Evidence (6491 claims)
Adoption
8570 claims
Productivity
7631 claims
Governance
6869 claims
Human-AI Collaboration
6491 claims
Org Design
4175 claims
Innovation
4114 claims
Labor Markets
3566 claims
Skills & Training
2966 claims
Inequality
2066 claims
Evidence Matrix
Claim counts by outcome category and direction of finding.
| Outcome | Positive | Negative | Mixed | Null | Total |
|---|---|---|---|---|---|
| Other | 758 | 199 | 100 | 900 | 2007 |
| Governance & Regulation | 826 | 400 | 191 | 122 | 1563 |
| Organizational Efficiency | 777 | 193 | 124 | 84 | 1189 |
| Technology Adoption Rate | 635 | 233 | 124 | 97 | 1098 |
| Research Productivity | 422 | 128 | 57 | 336 | 954 |
| Output Quality | 476 | 179 | 59 | 47 | 761 |
| Decision Quality | 328 | 177 | 81 | 47 | 640 |
| Firm Productivity | 435 | 57 | 88 | 20 | 606 |
| AI Safety & Ethics | 218 | 277 | 65 | 33 | 599 |
| Market Structure | 180 | 170 | 123 | 24 | 502 |
| Task Allocation | 213 | 64 | 72 | 33 | 387 |
| Skill Acquisition | 170 | 61 | 61 | 17 | 309 |
| Innovation Output | 203 | 27 | 43 | 18 | 292 |
| Employment Level | 105 | 54 | 107 | 13 | 281 |
| Fiscal & Macroeconomic | 131 | 69 | 43 | 26 | 276 |
| Consumer Welfare | 117 | 63 | 42 | 11 | 233 |
| Firm Revenue | 153 | 48 | 26 | 3 | 230 |
| Task Completion Time | 173 | 31 | 8 | 12 | 225 |
| Inequality Measures | 44 | 122 | 49 | 6 | 221 |
| Worker Satisfaction | 89 | 65 | 22 | 12 | 188 |
| Error Rate | 69 | 92 | 10 | 2 | 173 |
| Regulatory Compliance | 77 | 69 | 14 | 5 | 165 |
| Automation Exposure | 56 | 56 | 26 | 13 | 154 |
| Training Effectiveness | 94 | 21 | 13 | 19 | 149 |
| Wages & Compensation | 77 | 36 | 25 | 6 | 144 |
| Team Performance | 86 | 17 | 27 | 10 | 141 |
| Developer Productivity | 95 | 17 | 14 | 6 | 133 |
| Job Displacement | 12 | 80 | 20 | 1 | 113 |
| Hiring & Recruitment | 52 | 7 | 8 | 3 | 70 |
| Creative Output | 31 | 18 | 8 | 3 | 61 |
| Skill Obsolescence | 5 | 46 | 6 | 1 | 58 |
| Social Protection | 27 | 16 | 8 | 2 | 53 |
| Labor Share of Income | 17 | 19 | 17 | — | 53 |
| Worker Turnover | 11 | 12 | — | 3 | 26 |
| Industry | — | — | — | 1 | 1 |
Human Ai Collab
Remove filter
We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs.
Paper claim of introducing a new benchmark named ImplicitMemBench; it states novelty ('first systematic benchmark') and describes design around three constructs (Procedural Memory, Priming, Classical Conditioning).
Above the Accountability Horizon, distributed accountability mechanisms become necessary.
Derived implication from the Accountability Incompleteness Theorem and the paper's discussion of policy responses; theoretical argument rather than empirical evidence.
Experiments on 3,000 synthetic collectives confirm all predictions with zero violations.
Reported simulation experiments: N = 3,000 synthetic Human-Agent Collectives evaluated against the theoretical predictions; reported outcome was zero violations of the predicted impossibility/conditions.
Below the threshold (Accountability Horizon), legitimate frameworks exist, establishing a sharp phase transition between regimes where the four properties can and cannot be satisfied.
Constructive existence results and theoretical arguments in the paper showing frameworks that satisfy the axioms when compound autonomy is below the defined threshold.
We introduce Human-Agent Collectives, a formalisation of joint human-AI systems where agents are modelled as state-policy tuples within a shared structural causal model.
Paper provides a formal model/definition called Human-Agent Collectives (mathematical formalisation and definitions).
Existing accountability frameworks for AI systems, legal, ethical, and regulatory, rest on a shared assumption: for any consequential outcome, at least one identifiable person had enough involvement and foresight to bear meaningful responsibility.
Stated as background assumption in the paper's introduction/abstract; supported by citation to prior legal/ethical/regulatory frameworks (normative claim about literature). No empirical test reported in this paper.
Tiny sharing incentives improve models with weak cooperation.
Experimental intervention reported in the paper: adding small sharing incentives and observing improved cooperation among weakly-cooperative models (stated in abstract; no quantitative effect size or sample size provided there).
Explicit protocols double performance for low-competence models.
Experimental intervention reported in the paper: introducing explicit protocols in the multi-agent setup and observing a doubling of performance for low-competence models (stated in abstract; no sample size reported there).
OpenAI o3-mini reaches 50% of optimal collective performance.
Experimental measurement of collective performance for OpenAI o3-mini in the paper's multi-agent setup (value reported in abstract; no sample size provided there).
The core thesis is alignment-through-accountability: if each agent is aligned with its human owner through the accountability chain, then the collective converges on behavior aligned with human intent -- without top-down rules.
Central theoretical thesis of the paper; presented as a hypothesis to be evaluated rather than as an empirically demonstrated result in the excerpt.
We propose the Separation of Power (SoP) model, a constitutional governance architecture deployed on public blockchain that breaks this monopoly through three structural separations: agents legislate operational rules as smart contracts, deterministic software executes within those contracts, and humans adjudicate through a complete ownership chain binding every agent to a responsible principal.
Design proposal / governance architecture presented in the paper; the text asserts that the model 'breaks this monopoly' but provides no experimental results in the excerpt to validate that claim.
Those incentivized for originality rely on the model more selectively for brainstorming, proofreading, and targeted edits.
Behavioral/usage measures from the RCT indicating task-level patterns of AI use (described qualitatively in excerpt; no quantitative task-level usage breakdown provided).
Participants rewarded for originality relative to peers produce collectively more diverse writing than those rewarded for quality alone.
Randomized assignment to incentive conditions (originality reward vs. quality reward) in the pre-registered RCT on a creative writing task (no sample size or numerical effect provided in excerpt).
Early evidence has shown that generative AI can increase individual-level productivity.
Statement refers to prior literature/early studies (no specific study, sample size, or method reported in the excerpt).
Much of the business and management literature approaches artificial intelligence primarily as a technological capability that enhances efficiency and productivity.
Literature review / characterization of existing business and management literature cited in the paper; no quantitative synthesis or meta-analysis reported.
The paper documents best practices for iteratively generating tests to capture existing system behavior before model-assisted refactoring.
Methodological contributions in the paper: recommended workflow and practices for iterative test generation to lock down behavior prior to refactoring.
The described workflow constrained refactoring changes and enabled model-assisted refactoring under developer supervision, with proposed code changes validated by passing tests.
Methodological description in the paper: iterative test generation to capture existing behavior, then model-assisted refactoring with developer oversight and test-based validation.
The generated tests achieved up to 78% branch coverage in critical modules.
Measured branch coverage reported in the case study for critical modules after running the generated tests.
Using coding models, we generated nearly 16,000 lines of reliable unit tests in hours rather than weeks.
Single case study reported in the paper: automated unit test generation using coding models; reported aggregate output of generated tests and a qualitative time comparison (hours vs weeks).
This work demonstrates how energy considerations can be embedded directly into AI-assisted coding workflows, supporting developers as they engage with energy implications through actionable feedback.
Concluding claim based on the system implementation and evaluation described (benchmarks and controlled study).
EcoAssist reduced per-website energy by 13-16% on average.
Reported result from the benchmark evaluation of 500 websites (effect size reported as 13-16%).
We introduce EcoAssist, an energy-aware assistant integrated into an IDE that analyzes AI-generated frontend code, estimates its energy footprint, and proposes targeted optimizations.
Description of the system introduced by the authors (implementation claim).
AI assistance improves short-term performance on tasks (people do better while using the AI).
Randomized controlled trials (N = 1,222) showing better immediate task outcomes when participants used AI assistance.
Analyses use fixed-effects regression and structural equation modeling (SEM) on panel data from OECD countries.
Methods statement in the paper indicating use of fixed-effects and SEM applied to OECD-country panel data.
This paper provides the first cross-country empirical validation of AI-augmented scientific evaluation systems.
Authors' stated novelty claim that prior work lacked cross-country empirical quantification and that their OECD panel study is the first such validation.
A one standard deviation increase in AIRC is associated with an 18–25% increase in scientific productivity.
Reported point estimate/range from regression/SEM results linking a 1 SD change in the constructed AIRC to productivity outcomes in the OECD panel.
AI-assisted evaluation significantly enhances scientific productivity.
Fixed-effects regression and structural equation modeling (SEM) applied to panel data from OECD countries; reported association between AIRC and research output.
We construct a novel AI Review Capability Index (AIRC).
Paper reports creation of a new composite index (AIRC) to measure national-level AI capability in peer review; constructed and applied to panel data from OECD countries.
China's 'Global Community of Shared Future' white paper and Putin's 2024 Valdai address provide empirical evidence for an articulated alternative vision to the Western‑led global order.
Qualitative textual/readings of the cited official documents (the white paper and the Valdai address) used in the paper as empirical support; no quantitative content analysis or sample coding is reported.
Technical workers' potential for progressive transformation lies not just in their strategic importance and specialized knowledge but in their ability to build solidarity across the broader ecosystem of AI labour while operating between otherwise incommensurable philosophical and infrastructural systems.
Normative/theoretical claim combining philosophical analysis (Chinese Marxism, Bauman) with empirical literature on hidden AI labour and infrastructure competition (Muldoon et al., 2024); offered as an interpretive synthesis rather than empirically validated causal finding.
Technical workers occupy a strategic position at the intersection of competing infrastructural systems and alternative visions of global order, making them potentially crucial actors in determining the outcome of the current interregnum.
Argumentative claim supported by secondary empirical literature cited in the paper (Muldoon, Graham, and Cant, 2024) on hidden labour supporting AI systems and on geopolitical competition over digital infrastructure; presented as qualitative/interpretive evidence rather than primary quantitative measurement.
The semi-core's challenge to Western hegemony creates unique conditions for systemic transformation.
The paper advances this as a theoretical argument synthesizing World‑Systems theory, Demirel (2024), Bauman's philosophical work, and interpretive readings of official Chinese and Russian documents; no quantitative causal test is reported.
The emergence of a 'semi-core' is represented most prominently by China and Russia.
The paper cites Ege Demirel (2024) as the primary conceptual source and draws on textual evidence from China's 'Global Community of Shared Future' white paper and Putin's 2024 Valdai address; presented via World‑Systems theoretical framing and qualitative/discourse analysis.
AI agents autonomously plan, invoke external tools, and execute multi-step action chains with reduced human involvement.
Definitional framing provided by the authors describing the technical/functional characteristics of 'AI agents' as used in the paper.
The provider's foundational compliance task is an exhaustive inventory of the agent's external actions, data flows, connected systems, and affected persons.
Authors' recommendation/practical conclusion derived from the regulatory mapping (prescriptive guidance rather than empirical measurement).
We propose a twelve-step compliance architecture and a regulatory trigger mapping connecting agent actions to applicable legislation.
Paper asserts it includes a proposed 12-step compliance architecture and a mapping between agent actions and regulatory triggers (explicit step count provided).
We present a practical taxonomy of nine agent deployment categories mapping concrete actions to regulatory triggers.
Paper states it includes a taxonomy comprising nine deployment categories (explicit count provided).
This paper provides the first systematic regulatory mapping for AI agent providers integrating (a) draft harmonised standards under Standardisation Request M/613 to CEN/CENELEC JTC 21 as of January 2026, (b) the GPAI Code of Practice published in July 2025, (c) the CRA harmonised standards programme under Mandate M/606 accepted in April 2025, and (d) the Digital Omnibus proposals of November 2025.
Author claim about the paper's contribution and scope (novelty/first-of-its-kind mapping integrating specified standards and documents).
AI agents - i.e. AI systems that autonomously plan, invoke external tools, and execute multi-step action chains with reduced human involvement - are being deployed at scale across enterprise functions ranging from customer service and recruitment to clinical decision support and critical infrastructure management.
Author assertion in the paper's introductory framing; no empirical sample size or quantified deployment statistics provided in the excerpt.
Rather than indiscriminate collection of context-relevant data, researchers and practitioners should adopt interactional practices to embed generative AI systems more appropriately into users' contexts of use.
Normative conclusion/provocation drawn from the paper's empirical findings and analysis of failure modes; presented as a recommendation (not an empirical effect; based on qualitative synthesis).
Users deploy concrete strategies to address failures of generative AI systems to account for context.
Empirical observations from interviews describing user-devised workarounds and strategies; qualitative cases/examples (sample size not provided).
We hypothesize the emergent necessity of a 'Compliance Premium,' indicating wage resilience increasingly tied to risk-absorption capacity.
Hypothesis proposed by authors based on observed institutional/business risk differentials from HITL validation and OAI patterns; framed as a forward-looking interpretation rather than demonstrated empirical result.
Non-routine cognitive roles highly dependent on symbolic manipulation (e.g., Data Scientists) face unprecedented exposure, with OAI ≈ 0.70.
Reported OAI value for example occupation(s) (Data Scientists) derived from the algorithmic aggregation across DWAs; claim presented as a key empirical finding.
We utilize a multi-agent LLM ensemble to score both technical feasibility and business risk for DWAs.
Method description: deployment of a multi-agent LLM ensemble to produce scores on technical feasibility and business risk per DWA. Specific ensemble composition and hyperparameters not provided in the excerpt.
We introduce a Tech-Risk Dual-Factor Model that jointly scores technical feasibility and business risk to re-evaluate occupational exposure to LLMs.
Methodological contribution described in the paper (model specification). Implementation details described elsewhere in paper (see multi-agent scoring and aggregation), but claim itself is the introduction of the model.
All code, infrastructure, and benchmark data are released to facilitate future research in realistic computer-use agents.
Statement of release in paper (availability claim).
Applying the same auditing principle at test time — a separate VLM reviews completed trajectories and provides feedback — improves Gemini-3-Flash on CUA-World-Long from 11.5% to 14.0%.
Experimental result reported in paper: evaluation of Gemini-3-Flash with/without test-time VLM auditing on CUA-World-Long, reported scores 11.5% -> 14.0%.
Distilling successful trajectories from the training split into a 2B vision-language model outperforms models 2× its size.
Modeling experiments reported in paper: distilled 2B VLM evaluated against larger models (2× size). Exact evaluation metrics and baseline model sizes not specified in excerpt.
CUA-World-Long is a challenging long-horizon benchmark with tasks often requiring over 500 steps, far exceeding existing benchmarks.
Benchmark description in paper reporting typical task lengths ("often requiring over 500 steps") and comparison to existing benchmarks.
The result is CUA-World, a collection of over 10K long-horizon tasks spanning domains from medical science and astronomy to engineering and enterprise systems, each configured with realistic data along with train and test splits.
Dataset release / creation claim specifying >10,000 tasks and train/test splits.