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
There is a structural shift from 'street level' bureaucracies to 'system-level' architectures that can be defined as the institutional division of 'Artificial Discretion' to algorithmic infrastructures.
Synthesis from the PRISMA-guided systematic review of literature (2018-2026) reporting observed changes in administrative architectures; specific studies not enumerated in abstract.
As a General-Purpose Technology (GPT), Artificial Intelligence (AI) is fundamentally reconfiguring state capacity, as well as the mechanics of global economic management.
Systematic review of current research studies (2018-2026) conducted following PRISMA guidelines; synthesis of literature claiming broad institutional and macroeconomic effects. Number of studies not specified in abstract.
For LLM agents, memory management critically impacts efficiency, quality, and security.
Statement in paper framing and motivation; supported conceptually by literature linking memory design to system properties (no specific experimental details provided in abstract).
Coding patterns are bimodal: in 41% of sessions, agents author virtually all committed code ("vibe coding"), while in 23%, humans write all code themselves.
Empirical analysis of authorship attribution across the 6,000 sessions in the SWE-chat dataset; percentages derived from session-level classification.
A determinism study of 10 replays per case at temperature zero shows both architectures inherit residual API-level nondeterminism, but DPM exposes one nondeterministic call while summarization exposes N compounding calls.
Determinism experiment with 10 replays per case at temperature zero; qualitative/quantitative observation about number of nondeterministic LLM calls exposed by each architecture.
Multi-agent workflows and benchmark evaluation reveal current capabilities, limitations, and research frontiers in agentic AI for physical design.
The paper states it analyzes recent experience with multi-agent workflows and benchmark evaluation; the abstract does not provide specific benchmark names, metrics, or sample sizes.
AI is associated with a shift toward younger, relatively less educated workers.
Reported association in the paper's baseline empirical results linking AI presence/pervasiveness to changes in workforce composition (age and education).
Given the results, educators should revisit pair programming as an educational tool in addition to embracing modern AI.
Authors' recommendation in the paper's conclusion based on experimental findings (performance, workload, emotion, retention outcomes).
Formal network verification has made substantial progress in proving correctness properties but is typically applied in offline, pre-deployment settings and faces challenges in accommodating continuous changes and validating live production behavior.
Authors' summary of the state of the art in network verification (assertion in paper; no empirical data in abstract).
Overall, the proposed HRL framework improves learning efficiency and scalability, outperforming heuristic baselines while remaining below the perfect-information oracle bound.
Results reported in the paper from simulation experiments comparing the HRL framework to heuristic baselines and the oracle; pairwise differences analyzed (Wilcoxon tests referenced). The paper asserts better performance than heuristics but still worse than the oracle.
How software developers interact with AI-powered tools, including Large Language Models (LLMs), plays a vital role in how these AI-powered tools impact them.
Based on qualitative analysis of twenty-two interviews with software developers about using LLMs for software development; asserted as a central finding in the paper's analysis.
Benefits of technology and data analytics are context-dependent, with emerging markets facing unique regulatory and infrastructural barriers.
Narrative synthesis of included studies noting heterogeneity by context and reports of regulatory/infrastructural constraints in emerging markets.
Cybersecurity has a moderating effect on audit data analytics.
Synthesis statement in the review summarizing included studies that report cybersecurity influences the effectiveness/usability of audit data analytics.
CLARITI matches GPT-5's resolution rate on underspecified issues while generating 41% fewer questions.
Empirical evaluation comparing CLARITI and GPT-5 on a task set of underspecified software engineering issues; the result reported in the abstract indicates parity in resolution rate and a quantified reduction in questions (41%) but the abstract does not report sample size, test set composition, or statistical significance.
They can produce fluent outputs that resemble reflection, but lack temporal continuity, causal feedback, and anchoring in real-world interaction.
Descriptive claim made in the text contrasting surface-level fluency with missing properties; no empirical data or experiments provided.
This work establishes a foundation for understanding how generative AI systems not only augment cognitive performance but also reshape self-perception and perceived expertise.
Paper's stated contribution presenting theory and conceptual groundwork; no empirical validation provided in the abstract.
The LLM fallacy has implications for education, hiring, and AI literacy.
Implications and argumentation presented in the paper; these are prospective and conceptual rather than supported by empirical data in the abstract.
Further research is needed to explore the longitudinal impact of these AI deployments on local labor markets and the creation of indigenous datasets that reflect Cameroon’s unique linguistic diversity.
Authors' identified research gaps and recommendations; statement of future research needs rather than empirical result.
Removing safety layers made the system less useful: structured validation feedback guided the model to correct outcomes in fewer turns, while the unconstrained system hallucinated success.
Qualitative and quantitative comparisons from the deployed evaluation across the three conditions (observations about turn counts, validation-feedback loops, and model hallucinations in unconstrained condition over the 25 scenario trials).
Across all settings, AI Organizations composed of aligned models produce solutions with higher utility but greater misalignment compared to a single aligned model.
Reported experimental results aggregated across two practical settings (AI consultancy and AI software team) and 12 tasks; direct comparison between AI Organizations of aligned models and a single aligned model.
Multi-agent "AI organizations" are simultaneously more effective at achieving business goals, but less aligned, than individual AI agents.
Experimental comparison reported in the paper: experiments comparing multi-agent AI organizations to single aligned agents across tasks and settings (described below).
Although some frontier models exceed human performance, model accuracy is still far below what would enable reliable experimental guidance.
Paper reports instances where top-performing (frontier) models outperform aggregate human expert accuracy on SciPredict, but concludes overall accuracies are insufficient for reliable experimental guidance.
Professional and Technical Services, Information, and Finance and Insurance account for approximately 86 percent of the base-case direct contribution.
Sectoral decomposition of base-case direct contribution in the model; paper explicitly reports the three sectors' combined share as ~86%.
Subjectivity persisted in AI-powered recruitment decisions; human judgment remained an important factor.
Theme 2 (subjectivity in AI-powered recruitment) from interviews indicating retained human subjectivity and judgement in recruitment processes (n = 22).
Sensitivity analyses indicate the observed positive belief changes likely reflect recovery from carry-over effects rather than genuine training-induced shifts.
Authors' sensitivity analyses discussed in the paper that examined alternative explanations (e.g., carry-over effects) and concluded the belief-change result is likely due to recovery from such effects.
Simulations demonstrate that standard methods, such as principal components analysis and inverse covariance weighting, can generate spurious cross-study differences, whereas our approach recovers comparable latent treatment effects.
Simulation experiments reported in the paper comparing the proposed method to PCA and inverse covariance weighting; results show PCA and inverse-covariance-weighted estimators can produce spurious cross-study differences while the proposed method recovers comparable latent treatment effects (no simulation sample sizes provided in the abstract).
Big data analytics (BDA) adoption is a risky strategy with potentially high rewards for start-ups.
Stated as a summary conclusion based on empirical analysis of a large sample of start-ups in Germany comparing adopters and non-adopters across multiple performance measures (survival, costs, sales, employee growth, access to financing).
Bounded agents act as an amplifying but not necessary extension to the foundation-model stack for changing work coordination.
Conceptual argument within the paper distinguishing bounded agents from the core stack; no empirical comparison or measurement reported.
The effects of generative AI on work and organisations are heterogeneous and context-dependent, shaped by job roles, skill levels, and institutional environments.
Synthesis across the included studies noting variation in outcomes conditional on role, skill, and institutional context.
The positive effect of big data applications on firms' markups exhibits heterogeneity across organizational, technological, and environmental dimensions.
Paper reports heterogeneity analysis showing variation in the magnitude of the positive markup effect across organizational, technological and environmental factors; based on model implications and empirical subgroup/interaction tests using micro-level firm data (sample size not reported).
If employment losses are relatively small and productivity gains are realised, AI adoption could boost Exchequer revenues. But if job displacement is sizeable, tax receipts fall while welfare spending rises, resulting in potentially large pressures on the public finances.
Conditional fiscal scenarios simulated in the report combining employment, wage and benefit changes with the public finance implications (tax receipts and welfare spending); reported as scenario-based outcomes.
Ireland’s tax and welfare system absorbs most of the income loss for lower income households, and roughly half of the loss for households at the top of the income distribution.
Microsimulation using SWITCH to model taxes and transfers applied to simulated income changes across income groups; reported as a finding in the report.
Qualitative results underscored both perceived benefits in comprehension and challenges when interpretations of gaze behaviors were inaccurate.
Qualitative analysis of participant feedback from the study (n=36) reporting themes of improved comprehension and occasional problems when the assistant misinterpreted gaze.
The productivity decomposition classifies deployments into five regimes that separate beneficial adoption from harmful adoption and identifies which deployments are vulnerable to the augmentation trap.
Model-based taxonomy produced from the analytical decomposition (classification into five regimes described in the paper).
Small differences in managerial incentives can determine which skill path a worker takes (whether they realize full potential or deskill).
Comparative statics / theoretical sensitivity analysis in the dynamic model indicating tipping behavior based on managerial incentives.
Result 3: When AI productivity depends less on worker expertise, workers can permanently diverge in skill: experienced workers realize their full potential while less experienced workers deskill to zero.
Analytical result from the dynamic model showing path-dependent divergence in skill levels under particular parameterizations (lower dependence of AI on worker expertise).
The rise of agentic AI development, where LLM-based agents autonomously read, write, navigate, and debug codebases, introduces a new primary consumer with fundamentally different constraints.
Conceptual claim argued in the paper; refers to the emergence of agentic LLM-based tools as new consumers of software artifacts rather than an empirical measurement; no sample size reported.
Analysis uncovers dramatic asymmetries: inhibition 17.6% vs. preference 75.0%.
Paper reports specific aggregated percentages for two types of implicit effects (inhibition and preference) observed in their analysis; methodology context implies these are results from the benchmark evaluation (300 items / 17 models).
These results suggest the need for AI model development to prioritize scaffolding long-term competence alongside immediate task completion.
Authors' policy/research recommendation based on experimental findings showing short-term gains but longer-term harms.
These effects are observed across a variety of tasks, including mathematical reasoning and reading comprehension.
Trials included multiple task types (explicitly naming mathematical reasoning and reading comprehension); cross-task analysis reported.
Providing issue-specific design guidance reduces design violations, but substantial non-compliance remains.
Intervention experiments in paper: agents were given issue-specific design guidance and resulting patch compliance measured; reported reduction in violations but remaining non-compliance.
Policy implication: encouraging public sharing of AI-assisted solutions offsets the decline associated with private diversion (flow margin) but cannot repair participation-driven deterioration in conditional resolution; the latter requires directly maintaining contributor engagement.
Prescriptive conclusion from the theoretical model comparing interventions: public-sharing encouragement helps with flow-margin diversion but not with supply-side contributor thinning.
Diagnostic prediction: in a congested regime, observing a joint decline in posted volume and conditional resolution implies supply-side pool thinning is quantitatively present; by contrast, volume decline with stable or rising resolution indicates private diversion (flow margin) alone is the dominant force.
Analytical diagnostic derived from the model that links empirical patterns (volume and conditional resolution) to underlying mechanisms; no empirical validation given in the excerpt.
For the short-run optimization problem of AI deployment given fixed job responsibilities and worker skill levels, the firm’s optimal strategy for an m-step job can be computed in time O(m^2) using dynamic programming; the long-run joint optimization including task assignment to workers can also be solved in polynomial time up to an arbitrarily small error term.
Algorithmic results and complexity analysis derived in the theoretical sections and appendices of the paper (dynamic programming construction and polynomial-time solution statements).
Appending a neighboring step to an existing AI chain adds no additional human verification burden (verification is a fixed cost at the chain level), which can make appending steps to a chain optimal even if manual execution is individually preferable for the appended step.
Theoretical model setup and formal argument showing verification is incurred only at the last augmented step of a chain; illustrative examples (data scientist workflow) and comparative-cost reasoning in the paper.
AI chaining can overturn standard comparative advantage logic in assignment: when multiple adjacent steps are executed as an AI chain, a step may be assigned to AI (as part of the chain) even if manual human execution would be preferred for that step in isolation.
Theoretical model of production as an ordered sequence of steps with firms endogenously bundling contiguous steps into tasks and jobs; formal comparative-static arguments and illustrative examples in the paper showing how fixed verification costs per chain change marginal assignment incentives.
Automation leads economic growth to accelerate, but the acceleration is remarkably slow because of the prominence of 'weak links' (an elasticity of substitution among tasks substantially less than one); even when most tasks are automated by rapidly-improving capital, output is constrained by the tasks performed by slowly-improving labor.
Theoretical mechanism from the task-based model (σ < 1 weak-links structure) combined with calibrated simulations that incorporate historical accounting results.
The general public supports both targeted programs and broader interventions (including job guarantees and UBI), contrasting with economists' preferences.
Survey comparisons across groups contrasting normative policy support (textual summary in Key Findings; exact public-group percentages not provided in excerpt).
Unconditional forecasts are relatively close to historical trends, but under the rapid scenario the range of plausible outcomes expands (greater uncertainty).
Comparison of unconditional (all-things-considered) survey forecasts to conditional rapid-scenario forecasts; dispersion metrics referenced qualitatively in Key Findings (detailed variance numbers not provided in excerpt).
Both rapid model improvement and benchmark quality issues contributed to underestimating agent capabilities.
Synthesis of results: improved LLM performance plus audit findings showing benchmark errors together explain the prior underestimation; based on the re-evaluation and audit described in the paper.