Evidence (7560 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
Filter claims →
Governance
7870 claims
Filter claims →
Human-AI Collaboration
7560 claims
Filtered →
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 |
Human Ai Collab
Remove filter
Developers actively manage the collaboration, externalizing plans into persistent artifacts, and negotiating AI autonomy through context injection and behavioral constraints.
Observed behaviors in chat transcripts and committed artifacts showing developers creating persistent plans, injecting context, and specifying constraints to shape AI behavior.
Developers redistribute cognitive work to AI, delegating diagnosis, comprehension, and validation rather than engaging with code and outputs directly.
Content and interaction analyses of chat sessions showing developer prompts delegating diagnosis, comprehension, and validation tasks to the AI assistants (Cursor and GitHub Copilot) across the dataset.
Conversational programming operates as progressive specification, with developers iteratively refining outputs rather than specifying complete tasks upfront.
Qualitative/content analysis of the 74,998 messages across 11,579 sessions indicating patterns of iterative prompts and refinements rather than one-shot complete specifications.
An Evolutionary Game Theory (EGT) framework produces a 'Red Queen' co-evolutionary dynamic between platforms' algorithmic control and worker behavior in which neither side reaches a stable static equilibrium.
Analytical EGT model and numerical simulations of a population-level game between workers (choices: compliance vs. algorithmic gaming) and a platform varying surveillance strictness; model-based result (no empirical sample size).
This paper proposes three archetypal AI technology types: AI for effort reduction, AI to increase observability, and mechanism-level incentive change AI.
Conceptual taxonomy introduced by the authors (theoretical classification presented in the paper).
AI-driven conversational coaching is increasingly used to support workplace negotiation, yet prior work assumes uniform effectiveness across users.
Background claim in paper indicating prior literature trends and assumptions (stated in introduction/motivation).
Participants were clustered into three profiles -- resilient, overcontrolled, and undercontrolled -- based on the Big-Five personality traits and ARC typology.
Paper reports clustering analysis on participants using Big-Five trait measures and ARC typology; clustering result described as three profiles. Total sample reported as N=267.
We conducted a between-subjects experiment (N=267) comparing theory-driven AI (Trucey), general-purpose AI (Control-AI), and a traditional negotiation handbook (Control-NoAI).
Stated experimental design in paper: between-subjects randomized comparison across three conditions with total sample N=267.
We provide empirical evidence for the inverse parametric knowledge effect: ontological grounding value is inversely proportional to LLM training data coverage of the domain.
Empirical claim based on the controlled experiment (pattern linking grounding value to parametric knowledge coverage reported in paper).
These findings carry implications for workforce transition policy, regional economic planning, and the temporal dynamics of labor market adjustment.
Paper's discussion/interpretation of modeled ATE results and their policy/economic implications; no empirical test provided for policy outcomes.
Practitioners see the socio-emotional gap not as AI's failure to exhibit SEI traits, but as a functional gap in collaborative capabilities.
Reported interpretation from interview data (10 practitioners) indicating practitioners framed the gap functionally rather than as missing emotional traits.
AI technologies and digital platforms have fundamentally altered the organization of work and modes of value realization.
Synthesis of contemporary literature and theoretical analysis in a conceptual study (no empirical sample reported).
Leader emotional intelligence (EI) moderates decision quality, delegation, and managerial communication when generative AI tools (Copilot/ChatGPT) are used in corporate management.
Theoretical EI-moderated human–AI model described in the paper and proposal to test it using a randomized online experiment.
The four-variable account (produced output, underlying understanding, calibration accuracy, self-assessed ability) better explains phenomena like overconfidence, over- and under-reliance on AI, 'crutch' effects, and weak transfer than the simpler claim that generative AI merely amplifies the Dunning–Kruger effect.
Argumentative synthesis in the paper comparing explanatory power of the proposed four-variable framework against the more general Dunning–Kruger metaphor; draws on examples and empirical patterns from the reviewed literature rather than a single empirical test.
A useful working model is 'AI-mediated metacognitive decoupling': LLM use widens the gap among produced output, underlying understanding, calibration accuracy, and self-assessed ability.
Conceptual synthesis and theoretical proposal grounded in reviewed empirical findings from multiple literatures (human–AI interaction, learning research, model evaluation); presented as the paper's working model rather than as a single empirical estimate.
All models exhibit task-dependent confabulation: they perform well on standardized legislative templates (e.g., EU directive transpositions) but generate plausible yet unfounded reasoning for politically idiosyncratic proposals.
Qualitative and quantitative analysis across the 15 proposals showing high-fidelity outputs for standardized/template-like proposals and instances of fabricated or unsupported rationale for idiosyncratic proposals; based on model outputs compared to official explanatory memoranda using the dual evaluation framework.
As technological progress devalues labor, the welfare benefits of steering are at first increased but, beyond a critical threshold, decline and optimal policy shifts toward greater redistribution.
Theoretical model extension analyzing planner's optimal choice as labor's economic value changes; the paper states a non-monotonic relationship with a critical threshold.
Using pre-existing exposure as an instrument for ChatGPT adoption in a long-difference IV design, ChatGPT adoption causes households to spend more time on digital leisure activities while leaving total time spent on productive online activities unchanged.
IV long-difference empirical design: instrumenting household adoption with pre-ChatGPT exposure (2021 browsing); outcome measured as changes in categorized browsing durations (LLM-based classification into 'leisure' vs 'productive' sites); controls include demographic-by-region fixed effects and browsing composition controls.
These patterns are consistent with a reorganization of the scientific production process rather than immediate efficiency gains, in line with theories of general-purpose technologies.
Interpretation linking observed changes in budget allocation, team size, and task breadth (from the proposal dataset and task-level analyses) to theoretical predictions about general-purpose technologies (GPTs); empirical findings show organizational change rather than large average short-run productivity gains.
This paper offers a forward-looking framework that emphasizes the decentralizing potential of AI on labor markets, moving beyond the traditional displacement-versus-creation dichotomy.
Paper's stated contribution; based on conceptual framework and synthesis of historical and contemporary analyses (no empirical validation presented in the abstract).
The emergence of artificial intelligence and robotics is catalyzing a profound transformation in the nature of human labor.
Stated as a central premise in the paper's abstract; supported by the paper's synthesis of economic history, contemporary labor market data, and analysis of digital platform growth (no specific datasets or sample sizes reported in the abstract).
AI agents are approaching an inflection point where the binding constraint shifts from raw capability to how work is delegated, verified, and rewarded at scale.
Conceptual argument presented in the paper's introduction/positioning; no empirical data, experiments, or sample reported.
The resulting AI safety profile is asymmetric: AI is bottlenecked on frontier research (novel tasks) but unbottlenecked on exploiting existing knowledge.
Theoretical implication of the novelty-bottleneck model distinguishing novel (human-judgment) vs. routine (covered by agent prior) components of tasks.
Wall-clock time can be reduced to O(√E) through team parallelism, but total human effort remains O(E).
Model-derived result showing parallelism across humans can speed wall-clock completion time while aggregate human effort does not drop asymptotically.
Better agents improve the coefficient on human effort but not the exponent (i.e., they reduce the constant factor but do not change the asymptotic scaling class).
Analytic result from the stylized model under the paper's assumptions about task decomposition and novelty fraction ν.
The top four models are statistically indistinguishable (mean score 0.147–0.153) while a clear tier gap separates them from the remaining four models (mean score <= 0.113).
Reported mean performance scores across 8 models and statement of statistical indistinguishability for the top four vs lower-tier four; numerical means provided.
Behavioral factors — specifically trust calibration, cognitive load, and affective reactions — shape the transition of corporate AI initiatives from pilot deployments to scalable, sustained use.
Synthesis of human-AI interaction literature integrated with adoption frameworks (TAM and TOE); conceptual linkage rather than new empirical testing in this paper.
AI accelerates value-chain maturation while creating distinct risks — including professional responsibility tensions and potential system-level externalities.
Conceptual argument and risk analysis in the Article (theoretical reasoning and synthesis of management/ethics literature). No empirical causal estimate reported in the excerpt.
The legal profession is at a crossroads, caught between intensifying fears of AI-driven displacement and a generational opportunity for transformation.
Author's synthesis and framing in the Article (conceptual assessment; literature/contextual synthesis). No empirical sample or experiment reported in the excerpt.
This advantage is contingent upon robust AI governance, ethical frameworks, and the transition from 'pilot-lite' projects to integrated, data-driven 'AI-first' business models.
Conditional claim in the paper linking success to governance, ethics, and organizational integration; appears to be normative/analytical rather than empirical in the abstract.
Actual sharing often contradicted willingness to share (the privacy paradox), with consistently high sharing rates across all conditions.
Observed discrepancy reported in the experimental results (N=240): despite variation in willingness-to-share, behavioral sharing rates remained high and similar across human, white-box AI, and black-box AI conditions.
Machine-readable metrics and open scholarly infrastructure are reshaping scholarly profiles and incentives.
Conceptual and historical discussion referring to platforms and metrics (e.g., arXiv, Google Scholar, ORCID) as mechanisms changing incentives; no new empirical estimates provided.
That interconnected ecosystem is fundamentally restructuring who can do science (access), how fast discoveries propagate, and what counts as a valid scientific contribution.
Argumentative claim linking infrastructural and tool changes to changes in access, dissemination speed, and norms of contribution. The paper presents examples and narrative but no systematic empirical evaluation or sample.
The most consequential development is not any single tool but the emergence of an interconnected ecosystem—AI agents, preprint platforms, open source codebases, and citation infrastructure—that forms a feedback loop.
Synthesis/argument based on multiple examples (LLM agents, preprint servers like arXiv, open-source code repositories, citation indices). No quantitative measurement or causal identification reported.
The central tension in AI for science is between automation (building systems that replace human researchers) and augmentation (tools that amplify human creativity and judgement).
Analytical claim based on the paper's review of historical examples and conceptual discussion; no primary data or experimental design reported.
Science has repeatedly delegated its bottlenecks to machines—first inference, then search, then measurement, then the full workflow—and each delegation solves one problem while exposing a harder one underneath.
Interpretive historical argument drawing on examples across AI-for-science milestones (e.g., DENDRAL, search and inference systems, measurement automation, and contemporary end-to-end workflows). No quantitative sample or experimental method reported.
Testing revealed AI excels at computational tasks but consistently misses nuanced factors like new construction rent premiums and infrastructure proximity impacts, validating the framework's hybrid structure as essential for professional-grade underwriting.
Findings from the controlled ChatGPT-4 test on the single 150-unit scenario: qualitative and comparative observations showing AI handled computations well but failed to capture specific local-market nuances, leading authors to endorse a hybrid human-AI framework.
Phase Two requires human-led professional validation to correct AI limitations, apply local market knowledge, and integrate risk factors.
Framework description supported by observations from the controlled test where human review was used to correct AI outputs and apply local knowledge (e.g., adjusting for nuanced market factors).
AI assistance in safety engineering is fundamentally a collaboration design problem rather than merely a software procurement decision: the same tool can either degrade or improve analysis quality depending entirely on how it is used.
Synthesis of the formal framework and analytic results in the paper (theoretical argument; no empirical sample reported).
The paper concludes by discussing open challenges in evaluating harmful manipulation by AI models.
Paper includes a discussion/conclusion section enumerating open challenges; stated in abstract.
We identify significant differences across our tested geographies, suggesting that AI manipulation results from one geographic region may not generalise to others.
Empirical comparison across three locales (US, UK, India) showing statistically significant differences in manipulation outcomes by geography.
Context matters: AI manipulation differs between domains, suggesting that it needs to be evaluated in the high-stakes context(s) in which an AI system is likely to be used.
Comparative analysis across three domains (public policy, finance, health) showing differences in manipulative behaviour and/or impact by domain in the empirical study.
AUROC_2 and M-ratio produce fully inverted model rankings, demonstrating these metrics answer fundamentally different evaluation questions.
Metric comparison across models showing that AUROC_2-based ranking and M-ratio-based ranking are fully inverted in the reported results on the evaluated dataset.
Temperature manipulation shifts Type-2 criterion while meta-d' remains stable for two of four models, dissociating confidence policy from metacognitive capacity.
Experimental manipulation (temperature changes) applied to models; reported result that Type-2 criterion shifted with temperature while meta-d' was stable for two models (out of four) in the 224,000-trial dataset.
Metacognitive efficiency is domain-specific, with different models showing different weakest domains, invisible to aggregate metrics.
Domain-level analyses reported in the paper showing per-domain M-ratio results and identification of different weakest domains per model, contrasted with aggregate metric behavior.
Metacognitive efficiency varies substantially across models even when Type-1 sensitivity is similar — Mistral achieves the highest d' but the lowest M-ratio.
Empirical comparison of Type-1 sensitivity (d') and metacognitive efficiency (M-ratio) across the four evaluated LLMs on the 224,000 QA trials; explicit statement that Mistral had highest d' but lowest M-ratio.
Organizational culture and technological readiness moderate the effectiveness of generative AI integration in decision-making processes.
The paper reports moderation effects tested in the SEM framework using survey data from senior managers, decision-makers, and AI adoption specialists (SmartPLS). No numeric moderator effect sizes or sample size provided in the excerpt.
Implementation of human-replacing technologies leads to significant transformations in skill demand: it reduces reliance on low-skilled labour while increasing demand for qualified engineers, system operators and specialists in digital technologies.
Sector-specific analysis and review of international labour-market studies cited in the article documenting skill-biased effects of automation and digitalization; qualitative assessment for Ukraine's mining and metallurgical sector under workforce shortage conditions.
The framework implies threshold effects in training and capability acquisition: when the teaching horizon lies below the prerequisite depth of the target, additional instruction cannot produce successful completion of teaching; once that depth is reached, completion becomes feasible.
Model-derived threshold result described in the abstract (mathematical analysis of prerequisite depth vs. teaching horizon).
The value of information depends on whether downstream users can absorb and act on it: a signal conveys meaning only to a learner with the structural capacity to decode it (an explanation that clarifies a concept for one user may be indistinguishable from noise to another who lacks the relevant prerequisites).
Conceptual argument motivating the model; theoretical reasoning described in the paper's intro/abstract.