative statics).
Adjacency to AI-executed steps increases the likelihood that a given step is executed by AI (local complementarities): a step is more likely to be AI-executed in occupations where its neighboring steps are also AI-executed.
Empirical comparisons of conceptually similar steps across occupations paired with workflow adjacency information and realized AI execution outcomes from Anthropic’s Economic Index; statistical tests reported in the paper.
AI-executed steps co-occur in contiguous chains rather than being randomly scattered across a production workflow.
Empirical analysis linking O*NET tasks to human assessments of AI exposure (Eloundou et al., 2024), realized AI execution outcomes from Anthropic’s Economic Index (Handa et al., 2025), and GPT-generated workflow orderings for occupations; statistical tests comparing observed contiguity to random/scaled baselines reported in the paper.
Instrumenting AI use cases with treatment assignment suggests each additional AI use case prompted by treatment leads to approximately 26% higher revenue.
Instrumental variable analysis using randomized treatment as instrument for number of AI use cases in the 515-firm sample; outcome measured as revenue.
Instrumenting AI use cases with treatment assignment suggests each additional AI use case prompted by treatment leads to 0.85 more completed tasks.
Instrumental variable analysis using randomized treatment as instrument for number of AI use cases in the 515-firm sample; outcome measured as completed tasks.
Revenue and investment gains are largest at the 90th percentile and above, suggesting AI expands the upper range of what firms achieve.
Quantile/upper-tail analysis of revenue and investment outcomes in the randomized sample (515 firms); reported concentration of gains at the 90th percentile+.
Treated firms generate 1.9x higher revenue compared to control firms.
RCT with 515 firms; revenue reported by firms during and after the accelerator; comparison of mean revenues between treated and control groups.
Treated firms are 11 percentage points (18%) more likely to acquire paying customers.
RCT with 515 firms; customer acquisition measured in weekly reports / traction outcomes; treatment vs control comparison.
Treated firms complete 12% more tasks.
RCT with 515 firms; weekly progress reports used to measure tasks completed; comparison of completed tasks between treatment (255) and control (260) groups.
The additional AI use cases discovered by treated firms are concentrated in product development and strategy-related domains.
Analysis of categorized AI use cases reported in weekly progress reports from the randomized accelerator sample (515 firms); comparison of functional distribution of use cases between treated and control firms.
Treated firms discover 2.7 additional AI use cases (a 44% increase).
Randomized field experiment in a 3-month accelerator; sample of 515 high-growth startups, 255 treatment and 260 control; weekly progress reports capturing AI use cases; treatment delivered case-study workshops prompting broader search for AI use cases.
Under an extreme calibration where A.I. makes the entire economy grow like the computer industry, growth 'explodes' with incomes becoming infinite in finite time; infinite income does not occur until around 2060 even in this extreme calibration.
Simulation of the endogenous-automation endogenous-growth model calibrated to the fast-automation (computer industry) scenario.
Simulating the calibrated endogenous-automation model under an 'A.I. as a continuation of historical patterns' calibration yields growth rates reaching only 2.5% by 2075.
Forward simulations of an endogenous-growth model calibrated to historical private business sector patterns (model + calibration + simulation).
The main benefit of automation is that it allows production of a task to shift from slowly-improving human labor to rapidly-improving machines.
Theoretical argument within the task-based model and supporting historical accounting showing faster capital-augmenting productivity growth relative to labor.
At the task level, capital productivity has grown at least 3 percentage points per year faster than labor productivity.
Historical task-level growth accounting across sectors using BEA/BLS data and the paper's task-based decomposition; statement appears in abstract and introduction summarizing empirical findings across sectors.
Historically, TFP growth is driven primarily by improvements in capital productivity.
Growth accounting using a task-based model applied to aggregate U.S. data (BEA and BLS) and industry-level data; theoretical decomposition separating capital-augmenting, labor-augmenting, and "other" productivity components.
Economists strongly favor targeted policy interventions such as AI-focused worker retraining (71.8% support) over broad structural interventions like job guarantees (13.7% support) or universal basic income (37.4% support).
Survey items asking respondents to indicate normative support for six policy proposals; reported support percentages for the economist group for specific policies (retraining, job guarantee, UBI).
Economists (as a group) forecast GDP growth of 3.5% under the rapid AI scenario.
Conditional forecasts reported in Key Findings (economist subgroup forecasts under the rapid progress scenario).
The median respondent in each group expects annual U.S. GDP growth of about 2.5% (unconditional forecast).
Unconditional (all-things-considered) survey forecasts of annual GDP growth elicited from respondents across five groups; compared in text to government and private-sector baseline forecasts (typical medium-run 2.0% and long-run 1.7%).
The average economist assigns a 61.4% probability to moderate or rapid AI progress by 2030.
Survey responses from the economist respondent group reporting the mean/average subjective probability for the combined 'moderate' and 'rapid' scenario categories.
The median respondent in each group expects substantial advances in AI capabilities by 2030.
Survey of five respondent groups (academic economists, AI-company employees, AI policy researchers, highly accurate forecasters, and the general public) eliciting unconditional and conditional forecasts about AI capabilities and economic outcomes (details and sample sizes referenced in Section 2.1, not provided in excerpt).
Organizations and policymakers that treat work-time policy as foundational economic planning will better position their economies to harness AI's benefits while mitigating systemic instability.
Policy-prescriptive conclusion based on cross-disciplinary analysis; no empirical trial or quantification offered in the summary.
Work-time reduction can distribute productivity gains more equitably.
Argument supported by examination of historical work-time transitions and pilot programs referenced in the article; no empirical effect sizes or sample details in the summary.