Evidence (8974 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
10085 claims
Filter claims →
Productivity
8974 claims
Filtered →
Governance
8062 claims
Filter claims →
Human-AI Collaboration
7749 claims
Filter claims →
Org Design
5057 claims
Filter claims →
Innovation
4896 claims
Filter claims →
Labor Markets
4088 claims
Filter claims →
Skills & Training
3372 claims
Filter claims →
Inequality
2377 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 | 882 | 244 | 117 | 1097 | 2424 |
| Governance & Regulation | 1010 | 469 | 229 | 135 | 1875 |
| Organizational Efficiency | 977 | 235 | 149 | 90 | 1462 |
| Technology Adoption Rate | 781 | 299 | 143 | 128 | 1362 |
| Research Productivity | 506 | 155 | 74 | 363 | 1110 |
| Output Quality | 555 | 219 | 71 | 70 | 915 |
| Decision Quality | 395 | 200 | 95 | 54 | 751 |
| Firm Productivity | 523 | 67 | 101 | 27 | 724 |
| AI Safety & Ethics | 262 | 309 | 75 | 36 | 688 |
| Market Structure | 195 | 201 | 135 | 30 | 566 |
| Task Allocation | 248 | 77 | 96 | 38 | 464 |
| Innovation Output | 300 | 34 | 55 | 20 | 411 |
| Skill Acquisition | 207 | 75 | 65 | 21 | 368 |
| Employment Level | 138 | 67 | 119 | 24 | 350 |
| Fiscal & Macroeconomic | 156 | 80 | 53 | 33 | 329 |
| Task Completion Time | 211 | 38 | 13 | 16 | 280 |
| Firm Revenue | 183 | 52 | 29 | 5 | 270 |
| Consumer Welfare | 131 | 77 | 48 | 13 | 269 |
| Inequality Measures | 50 | 141 | 54 | 9 | 254 |
| Worker Satisfaction | 104 | 85 | 25 | 13 | 227 |
| Error Rate | 87 | 112 | 11 | 5 | 215 |
| Automation Exposure | 69 | 69 | 37 | 20 | 198 |
| Wages & Compensation | 102 | 49 | 31 | 11 | 193 |
| Team Performance | 115 | 30 | 30 | 11 | 187 |
| Regulatory Compliance | 88 | 74 | 17 | 7 | 186 |
| Training Effectiveness | 109 | 22 | 14 | 21 | 168 |
| Developer Productivity | 116 | 21 | 15 | 8 | 161 |
| Job Displacement | 12 | 92 | 26 | 1 | 131 |
| Hiring & Recruitment | 57 | 12 | 9 | 5 | 83 |
| Skill Obsolescence | 6 | 59 | 10 | 2 | 77 |
| Social Protection | 43 | 17 | 8 | 2 | 70 |
| Creative Output | 35 | 21 | 9 | 4 | 70 |
| Labor Share of Income | 18 | 23 | 17 | 1 | 59 |
| Worker Turnover | 15 | 16 | — | 4 | 35 |
| Industry | — | — | — | 1 | 1 |
Productivity
Remove filter
LLM agents can perform in silico biology tasks that previously required experienced human biologists.
Results from the ABC-Bench evaluation reported in the paper, where LLM agents completed multiple bio-relevant tasks (coding for liquid handlers, DNA fragment design, evasion of DNA screening) and outperformed a median expert human baseline.
Large language models (LLMs) are rapidly acquiring capabilities relevant to biological research, from literature synthesis to interpretation of experimental data.
Statement in paper's introduction framing observed trends in LLM capabilities; general literature and examples implied but no specific quantitative study in this paper reported to support the broad trend.
The paper proposes an evolutionary framework of AI-Economy transformation and calls for further research on governance, sustainability, and inclusive growth.
Abstract states the paper suggests an evolutionary framework and points to future research directions (governance, sustainability, inclusive growth); this is a conceptual recommendation rather than an empirical result.
Generative AI can transform value generation by enriching cognitive work instead of automating habitual processes.
Abstract claim synthesizing reviewed literature that generative models augment cognitive work; no empirical effect sizes or study counts given in abstract.
Deep Learning (DL) hastens automation and capital deepening in high-skill industries.
Synthesis claim in abstract from reviewed literature; no specific empirical estimates or sample sizes provided in abstract.
Machine Learning (ML) mainly boosts productivity by increasing predictive efficiency.
Synthesis claim in abstract based on the systematic review of peer-reviewed literature (Scopus and SCI); no specific empirical studies or sample sizes cited in abstract.
The review estimates sectoral, macroeconomic, and labor market effects of ML, DL, and Generative AI.
Stated scope in abstract: review used to estimate sectoral, macroeconomic, and labor market effects; no quantitative details provided in abstract.
This paper systematically reviewed peer-reviewed journal articles indexed in the Scopus and SCI databases.
Stated method in abstract: systematic review of peer-reviewed journal articles indexed in Scopus and SCI; no sample size or study count reported in abstract.
AI agents (like Computer) accelerate workflows, enhance output quality, reduce costs, and expand the breadth and depth of automated work.
Summary conclusion synthesizing the empirical findings from Perplexity production data comparisons (matched sessions, per-query dissatisfaction, completion time and cost estimates, and content analyses).
Computer lowers estimated time by 87% and estimated cost by 94% compared to humans equipped with Search alone.
Estimated time and cost comparisons derived from Perplexity matched-task completion times and cost models comparing Computer-enabled workflows to human workflows using Search.
Computer reduces completion time from 269 to 36 minutes on matched tasks.
Matched-task analysis of Perplexity production data comparing task completion times when users used Computer versus Search.
Autonomy increases execution quality, with per-query dissatisfaction rates 55% lower on Computer than on Search.
Per-query dissatisfaction metric computed from Perplexity product usage data comparing Computer and Search queries (matched comparisons).
Computer shifts follow-up query distribution toward higher-order work such as verification and extension.
Comparison of follow-up query types in Perplexity production data for matched tasks across Computer and Search sessions, showing relative increases in verification/extension queries after Computer use.
Computer performs 26 minutes of autonomous work per user session, versus 33 seconds for Search.
Analysis of Perplexity production data comparing sessions with near-identical initial query pairs as natural experiments for the same underlying task attempted with both products (matched-session comparison).
Ultimately, this framework provides technology managers with a verifiable, evidence-based pathway toward resilient, net-zero Industry 5.0 ecosystems.
Conclusion/assertion in paper positioning the framework as a practical pathway; described qualitatively without empirical outcome measures or quantified evidence.
The architecture demonstrates how fabs can export cryptographically signed compliance tokens via International Data Spaces (IDS) connectors without exposing proprietary process recipes.
Claim of demonstration in paper; implies a prototype or illustrative workflow using cryptographic signing and IDS connectors, but no empirical deployment, sample, or measured disclosure-risk reduction reported.
By executing Virtual Metrology (VM) predictions and Federated Machine Learning (FML) inside hardware-rooted Trusted Execution Environments (TEEs), this architecture resolves the Data Sovereignty Paradox.
Technical claim based on proposed use of TEEs with VM and FML in the paper; presented as conceptual/architectural resolution rather than empirically validated result.
Structured as an interoperable network protocol stack, the framework coordinates an automated, five-step "relay race" between Facility, Process Engineering, and Finance proxy teams to align factory-floor yield models with macro-level sustainability mandates.
Architectural and protocol-level description in the paper (system design); no quantitative alignment metrics or empirical validation reported.
We propose a shift from reactive automation to autonomous governance through "Professional Proxies"—role-based agentic workflows executing within hardware-isolated trust zones.
Design proposal and conceptual workflow model presented in the paper; no field trial or user study reported.
We introduce a zero-trust socio-technical orchestration framework that operationalizes a six-layer SSbD reference architecture within trustworthy industrial data spaces.
Proposed system architecture described in the paper (design/proposal); no reported empirical deployment or quantitative evaluation.
Lightweight, self-improvable grounding layers can turn general coding agents into practical operators of scientific software.
Conclusion drawn from empirical results across GEOS, OpenFOAM, and LAMMPS experiments showing SIGA enabling agents to operate simulators effectively.
Grounding can reduce the across-seed standard deviation by 16x on the held-out GEOS set.
Reported reduction in across-seed standard deviation in held-out GEOS experiments (numerical multiplier provided).
On a harder held-out set, grounding raises TreeSim from 0.720 to 0.789, a roughly 10% relative gain over the bare agent.
Held-out evaluation comparing bare agent vs grounded agent on GEOS reported in the paper (numerical TreeSim scores provided).
On GEOS, SIGA attains TreeSim above 0.90, matching the extended-budget human expert.
Empirical evaluation on GEOS reporting TreeSim score (quality metric) above 0.90 for SIGA and parity with a human expert.
On GEOS, SIGA produces a complete GEOS deck in about five minutes, matching an extended-budget human expert who took about three hours—a roughly 36x wall-clock speedup.
Empirical evaluation on GEOS reported in the paper (timing comparison between SIGA and an extended-budget human expert). Sample size not stated in abstract.
SIGA (Simulator-Interface Grounding Adapter) supplies a simulator's executable contract through retrieval, procedural memory, in-trajectory validation, and validation-enforced termination.
Method description in the paper introducing SIGA (architectural/components claim).
The system maintains safety guarantees through layered authorization and rollback mechanisms.
Stated safety design: layered authorization and rollback mechanisms described as part of deployment; no empirical safety metrics provided in the abstract.
Agentic AI systems can achieve autonomous resolution rates exceeding 90% for common incident categories.
Reported production result in the paper claiming >90% autonomous resolution for common incident categories; the abstract does not provide sample size, time period, or incident counts.
The architecture has been deployed in production at a major cloud provider.
Stated deployment claim in the paper (production deployment at a major cloud provider); no further deployment metrics or number of sites provided in the abstract.
Architectural principles include hierarchical agent decomposition, skills-based tool invocation via standardized protocols, structured knowledge encoding from operational runbooks, progressive autonomy with safety boundaries, and closed-loop verification.
Paper lists these principles as the design foundation for the architecture; descriptive claim about system design.
The system employs a multi-agent orchestration framework where specialized AI agents collaborate to detect, diagnose, and remediate network incidents without human intervention.
Architectural/system description in the paper; stated implementation details of multi-agent orchestration and agent specialization.
We present an agentic AI architecture for autonomous incident resolution in large-scale network operations.
Description of a proposed system architecture in the paper (design and architectural principles), presented as the main contribution.
The AARRI-Bench dataset and code are released at https://github.com/AARR-bench/AARRI-bench.
Explicit statement in the paper providing a GitHub URL for released data.
Our results indicate that developing researcher-like AI requires further exploration of research behavior, rather than merely complex scaffolding.
Conclusion/recommendation derived from the benchmark results reported in the paper.
We propose AARRI-Bench (Act As a Real Research Intern), the first benchmark in the AARR series that focuses on granular research scenarios rather than macro-level execution capabilities.
Paper introduces AARRI-Bench as the initial benchmark in the proposed series; claim is descriptive about the paper's contribution.
We conceptualize the AARR (Act As a Real Researcher) benchmark series to evaluate whether agents can emulate the professionalism, thoroughness, and nuanced reasoning of human researchers in granular research scenarios.
Paper proposes the AARR benchmark series and describes its focus and motivation; this is a methodological/conceptual contribution of the paper.
Agents have demonstrated remarkable proficiency in complex, long-horizon coding tasks and even autonomous experiment execution.
Stated as background/motivating claim in the paper's introduction; no specific experimental details or sample sizes provided in the excerpt.
Organizations increasingly use intelligent systems for high-stakes strategic decision-making (SDM).
Introductory/background statement in the paper summarizing trends in practice and motivating the research; based on literature review and observed practice (no study data reported here).
Given modest efficiency gains and limited costs, the authors recommend adjusting for AI-generated predictions as a regular empirical practice.
Policy/recommendation statement in the paper's abstract based on theoretical arguments, simulations, and three empirical applications; not a quantified causal claim.
Overall, efficiency gains from adjusting for AI-generated predictions are real but modest, with greater benefits in studies that contain substantial text and other unstructured data.
Summary claim in abstract based on the paper's simulations and three empirical applications; no numeric effect sizes reported in abstract.
The ideas are demonstrated in simulations and three empirical applications: a survey mega-study, an email marketing A/B test, and a large-scale technology platform experiment.
Statement in abstract claiming empirical demonstrations; the specific applications are named but sample sizes are not reported in the abstract.
The paper provides implementation guidance, including how to obtain continuous scores from discrete LLM outputs and how to use LLMs to featurize unstructured inputs as auxiliary covariates.
Stated in abstract that the paper contains practical implementation guidance; evidence basis is descriptive (contents of the paper) rather than empirical.
No new estimators are required: researchers can simply include AI predictions as covariates in standard regression adjustment, analogous to adjusting for a prognostic score.
Methodological claim in the paper (abstract) presenting the approach as compatible with standard regression adjustment; presumably supported by theoretical exposition in the paper.
AI-generated predictions can be used to reduce variance in randomized experiments by including them as covariates in regression adjustment.
Theoretical argument in the paper plus demonstrations via simulations and three empirical applications (survey mega-study, email marketing A/B test, large-scale platform experiment) as stated in the abstract.
Generative AI and large language models can produce realistic predictions of human behavior from rich, unstructured inputs with little to no task-specific training data.
Statement in paper's introduction/abstract referencing recent work that uses LLMs to predict human responses from unstructured inputs; likely supported by literature citations and examples but no sample size reported in abstract.
Introducing completeness-aware decision utility (informed decision-quality = decision-quality × gold-coverage), C reaches 7.43 vs. A 1.76 and B 2.57 on this metric.
Computed metric (decision-quality × gold-coverage) applied to agents' measured decision-quality and gold-coverage on the benchmark (13 assets).
On the curated long-tail subset, agent C reaches 0.93 recovered coverage vs. A 0.26 and B 0.30.
Subset analysis (long-tail subset) of the curated gold record recovery fractions comparing A, B, and C; subset size not specified in abstract.
Across the same benchmark, B increases objectivity from 3.16 to 3.30 (vs. A).
Measured objectivity score comparison between agent A and agent B on the 13-asset stratified benchmark.
Across a 13-asset stratified benchmark, adding the public structured tools and playbook (B) improves tier-in-range accuracy from 0.80 to 0.89 (vs. A).
Comparison of tier-in-range accuracy between agent A and agent B on the 13-asset stratified benchmark (controlled ablation).
For knowledge-intensive scientific decisions, the limiting factor is often the evidence substrate the agent can access.
Controlled three-arm ablation experiment on a production drug-asset valuation agent across a 13-asset stratified benchmark comparing variants with differing access to structured/public/proprietary evidence.