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
Small language models (SLMs) are more cost-efficient and amenable to on-device inference.
Background claim in the paper's introduction/abstract; motivated by literature and the paper's focus on on-device SLMs.
Frontier large language models (LLMs), typically hosted in the cloud, offer strong performance across a wide range of tasks at substantially high cost.
Background claim in the paper's introduction/abstract; supported by literature context and framing rather than a specific experiment in this paper.
Managers can make both (Agentic Technical Debt and Stochastic Tax) visible through lightweight dashboards and governance controls.
Prescriptive/recommendation in the paper; authors state they 'outline' approaches for managers to surface these concepts using dashboards and governance controls. No empirical evaluation or case study evidence reported in the provided excerpt.
Using the three metrics (data product adoption, time-to-find, time-to-insight) ties platform success to measurable business value rather than internal activity.
Argument in the paper about metric selection and their role in assessing platform success (methodological rationale).
A staged framework that shifts ownership from hub to spokes avoids both centralized bottlenecks and uncoordinated decentralization.
Organizational/process recommendation presented in the paper as a way to manage decentralization (design rationale).
Natural-language conversational interfaces democratize access for business users and expose historically underutilized enterprise data.
Proposed UX/interaction benefit asserted in the paper (design claim; no empirical measurement reported in the excerpt).
Large language models (LLMs) that automate governance tasks also lower the barrier for domain practitioners to develop genuine cross-functional expertise spanning business and data engineering, enabling spoke teams to take on greater end-to-end ownership without proportionally increasing their dependence on the hub.
Argument in the paper linking AI/LLM capabilities to skill enablement and reduced hub dependence (conceptual claim; no empirical results in the excerpt).
Domain spokes own business semantics, product backlogs, and local iteration cadence, progressively assuming greater responsibility as they mature (shifting operational ownership outward over time).
Architectural/organizational design element described in the paper (procedural proposal for staged ownership transfer).
A central hub (Center of Excellence) can provide shared platform services, policy automation, and AI-enabled governance that automatically standardizes data products, generates quality rules, drafts data contracts, and reviews changes for regressions.
Functional capabilities described in the proposed architecture; presented as what the hub component will provide (design/specification).
An AI-augmented hub-and-spoke model layered on a modern lakehouse architecture can relax the flexibility-versus-control trade-off inherent in enterprise data platforms.
Proposed architectural solution and theoretical argument in the paper (design proposal; no reported experimental/field results provided in the text excerpt).
Ongoing efforts of the initiative aim to incorporate benchmarks that address concerns about bias by considering alternative perspectives and human centered use cases.
Statement of planned/ongoing work in the paper regarding future benchmark inclusion to address bias and human-centered use cases; no empirical results provided.
Implemented tests include causal translation, model iteration, causal reasoning, conformance, model behavior explanation, suggested model building steps, and suggested model fixes.
Specific list of implemented test categories provided in the paper; descriptive/reporting evidence from the initiative's work.
Tests for several distinct categories of evaluation have been implemented and applied to AI tools that support qualitative model building, quantitative model building, and model discussion.
Paper reports that a set of tests have been implemented and applied to AI tools across qualitative and quantitative modeling and discussion; no sample sizes or numeric evaluation results provided in the excerpt.
A steering group focuses on prioritizing potential benchmarks, while a technical group focuses on implementing the benchmarks in the form of automated tests.
Organizational description in the paper specifying roles (steering group and technical group); no quantitative evaluation reported.
The open source sd ai project hosted by the initiative establishes transparency and enables contributions to be shared broadly.
Descriptive statement about the open-source project hosted by the initiative; no empirical measures of transparency or contribution sharing provided.
The initiative uses open digital and organizational infrastructure to collaboratively evaluate AI tools for modeling and simulation.
Descriptive claim in the paper about organizational approach (open infrastructure and collaborative evaluation); no empirical testing or sample size reported.
The BEAMS Initiative aims to guide the development of AI tools for modeling and simulation toward forms that are responsible and ethical by establishing benchmarks for human centered modeling and simulation practices.
Descriptive statement about the Initiative's stated aims and purpose in the paper; organizational description rather than empirical evidence.
Tools that can automate aspects of modeling practice must complement human expertise, not replace it.
Normative claim made in the paper (argument about human-centered design); no empirical evidence or sample size reported.
AI tools to support real world decision making must be able to build simulation models that inform their recommendations and render them interpretable.
Normative assertion in the paper (position statement / requirement); no empirical study or sample size reported.
The agentic future is not predetermined; leaders must both skate to where the puck is going and actively steer it toward a good place, ensuring innovation delivers welfare gains felt by businesses and consumers around the world.
Normative recommendation offered by the authors; based on conceptual argument and interpretation of the framework rather than empirical testing in the excerpt.
These complementary investments produce the familiar 'productivity J-curve' of general-purpose technologies.
Stated as an economic analogy/claim drawing on general-purpose technology literature; presented as an asserted mechanism rather than shown with new empirical estimates in the excerpt.
The most consequential disruption resides in the third stage (Reconstruction) where workflows and markets are rebuilt around delegation, machine-to-machine interaction, continuous monitoring, and auditable constraints.
Theoretical claim in the paper backed by conceptual reasoning and illustrative sector examples; no quantitative evidence provided in the excerpt.
The system preserves human agency via override mechanisms.
Design description of the collaborative forecasting system that explicitly includes override controls for human users.
The paper provides a rigorous blueprint for designing synergistic, trustworthy, and diagnostic operational planning tools, contributing to the discourse on human-AI collaboration and sustainable information systems (IS).
Stated contribution in the paper's conclusions: presentation of a blueprint and implications for human-AI collaboration and sustainable IS.
Two think-aloud sessions show that human judgment remains critical for high-uncertainty events.
Qualitative evaluation consisting of two think-aloud sessions reported in the paper.
Algorithmic benchmarking reduced forecast errors by 30% over naive baselines.
Quantitative algorithmic benchmarking reported in the evaluation section of the paper (comparison vs. naive baselines).
Teams interacting with more embodied agents display conversational patterns that more closely resemble human–human dialogue.
Conversational analysis comparing dialogue patterns across teams interacting with different embodiment levels; the abstract reports greater similarity to human–human dialogue for teams with higher embodiment agents, but does not provide the similarity metric values or sample sizes.
Human-only teams are more likely to complete all tasks successfully (higher task completion success) than mixed human–AI teams.
Comparison of task completion success between human-only teams and mixed teams in the escape room experiment as reported in the paper; no numerical completion rates provided in the abstract.
Risk-aware layered automation can materially reduce review bottlenecks created by AI-driven code growth without compromising production safety.
Synthesis conclusion based on RADAR deployment results, telemetry (535K+ reviewed diffs, 331K+ landed), and comparative analyses (before-after and difference-in-differences) reported in the paper.
RADAR reduces median diff review wall time by 35%.
Efficiency outcomes reported via telemetry and difference-in-differences analysis stated in the paper; median diff review wall time reduction reported as 35%. Sample likely drawn from RADAR telemetry (535K+ diffs) though not explicitly stated for this metric in the excerpt.
RADAR reduces median time to close by over 330%.
Efficiency outcomes reported via telemetry and difference-in-differences analysis stated in the paper; median time-to-close reduction reported as 'over 330%'. Underlying sample for efficiency analysis likely from the RADAR telemetry (535K+ diffs), though the excerpt does not give the precise sample for this metric.
The Production Incident rate for RADAR-reviewed diffs is 1/50 that of non-RADAR diffs.
Comparative observational analysis reported in the paper; production incident rate for RADAR-reviewed diffs compared to non-RADAR diffs, with the relative rate given as 1/50. Exact absolute counts not provided in the excerpt; overall RADAR telemetry covers 535K+ diffs.
The revert rate for RADAR-reviewed diffs is 1/3 that of non-RADAR diffs.
Comparative observational analysis reported in the paper contrasting RADAR-reviewed diffs with non-RADAR diffs. Underlying counts and exact sample split not provided in the excerpt; overall RADAR telemetry covers 535K+ diffs.
Relaxing the Diff Risk Score threshold from the 25th to the 50th percentile increased the approve rate to 60.31%.
Policy threshold comparison reported in the paper using observational before-after comparisons and system telemetry; approval rate reported as 60.31% after threshold change.
RADAR has reviewed 535K+ diffs and landed 331K+ changes.
System deployment telemetry reported in the paper: 'RADAR has reviewed 535K+ diffs and landed 331K+.'
Agentic AI was responsible for over 80% of that growth in code volume.
Attribution analysis reported in the paper linking growth in code/diff volume to agentic AI sources; described as 'over 80% of that growth.' The underlying attribution method is not detailed in the excerpt.
Per-developer diff volume rose 51% (year over year) at Meta.
Internal telemetry/observational analysis reported in the paper; stated as a 51% increase in per-developer diff volume. No explicit sample size for this specific measure provided in the excerpt.
At Meta, significant lines of code per human-landed diff grew by 105.9% year over year.
Internal telemetry/observational analysis reported in the paper; stated as a year-over-year percentage growth for Meta. No sample size for this specific measure provided in the excerpt.
We discuss implications for Information Systems (IS) design and propose future field evaluations.
Paper includes a discussion section outlining IS design implications and suggestions for future empirical/field work.
The approach preserves statistical rigour, traceability, and nuanced Persevere/Iterate decisions when accelerating experimentation.
Reported outcomes of controlled simulations and description of system design that enforces statistical procedures and logging; stated in manuscript as findings.
Logs render capabilities observable at the feature level, turning 'agentic AI' into a disciplined experimentation infrastructure rather than a generic assistant.
Implementation logs and descriptions from the Node.js instantiation reported in the paper; qualitative claim about observability and traceability at the feature level.
The Multi Agent System reduces time-to-validated-learning by roughly an order of magnitude while preserving statistical rigour, traceability, and nuanced Persevere/Iterate decisions.
Results from the controlled simulations reported in the paper (comparison between agentic multi-agent system and manual B-M-L cycles).
Controlled simulations compare agentic and manual B-M-L cycles on feature ideas.
Reported controlled simulation experiments in the paper comparing agentic (multi-agent) and manual B-M-L cycles; methodological description present in manuscript.
We instantiate them in a Node.js package instrumenting a production-grade SaaS codebase.
Implementation artifact reported in the paper (Node.js package) and description of instrumentation on a production-grade SaaS codebase.
Drawing on the Dynamic Capabilities View, we derive fifteen meta-requirements and thirty-three design principles (consolidated into seven goal-directed groups) for sensing, seizing, reconfiguring, orchestration, and governance.
Design-theory derivation reported in the paper (counts of meta-requirements and design principles are stated in the manuscript).
We propose a multi-agent artefact that operationalises the Build–Measure–Learn (B-M-L) cycle as a closed-loop control system.
Design science study described in the paper; conceptual derivation and artifact instantiation (Node.js package) reported in the manuscript.
AI adoption raises real output.
Panel local projections linking establishment-level AI adoption (share of job postings requiring AI skills) to real output across 13 industries over 2017-2025.
AI adoption raises labor productivity.
Panel local projections estimating the effect of establishment-level AI-skill posting share on labor productivity across 13 industries (2017-2025).
The positive impact of AI application on enterprise innovation efficiency is stronger in labor-intensive firms.
Heterogeneity/subsample analysis using the 2012–2023 A-share panel indicating larger AI effects for labor-intensive firms.
The positive impact of AI application on enterprise innovation efficiency is stronger in asset-intensive firms.
Heterogeneity/subsample analysis on 2012–2023 A-share firm panel showing larger estimated AI effects for firms characterized as asset-intensive.