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
AI-adopting firms anticipate smaller increases in their own prices and lower medium- to long-term inflation than non-adopters.
Survey questions on firms' price-change expectations and macro inflation expectations, comparing responses of adopting vs non-adopting firms.
AI adoption leads to a contraction of blue-collar employment.
Difference-in-differences analysis of administrative employer–employee records showing decreases in blue-collar employment associated with adoption.
Boundary conditions limit UCF applicability in contexts requiring human accountability or embodied knowledge.
Author-stated caveat in the abstract identifying contexts (accountability, embodied knowledge) where the framework may not apply; theoretical reasoning, no empirical tests.
Existing frameworks (Transaction Cost Economics and Electronic Markets Hypothesis) cannot explain emerging organizational phenomena like GitHub Copilot’s recursive value creation or AI-mediated expert networks.
Conceptual critique in the position paper using illustrative examples (GitHub Copilot, AI-mediated expert networks); no empirical testing or sample provided.
AI governance, ethical concerns, openness, workforce adjustment, and integration complexity are crucial concerns that managers must consider when implementing AI.
Synthesis of risks and challenges reported across the reviewed literature (paper's discussion/conclusion); no specific counts of studies or empirical measures provided in the abstract.
Conventional managerial practices usually encounter difficulties dealing with the flow of information, ineffectiveness of workflow, slow decision making, and redundant administrative processes.
Background statement in the paper's introduction / literature review (narrative claim based on surveyed literature); no specific empirical study or sample size reported in the abstract.
Vulnerable populations—including low-skill workers, aging labour forces, and developing economies—are especially affected by AI-driven changes.
Abstract highlights special attention to vulnerable populations in the review and asserts differential impacts; no specific empirical estimates or sample sizes provided in abstract.
AI displaces routine cognitive and manual tasks.
Explicit finding reported in abstract based on the paper's systematic review of empirical studies (no individual study sample sizes or quantitative estimates provided in abstract).
Persistent AI memory reduced to a retrieval problem (store prior interactions as text, embed them, and ask the model to recover relevant context later) is mismatched to the kinds of memory that agents need in production: exact facts, current state, updates and deletions, aggregation, relations, negative queries, and explicit unknowns.
Argument and conceptual analysis presented in the paper describing types of operations (exact facts, updates/deletions, aggregation, relations, negative queries, explicit unknowns) that retrieval-style memory fails to satisfy; no sample size or quantitative evaluation provided for this specific claim in the excerpt.
This stratification produces trust-based inequality in who can leverage AI while sustaining credibility, voice, and liveness.
Analytical claim based on patterns in 16 interviews indicating differential capacities to conceal/humanize AI lead to unequal ability to both use AI and maintain audience trust and perceived authenticity.
Passing capacity is stratified by educational and professional capital, economic resources and team support, and platform position.
Interview evidence (n=16) showing creators with higher education/professional capital, more economic resources, team support, or advantageous platform positions report greater ability to conceal and perform AI-assisted content.
These invisible authenticity practices reallocate work from generation to downstream repair and performance, complicating claims that AI simply improves efficiency.
Derived from creators' accounts in 16 interviews describing extra downstream editing, verification, and performance labor required after AI generation.
Creators associate legible AI assistance with intertwined trust vulnerabilities, including epistemic unreliability, anticipated relational penalties, and platform authenticity regimes.
Thematic findings from 16 interviews in which creators express concerns about AI-generated content being epistemically unreliable, damaging relationships with audiences, and conflicting with platform authenticity norms.
On authenticity-oriented platforms, visible use of AI can be discrediting for creators.
Reported by creators across 16 in-depth interviews on Xiaohongshu and Douyin; qualitative thematic analysis identifying platform-specific authenticity norms and reputational consequences.
Leaderboard rank alone is insufficient because models with similar pass rates can diverge in overall completion, and task-level discrimination concentrates in a middle band of tasks.
Analytical observations from benchmark results comparing pass rates, overall completion metrics, and per-task discrimination patterns across models; based on the 13-model leaderboard analysis.
Experiments reveal that reliable workflow automation remains far from solved: the leading model passes only 66.7% of tasks and no model reaches 70%.
Experimental evaluation of 13 frontier models on 105 tasks; reported pass rates from the benchmark runs (leading model pass rate 66.7%, no model >=70%).
Many agent benchmarks freeze a curated task set at release time and grade mainly the final response, making it difficult to evaluate agents against evolving workflow demand or verify whether a task was executed.
Qualitative critique in the paper comparing existing benchmark design choices; based on authors' survey/analysis of prevailing benchmark practices (no explicit systematic review sample size reported).
The 2026 Amazon outages illustrate how 'mechanized convergence' (homogenization of code/engineering practices via AI) leads to systemic fragility.
Case study analysis using the 2026 Amazon outages as a single illustrative example; implies qualitative examination of that event.
Recursive training on synthetic code threatens to homogenize the global software reservoir, diminishing the variance required for robust engineering.
Theoretical claim about dataset/model feedback loops; no empirical quantification provided in the text excerpt (argumentative risk assessment).
This epistemological debt erodes the mental models essential for root-cause analysis, widening the gap between system complexity and human comprehension.
Argumentative/theoretical claim supported by reasoning in the paper; no quantified measurement of mental-model erosion reported.
Substituting logical derivation with passive AI verification creates an 'Epistemological Debt' — a hidden carrying cost incurred by engineers.
Theoretical/conceptual assertion within the paper; argued qualitatively rather than demonstrated with controlled empirical data.
The integration of Large Language Models (LLMs) into the software development lifecycle (SDLC) masks a critical socio-technical failure the authors term 'Cognitive-Systemic Collapse.'
Conceptual/theoretical claim presented in the paper's argumentation; no empirical sample or quantitative study reported for this specific naming claim.
Most studies are exploratory (59%) and methodologically diverse, but there is a lack of longitudinal and team-based evaluations.
Authors report study typology counts and note the absence of longitudinal and team-based designs across the reviewed literature.
Studies highlight concerns around cognitive offloading and reduced team collaboration when using LLM-assistants.
Synthesis of reported negative effects in included studies (themes extracted by the authors).
A notable subset of studies identifies critical risks associated with LLM-assistants.
Synthesis across included studies noting reported risks (e.g., cognitive offloading, collaboration issues).
Answer completeness averages 0.40.
Reported average completeness metric for system answers on EnterpriseDocBench (method for computing completeness not given in excerpt).
Hallucination rate does not grow monotonically with document length: short documents and very long ones both hallucinate more than medium ones (28.1% and 23.8% vs. 9.2%).
Empirical measurement of hallucination rates by document-length buckets on EnterpriseDocBench; percentages reported in paper. Sample sizes per bucket not provided in excerpt.
Strong heuristic, single-agent RL, and multi-agent RL baselines (including Greedy, SAC, MAPPO, and MADDPG) achieved net profit in the range $0.58M--$0.70M in the same experiments.
Empirical comparison in the paper's experiments on the NYC-taxi-based EV fleet simulator listing baseline methods and their reported net profits ($0.58M--$0.70M).
Monthly operational cost of running the system is approximately USD 4,000.
Full-scale performance characterization reports monthly cost estimate of approximately USD 4,000.
Prior work has largely focused on developing novel cooperative architectures while overlooking the question of when joint training is necessary.
Literature-review style claim made in the paper asserting a gap in prior research emphasis (novel cooperative architectures) versus investigation of training modality necessity.
The coordination gap advantage (between joint and modular training) diminishes in bottleneck environments, particularly under severe transport and processing constraints.
Results from a sensitivity analysis varying resource scarcity and temporal dominance showing the relative performance gap shrinks under bottleneck conditions with tight transport and processing constraints. Details on experimental scenarios not provided in the abstract.
The framework addresses emerging tensions captured in the Creativity Paradox, whereby GenAI may weaken intrinsic motivation, conceptual risk-taking, and evaluative depth.
Theoretical extension of paradox theory and conceptual discussion of potential negative effects; presented as conceptual risks rather than empirically demonstrated outcomes.
Manual tools like mind maps support structure creation but lack intelligent (AI) assistance.
Paper's comparison of manual tools versus AI-augmented tools (background/related-work discussion; no empirical evaluation reported for this claim).
Current LLM-based systems let users query information but do not let users shape how knowledge is organized.
Paper's analysis of existing tools and limitations (literature/feature comparison described in introduction; no new empirical test reported).
Knowledge workers face increasing challenges in synthesizing information from multiple documents into structured conceptual understanding.
Statement in paper's introduction/motivation; conceptual observation (no empirical data reported here).
The near-uncorrelated rankings and rank shifts on the n=11 subset are driven by a strong negative Adoption-Capability correlation among closed-source high-capability agents within this subset.
Subgroup analysis/observation within the 11-agent SWE-bench overlap indicating a negative correlation between Adoption and Capability for closed-source high-capability agents (no numerical coefficient reported in the excerpt).
Static benchmarks measure what AI agents can do at a fixed point in time but not how they are adopted, maintained, or experienced in deployment.
Conceptual statement in the paper; no empirical sample cited for this specific claim (framing/argumentation).
Standard PayGo degrades substantially under classroom-scale concurrency.
Empirical latency measurements and comparative analysis across throughput tiers and concurrency levels in the instrumented deployment.
Each student query triggers several concurrent API calls whose latencies compound through a parallel-phase maximum effect that single-agent systems do not face.
Architectural description and instrumentation of the four-agent ITAS system (paper reports measurements and latency analysis across tiers and concurrency levels).
In the absence of intervention, individually rational adoption of genAI will assuredly and profoundly reduce collective welfare.
Conclusion drawn from the paper's theoretical model (normative/predictive claim based on model dynamics; no empirical validation or sample reported in abstract).
Habit formation around genAI use can couple otherwise separate domains, so that adoption in low-stakes tasks spills over into high-value tasks and amplifies welfare losses.
Theoretical/model-based claim showing coupling across domains via habit formation (model extension; no empirical sample reported in abstract).
The introduction of genAI—while initially beneficial at the individual level—will reduce social welfare for the most important types of tasks.
Model-derived result: theoretical analysis indicates social-welfare reductions in high-value tasks despite individual gains (no empirical sample reported in abstract).
Generative models are vulnerable to model collapse: when trained on data generated by earlier versions of themselves, their outputs can lose diversity and accuracy.
Theoretical claim / conceptual claim presented in the paper (no empirical sample size given in abstract); refers to degradation of model outputs when trained on self-generated data.
Frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs.
Evaluation of models' self-predicted token cost versus realized token usage across agentic runs on SWE-bench Verified; reported correlations up to 0.39 and systematic underestimation bias.
Models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5.
Cross-model comparisons of average total token consumption per task run across the eight evaluated LLMs on SWE-bench Verified; paper reports average differential between named models and GPT-5.
Input tokens rather than output tokens drive the overall cost of agentic tasks.
Breakdown of token usage into input vs output token components from the analyzed agentic task trajectories on SWE-bench Verified (across the eight LLMs evaluated).
Agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat.
Empirical measurement of token counts from agentic coding task runs compared to runs labeled as code reasoning and code chat across the evaluated trajectories (paper reports comparisons on SWE-bench Verified across eight frontier LLMs).
Industrial robots are widely used in manufacturing, yet most manipulation still depends on fixed waypoint scripts that are brittle to environmental changes.
Background statement in the paper's introduction; general literature/field observation (no new primary data reported for this claim in the abstract).
Each new task domain requires painstaking, expert-driven harness engineering: designing the prompts, tools, orchestration logic, and evaluation criteria that make a foundation model effective.
Author assertion in the paper's introduction/abstract describing the state of practice; no empirical method, dataset, or sample size reported in the excerpt.
Vibe coding (unstructured GenAI-driven coding) promises rapid prototyping but often suffers from architectural drift, limited traceability, and reduced maintainability.
Paper asserts this as a motivating observation and characterizes vibe coding's weaknesses; the abstract frames these as commonly observed problems motivating the Shift-Up approach (no sample size given in abstract).