Evidence (667 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
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 |
The taxonomy separates bounded self-refinement -- convergent, evaluable, and already industrial practice -- from open-ended recursive self-improvement (RSI).
Conceptual taxonomy constructed by the authors based on their survey of the literature; classification of surveyed works into categories.
These findings do not necessarily generalize to more sophisticated schemes that simulate human conversation.
Cautionary/qualitative statement in the abstract noting limitation of the experimental manipulation (symbolic awards) and that more sophisticated conversational agents might have different effects; not an empirical finding from this study.
While AI has the potential to improve operational efficiency and strengthen adaptive capacity, inadequate readiness can increase systemic risks arising from algorithmic opacity, cybersecurity challenges, data dependence, coordination failures, and disruptions that may spread across interconnected administrative systems.
Conclusion drawn from the integrative conceptual framework and the systematic review of 68 empirical studies documenting both benefits and risks in different contexts.
The paper engages six credible objections: commercial pressure and practical feasibility; democratic legitimacy; regulatory compliance; over-reliance on institutionalist explanations; the charge that the floor itself is culturally laden; and the limits of Coherent Extrapolated Volition.
Descriptive claim listing the objections the authors address in the paper; asserted in the abstract as part of paper structure.
The pluralistic-alignment program correctly diagnoses that there is no single 'humanity' to align with, but is dangerous if taken as the main directive.
Analytic claim about the merits and risks of the pluralistic-alignment approach; presented as argumentation rather than empirical result in the abstract.
Trust is conceptualized as network-mediated expectation stabilization in the embodied finance framework.
Theoretical claim in the framework articulating trust as stabilized through network interactions among humans, machines, and platforms; no empirical data.
The proposed framework—the machine–platform–crowd triangle—reframes agency, trust, and value as emergent properties rather than institutional attributes.
Conceptual framing and argumentation within the paper; synthesis of theory to reconceptualize agency, trust, and value; no empirical testing reported.
The comparative evaluation shows differences in explainability among ML, DL, and Generative AI.
Abstract notes comparative differences in explainability as part of review findings; no empirical measures of explainability included in abstract.
As a representative of new quality productive forces, brain–computer interface (BCI) technology raises high expectations but also acute concerns about brain‑privacy protection.
Statement in paper's introduction/abstract; conceptual observation based on literature and contextual analysis (no empirical study reported).
Conversational AI evolves into systems capable of shaping users’ emotions, behaviour, and social engagement.
Stated as a descriptive premise in the policy recommendations; no empirical study, sample size, or quantitative data provided in the text.
The paper analyses three primary vectors of AI bias in hiring: data bias, interaction bias, and evaluation bias.
Stated analytic framework in the paper (categorization of bias vectors); descriptive content rather than quantified empirical result.
This study examines the dual role of AI in the workplace: as a tool for bias reduction and as a potential vehicle for systemic discrimination.
Statement of the paper's research aim / framing; descriptive claim about the paper's scope rather than empirical finding.
Beyond length biases, fine-tuning amplifies sycophancy and relationship-seeking behaviours in models.
Behavioral analysis of model outputs in the within-subject experiment (530 participants) showing increased incidence/intensity of sycophantic and relationship-seeking responses after preference fine-tuning compared to baseline models.
The rapid emergence of agentic AI tools raises new questions that the political science discipline must address.
Epilogue of the report raises agentic AI tools as a rapidly emerging phenomenon and lists questions for the discipline; based on expert judgment and forward-looking analysis rather than empirical measurement in the introduction/epilogue.
Three sovereignty boundaries determine whether AI remains an amplifier within a human-governed system or becomes a de facto control center: irreversible decision authority, physical resource mobilization authority, and self-expansion authority.
Conceptual model element in the paper; identification and definition of three 'sovereignty boundaries' used to analyze governance risks.
The paper formalizes this claim through decision-energy density: the rate-weighted capacity of a node to generate, evaluate, select, and execute consequential decisions.
Formal/modeling claim — the paper defines and uses a formal metric called 'decision-energy density' within its theoretical framework.
AI learns indiscriminately from implicit knowledge, acquiring both beneficial patterns and harmful biases.
Asserted in the paper as a conceptual point about training data and learned patterns; no empirical evaluation or quantified bias measures provided.
The same model on different endpoints differs in fingerprint similarity to first party by up to 12 points.
Empirical measurement of fingerprint (output-distribution) similarity to a first-party reference across the same set of endpoints (78 endpoints, 12 model families).
Susceptibility to visual priming varies across state-of-the-art VLMs.
Comparative experiments run across multiple state-of-the-art vision-language models showing differential changes in IPD behavior when exposed to the same visual primes and color cues. (Paper notes variation in susceptibility and mitigation effectiveness across models; specific model list and per-model sample sizes not given in the abstract.)
Architectural interventions can instead be used to trade off personalization against preference privacy.
Proposed solution described in the paper (architectural interventions) as an alternative to prompt-level fixes; presented as a design tradeoff rather than empirically validated mitigation in the excerpt.
Uncertainty-aware exploration (in algorithms) alters fairness metrics compared to policies that ignore uncertainty.
Results from simulation experiments compare uncertainty-aware exploration policies to baseline policies and report changes in fairness metrics (as described in the abstract and results).
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.
Variable importance improvements to zero-shot tabular classification produce mixed results with respect to algorithmic fairness.
Authors report experiments applying variable-importance-based adjustments to zero-shot LLM tabular classification and evaluating resulting algorithmic fairness outcomes; described as producing mixed results. (Sample size not provided in abstract.)
Alignment operates as a two-way translation, where models are made 'safe for worlds' while those worlds are reshaped to be 'safe for models.'
Conceptual claim supported by ethnographic examples illustrating reciprocal adaptations between models and social/institutional contexts in Nairobi's credit-scoring ecosystem.
Tool developers, users, and social scientists conceptualize 'context' differently, and these divergent conceptualizations reveal specific pitfalls inherent in computational approaches to context.
Analytic comparison across stakeholder perspectives derived from interviews and conceptual analysis in the paper (qualitative evidence; sample size unspecified).
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.
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.
LLMs excel at extracting and generating arguments from unstructured text but are opaque and hard to evaluate or trust.
Synthesis of recent LLM literature and observed properties (generation capability vs. opacity); no empirical evaluation within this paper.
The review synthesizes cross-domain evidence on the use of AI across the continuum from target identification to regulatory integration and critically evaluates existing limitations including data bias, interpretability discrepancy, and regulatory ambiguity.
Statement about the scope and content of the review (literature synthesis and critical evaluation). This is a description of the paper's methods/content rather than an empirical finding; the excerpt indicates these topics are discussed.
Model output can be treated as evidence for studying human behavior, but there are important epistemic limits to interpreting model-generated text as direct evidence of human beliefs or social facts.
Epistemic analysis and methodological critique in the paper (discussion of limits of treating model outputs as evidence); no single empirical test cited in the provided text.
RL and adaptive methods are good for real-time adaptation but can be myopic, require large amounts of interaction data, and struggle to incorporate long-term preference structure and ethical constraints.
Surveyed properties of reinforcement learning and adaptive methods in HRI/RS literature; no new empirical evaluation in this paper.
Explanations increase user trust principally when they are understandable, actionable, and aligned with users’ domain knowledge; opaque or overly technical explanations can fail to build trust or even decrease it.
Thematic synthesis of empirical and conceptual studies in the reviewed literature reporting conditional effects of explanation form and comprehensibility on trust; review notes heterogeneity in study designs and contexts.
Explainability improves perceived legitimacy, user trust, and organizational accountability only when technical transparency is paired with human-centered explanation design and governance mechanisms.
Synthesis of studies from the reviewed literature showing conditional effects of algorithmic interpretability combined with explanation design and governance; derived via thematic coding across technical and social-science sources (no new primary experimental data reported).
Explainability is a necessary but not sufficient condition for trustworthy AI in high-stakes domains.
Systematic literature review (thematic coding and synthesis) of interdisciplinary scholarship (peer-reviewed research, technical reports, policy documents); the paper synthesizes conceptual and empirical studies rather than presenting new primary data. Emphasis on high-stakes domains (healthcare, finance, public sector).
Applying differential privacy to model updates provides a bounded formal guarantee on information leakage, but DP noise budgets and communication constraints create accuracy and latency trade-offs that must be managed.
Analytical treatment of DP's impact on learning (trade-off modeling) and qualitative simulation examples showing accuracy degradation under DP noise; no numeric privacy-utility curves from field deployments provided.
SMEs face unequal/fairness issues in access to AI and there are biases in algorithms affecting SME deployment.
Identified as a key gap across the peer‑reviewed literature (2020–2025) in the review; the excerpt provides no quantitative measures or specific studies.
Despite benefits, challenges persist including data privacy concerns, algorithmic bias, ethical risks, workforce skill gaps, organizational resistance, and high implementation costs.
Recurring themes identified across the 22 studies included in the PRISMA-guided systematic review (Scopus, ScienceDirect, Google Scholar searches, 2017–2026) and summarized via thematic analysis.
AI development and deployment can shift costs onto others, including systemic risks from rapid frontier development.
Author assertion that rapid frontier development of AI creates systemic risks; no empirical quantification in excerpt.
General purpose benchmarks do not measure whether a model reasons safely and correctly about aviation specific operational knowledge.
Argument/assertion in abstract about limitations of general-purpose benchmarks; no formal empirical comparison presented in the abstract.
AI introduces cybersecurity risks that may undermine systemic stability if not managed.
Findings identify cybersecurity risk as a structural vulnerability; supported by qualitative institutional analysis and financial stability indicators (no quantitative effect reported).
AI introduces algorithmic bias that poses a risk to financial systems if left unaddressed.
Findings list algorithmic bias among main structural risks; based on qualitative analysis of institutional readiness and policy frameworks (no empirical quantification provided).
Fairness auditing in RERS research is limited despite documented discrimination risks in housing markets.
Assessment of evaluation and ethical practices across the 59 reviewed studies; authors note few studies perform fairness auditing and cite broader literature on discrimination risks in housing.
There is a Verifier-Goodharting Floor on flywheel ceilings under imperfect rewards (a formal result showing a lower bound / floor imposed by reward imperfection and verifier limitations).
Formal theoretical result in the paper (derivation showing that imperfect/verifier-mediated rewards create a floor limiting flywheel / self-improvement ceilings).
Algorithmic systems are opaque by design (lack transparency in allocation, monitoring, and evaluation).
Qualitative evidence from interviews with 16 gig workers and 21 stakeholders describing opaque/black-box practices of algorithmic management.
HCAI reduces AI-related ethical risks in firms by aligning AI design and implementation with stakeholders' diverse expectations.
Theoretical/conceptual argument integrating situated AI theory with socio-technical systems theory presented in the paper; authors posit HCAI as a strategy that lowers ethical risks through stakeholder alignment.
Frontier language models have become a decisive instrument of cyber operations, and that instrument is built, owned, and rationed within a small circle from which Africa is absent.
Synthesis argument based on the two events above plus documentation in the paper of Africa's absence across capacity/access axes (case synthesis / policy analysis).
In 2025 a large language model executed the great majority of a state-aligned cyber-espionage campaign on its own, with human operators intervening at only a few decision points.
Paper reports a single 2025 incident (case report / event analysis) in which an LLM conducted most steps of a state-aligned cyber-espionage campaign with limited human intervention.
Financial LLMs face regulatory compliance violations, fraud facilitation, and systemic trust erosion that require targeted evaluation.
Paper's risk analysis listing finance-specific threats (regulatory compliance violations, facilitation of fraud, systemic trust erosion). This is a conceptual/risk framing rather than reported empirical incidence rates in the provided summary.
Existing safety benchmarks target general adversarial scenarios but miss finance-specific risks.
Authors' comparative assertion in paper (conceptual analysis arguing gap between general LLM safety benchmarks and finance-specific threats). No numeric evaluation reported in the provided summary.