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
Authors propose the 'AI orchestra' concept: future development will involve coordinated ensembles of specialized AI agents (code generation, test generation, dependency analysis, security scanning) orchestrated by humans and higher-level controllers.
Theoretical/conceptual argument by the authors grounded in qualitative findings from Netlight (practitioner reports of multiple tools and coordination frictions); this is a forward-looking synthesis rather than an empirically established fact.
Modular and cell‑free platforms could enable decentralized, localized manufacturing of specialty compounds, potentially altering trade flows away from centralized petrochemical hubs.
Conceptual synthesis plus small-scale demonstrations of modular/cell-free units in the reviewed literature; limited pilot projects and discussion of potential scalability and portability.
Canvas Design Principles aimed at reducing algorithmic myopia matter for welfare and regulatory concerns: better adaptive behavior reduces mispricing/misattribution risks but raises questions about transparency, accountability, and systemic amplification of shocks.
Policy and governance implication inferred from the claimed reductions in algorithmic myopia and increased adaptivity; study does not report direct welfare/regulatory impact measurements.
Faster, more accurate identification of demand shifts can compress the window for first‑mover advantages, intensify competitive dynamics, and raise the premium on organizational agility and human–AI integration capabilities.
Theoretical implication derived from observed improvements in signal detection (~5.8×) and resilience; not directly measured as market‑level competitive outcomes in the study.
Product teams evaluating LLM-powered features rely on a spectrum of practices—from informal “vibe checks” to organizational meta-work—to cope with LLMs’ unpredictability.
Qualitative interview study with 19 practitioners; thematic coding of transcripts produced descriptions of a range of evaluation practices used by teams.
Adoption of C.A.P. may reduce demand for routine oversight/clarification roles and increase demand for higher-skill roles such as prompt/system designers and dialogue curators.
Labor demand and task composition analysis presented as a conceptual projection in the paper; no labor-market empirical study reported.
Because failure modes such as definition misalignment and hypothesis creep were observed, the authors argue for regulation/standards around disclosure of AI-assisted scientific claims and archival of verification artifacts.
Policy recommendation in the paper derived from the documented process-level failure modes in the single project; recommendation is prescriptive, not empirically validated beyond the project.
Lower data and compute requirements could decentralize innovation (reducing incumbent advantages tied to massive compute/data), but the complexity of embodied systems and real-world testing could create new specialized incumbents (robotics platforms, simulation providers).
Market-structure hypothesis based on trade-offs between resource needs and platform value; speculative and not empirically tested in the paper.
Improved recovery capability from LEAFE reduces brittle failure modes but may also enable more autonomous behavior in novel settings, increasing both benefits and potential misuse risks.
Safety/risk discussion in the paper linking enhanced recovery/autonomy to both reduced brittleness (benefit) and heightened autonomy-related risks; supported by observed improved recovery behavior in experiments and conceptual risk analysis.
Widespread adoption of LEAFE-like learning could accelerate diffusion of agentic automation across sectors, affecting wages, task allocation, and demand for complementary capital (tooling, monitoring, retraining systems).
High-level economic reasoning in Discussion/Implications section tying observed performance improvements and sample-efficiency gains to possible macroeconomic effects; no empirical macroeconomic data provided.
If smaller tuned models can capture most performance of much larger systems, market power may shift toward specialized, cheaper models plus toolchains, promoting niche competition and verticalized offerings.
Inference from empirical finding that a 7B tuned model achieves 91.2% of a larger model's quality; market-structure implication (theoretical/economic argument, not empirically tested).
Proprietary, high-quality surrogate models could create competitive advantage and barriers to entry, whereas open-source surrogates would democratize access.
This is an implication/policy argument in the paper's discussion about IP and market effects; it is a theoretical/qualitative claim rather than an empirical result from the experiments.
Improved throughput and lower travel costs can induce additional travel demand (rebound), partially offsetting congestion/emissions gains unless paired with demand-management measures.
Theoretical economic reasoning presented in the paper as a caveat; not directly measured in the simulation experiments (no induced-demand dynamic experiments reported).
Pretraining on diverse temporal resolutions increases upfront costs (data acquisition, storage, compute) but can raise model generalization and reduce downstream retraining costs, improving ROI for platform providers.
Paper discusses trade-offs in AI economics, claiming broader pretraining raises costs but yields returns through better generalization and lower adaptation cost. This is a theoretical/cost–benefit argument rather than an empirical finding reported in the summary.
Algorithms could formalize and expand gig opportunities but also risk entrenching platform-based segmentation of the labor market (lock-in effects).
Theoretical implication and cautionary note in the paper; not empirically tested in the pilot as summarized.
Organizational heterogeneity in strategic backing and mentoring explains variation in benefits from AI adoption across firms and sectors, contributing to cross-firm productivity dispersion.
Theoretical claim linking organizational moderators to heterogeneous adoption outcomes; proposed as an empirical research direction without data provided.
Managerial and peer mentoring styles (e.g., directive vs. developmental mentoring) influence how affordances are perceived and actualized, affecting learning, trust, and task allocation in human–AI collaboration.
Theoretical argument drawing on mentoring and organizational behavior literatures integrated with AST/AAT; no empirical tests or sample presented.
Continuous learning capabilities imply ongoing maintenance/data costs but can lower long-run performance degradation and retraining expenses.
Analytic implication derived from system design (continuous model updating) and standard ML maintenance considerations; not empirically quantified in the paper.
Partial substitution of routine diagnostic work by HADT may shift clinicians toward oversight, complex cases, and supervision, raising workforce and retraining considerations.
Paper's discussion of workforce effects and implications for job design (policy/implication statement; not empirically tested in the study).
Organizational forms may shift (e.g., flatter, more modular organizations; increased platform-mediated teams) because easier global coordination changes the cost-benefit calculus for outsourcing and insourcing.
Conceptual mapping from reduced coordination costs to organizational design implications and illustrative examples; no firm-level empirical case studies or panel data presented.
AI-mediated reduction in language frictions could compress wage premia tied to language skills, reduce demand for pure translation/transcription roles, and increase demand for AI-supervisory, verification, and model-prompting roles.
Theoretical labor-market implications and illustrative scenarios linking reduced language frictions to labor supply/demand shifts; no empirical labor-market analysis or sample data included.
Large fixed costs to build standardized databases and automated laboratories imply economies of scale that can favor well-capitalized firms and centralized public infrastructures, potentially increasing barriers to entry.
Economic analysis and reasoning in the implications section drawing on the costs of data/infrastructure discussed in the reviewed literature; not empirically measured in the paper.
Implication (interpretive): The positive association between AI adoption and resilience suggests AI can strengthen institutions’ ability to detect and respond to shocks, but model risks and correlated behaviours (e.g., common models) could create systemic vulnerabilities that need management.
Inference combining reported positive association (β = 0.35 for resilience) with theoretical considerations about model risk and systemic correlation discussed in the paper.
Traditional drivers—macroeconomic stability, public spending and physical investment—remain important determinants of economic progress; AI’s economic gains will likely require institutional readiness and supportive economic contexts and may emerge over time.
Conclusion drawn from the combination of empirical findings (significant positive effects for GFCF, government expenditure, population growth; non-positive/negative result for AI patents) and theoretical reasoning about adoption costs, complementary skills/infrastructure, and institutional factors. This is a conceptual inference rather than a direct empirical test in the reported models.
AI in higher education is not simply a technological shift but a structural transformation requiring deliberate, critically informed governance grounded in equity and human agency.
Normative/conceptual conclusion drawn by the author from the thematic analysis and the critical AI media literacy framing; presented as the paper's principal argument or recommendation. (Supported qualitatively by themes from the analyzed discussions rather than quantitative causal evidence.)
Findings have important implications for enterprise strategy and economic policy in early-stage AI adoption environments.
Discussion and policy implications drawn from the paper's theoretical framework and empirical results; not tested empirically within the paper.
Lower barriers to producing design concepts with GenAI could enable more freelancing and entry by non-traditional providers, altering market structure and intensifying competition at the lower end of the value chain.
Speculative implication extrapolated from interview findings and economic reasoning in the paper; not empirically tested within the study.
Demand for designers will likely shift toward individuals combining domain expertise with algorithmic/AI fluency (prompting strategies, tool orchestration), potentially increasing returns to these hybrid skills.
Inference and implication drawn from interview themes about algorithmic thinking and authors' policy/economics discussion; not empirically tested in study.
Standard productivity metrics (e.g., output per hour) may misprice value if temporal quality matters; firms will face trade‑offs between maximizing throughput and preserving richer subjective temporality that affects long‑run creativity, morale, and retention.
Conceptual economic reasoning and literature synthesis on attention and productivity; no empirical studies or longitudinal workplace data presented.
Investors and firms may need to include metrics of experiential quality (subjective well‑being, sustained attention quality) alongside productivity metrics when valuing neurotech and human–AI platforms.
Normative/economic implication argued from the framework; no empirical valuation studies or survey of investor behavior included.
Adoption of advanced simulation and AI could affect productivity, returns to capital versus labor, trade and outsourcing patterns, and distributional outcomes, with benefits potentially concentrated among large firms.
Theoretical implications and discussion in the paper's AI economics section; framed as suggested areas for future study rather than empirically established effects.
Reported pilot gains, if scaled, could shift firm‑level returns and industry productivity measures, but gains are contingent on coordinated adoption; uneven uptake may produce winner‑takes‑more dynamics among technologically advanced firms.
Inference from pilot results and economic reasoning in the reviewed literature; no large‑scale empirical validation provided in the review.
Topology is the dominant factor for price stability and scalability compared to other swept variables (load, presence of hybrid integrator, governance constraints).
Factor-ablation analysis within the 1,620-run simulation study showing the largest explanatory effect (largest changes in volatility and scalability metrics) attributable to graph topology rather than load, hybrid flag, or governance settings.
Adoption heterogeneity may widen productivity dispersion across firms and contribute to market concentration, since organizations with better data, processes, and training budgets will capture more benefit.
Economic interpretation of literature and survey findings; speculative projection rather than empirical measurement within the study.
Promoting AI without complementary policies for physical capital and labor may produce uneven outcomes; policy sequencing and complementarity (capital modernization, workforce upskilling) are recommended to produce more inclusive growth.
Interpretation of asymmetric leverage and sensitivity results; policy implications drawn from model behavior and sensitivity experiments, not from causal identification in the data.
Demand for mid-level, routine-focused developer roles could compress while demand rises for verification, security, and AI–human orchestration skills.
Theoretical task-replacement argument based on observed capabilities of LLMs and synthesized user study evidence; limited direct labor-market empirical evidence in the reviewed literature.
Routine coding tasks may be partially automated, shifting human labor toward verification, integration, architecture, and domain-specific tasks.
Task-composition studies, user studies showing LLMs handle boilerplate/routine work, and economic inference synthesized across studies.
Societal acceptance of AI-generated audiovisual media is uncertain and could range from widespread uptake to broad rejection.
Discussion drawing on mixed empirical studies and scenario construction in the review; the paper notes contradictory findings in existing studies but does not provide primary survey data or sample sizes.
If cognitive interlocks are widely adopted, many negative externalities can be internalized and AI-driven productivity gains can be realized more sustainably; absent such controls, equilibrium may drift toward higher error rates and systemic incidents.
Long-run equilibrium argument based on theoretical reasoning and conditional claims; no longitudinal or cross-firm empirical evidence presented.
If AI raises the quality and pace of research, social returns to public research funding could increase, but distributional concerns and negative externalities must be managed to realize aggregate welfare gains.
Welfare implication discussed in the paper. Framed as conditional and theoretical; not empirically quantified in the abstract.
Policy interventions (data governance, transparency, reproducibility standards, ethical guidelines) will shape adoption and externalities (misinformation, misuse, reproducibility crises).
Policy recommendation/implication stated in the paper. This is a normative and predictive claim grounded in governance literature; the abstract does not present empirical evaluation of specific policies.
Labor demand effects are ambiguous: junior/entry-level demand may be reduced for some tasks while demand for verification and higher-skill roles may rise.
Economic reasoning, early observational signals, and theoretical task-reallocation frameworks; empirical longitudinal evidence is limited or absent.
The effectiveness of generative AI depends critically on human-AI workflows: prompt design, iterative refinement, and human vetting materially affect outcomes.
Qualitative analyses of interaction patterns and experiments manipulating prompting/iteration showing variation in outcomes; many studies report improved outputs after iterative prompting and human-in-the-loop refinement.
CRAEA-style systems could increase household productivity and substitute for some routine in-home human labor, altering demand for certain service roles and increasing demand for higher-skill roles (e.g., maintenance, AI oversight).
Paper's implications/economic analysis and qualitative extrapolation based on observed performance improvements in simulation; no empirical labor-market or deployment data provided to substantiate real-world labor substitution claims.
Integrated ERP vendors embedding AI could strengthen vendor lock-in, while interoperable AI layers may foster ecosystems and specialized entrants; empirical work is needed to determine market outcomes.
Conceptual discussion and observed vendor behavior in practitioner literature; explicit statement in the paper that empirical analysis is required.
Market demand is likely to bifurcate: high-value clinical markets will require rigorous explainability and neuroscientific grounding (higher willingness-to-pay), while research and consumer segments may tolerate black-box models (lower margins).
Market segmentation argument built from differing end-user requirements and tolerance for opaque models; presented as a projected implication rather than an empirically tested market study.
Persistent declines in self-efficacy after passive AI exposure suggest potential for skill atrophy and slower reversion when tasks must be performed without AI.
Inference from observed persistent reductions in self-efficacy post-return in the experiment; skill atrophy and reversion costs not directly measured—this is an implied consequence.
Firms that adopt passive, copy-based AI workflows risk psychological costs that could offset short-run productivity gains from AI.
Inference drawn from experimental findings of reduced efficacy/ownership/meaningfulness under passive use and short-term enjoyment gains; not directly tested for firm-level productivity or turnover—extrapolation from individual-level psychological measures.
Teams often produce evaluation outputs (tests, metrics, user feedback) but lack mechanisms, processes, or technical levers to convert those outputs into actionable engineering or product changes—a novel “results-actionability gap.”
Recurring theme from the 19 practitioner interviews and coding; authors explicitly articulate and label this gap based on participants' reports.
The study confirms several previously documented evaluation challenges with LLMs: model unpredictability, metric mismatch, high human-evaluation costs, and difficulty reproducing failures.
Interview data from 19 practitioners; thematic analysis flagged these recurring problems as reported by participants and aligned with prior literature.