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
Bankers rate 0% of GPT-5.4's outputs as client-ready.
Human ratings by bankers reported in abstract indicating none of the evaluated outputs from GPT-5.4 were judged client-ready.
Even the best-performing model (GPT-5.4) fails nearly half of the rubric criteria.
Evaluation results reported in abstract: model-level rubric pass/fail aggregated to show best model failure rate approaching ~50% of criteria.
Existing AI benchmarks lack the fidelity to assess economically meaningful progress on professional workflows.
Author assertion in paper abstract arguing current benchmarks are insufficient; presented as motivation for developing BTB rather than empirically tested within the abstract.
Models fail to distinguish reliable predictions from unreliable ones, achieving only ≈20% accuracy regardless of their confidence or whether they judge outcomes as predictable without physical experimentation.
Analysis in the paper comparing model self-reported confidence / predictability judgments to actual accuracy across the 405 tasks; reports ≈20% accuracy irrespective of confidence/predictability judgments.
Human expert performance on the benchmark is approximately 20%.
Reported comparison between human experts and models on SciPredict tasks; the paper states human performance is ≈20% (evaluated on the benchmark tasks).
Model accuracies on SciPredict are 14-26%.
Empirical evaluation of multiple LLMs on the SciPredict benchmark (405 tasks); the paper reports aggregate model accuracy range 14–26%.
Low-skill roles in packaging, sorting, and basic assembly face a high risk of automation.
Paper's findings/prediction derived from task-level classification (routine/repetitive tasks) applied to jobs in Nagpur's medium enterprises; no reported sample size or quantified risk metrics in the excerpt.
Agent frameworks infer authority conversationally, reconstruct accountability from logs, and produce silent errors: incorrect determinations that execute without any human review signal.
Statement/argument in paper describing failure modes of general-purpose agent frameworks; no empirical sample or experiment reported for this claim in the excerpt.
The study's findings are subject to design limitations including an AM/PM session confound, differential attrition, and LLM grading sensitivity to document length.
Authors' reported limitations section citing specific threats to internal validity and measurement (session timing confound, differential attrition across conditions, and grading biases of the LLM used to evaluate documents).
The behavioral scaffolding intervention was associated with substantially lower document production.
Same field experiment (N=388); the behavioral scaffolding required joint AI use within pairs and was compared to unstructured use, with reported reductions in document production in the behavioral condition.
A behavioral scaffolding intervention (a structured protocol requiring joint AI use within pairs) was associated with lower document quality relative to unstructured use.
Field experiment with 388 employees at a Fortune 500 retailer; random/experimental assignment to scaffolding conditions while all participants had access to the same AI tool; comparison reported between behavioral scaffolding condition and unstructured use.
LLMs lag behind humans in sustaining heterogeneity when divergence is rewarded.
Empirical comparison from the experiment showing humans are better able than LLMs to maintain diverse actions when the payoff structure rewards divergence; stated qualitatively in the abstract without numeric effect sizes or sample sizes.
Findings suggest that previous results relying on attitudinal outcomes may generalize poorly to behaviour, and therefore risk substantially mischaracterizing the real-world behavioural impact of AI persuasion.
Interpretation/conclusion based on the paper's empirical results: discrepancy between attitudinal effects and behavioural effects observed in the preregistered experiments.
Rote learning will become obsolete in favor of contextual application.
Paper's forward-looking prediction based on synthesis of adult learning theory and workforce development literature; no empirical sample size or quantified trend data provided.
Foundation-model usage can increase compute-related emissions.
Conceptual/environmental concern highlighted in the paper about the carbon footprint of heavy model use and persistent storage; no quantified emissions analysis or lifecycle assessment presented.
These systems can cause skill atrophy.
Theoretical risk articulated in the paper that reliance on AI assistance may degrade human skills over time; no longitudinal skill-measurement or experimental evidence provided.
The same foundation-model systems can also intensify surveillance.
Cautionary claim in the paper noting the surveillance risk of durable, queryable traces and integrated tooling; presented as a conceptual risk rather than empirically measured increase in surveillance.
Baseline (non-structured) interactions had 16 of 50 accepted on first pass.
Reported counts in the paper for the baseline group (16 accepted of 50 baseline interactions).
In an observational study of documented interactions across four AI tools (Claude, ChatGPT, Cowork, Codex), incomplete context was associated with 72% of iteration cycles.
Observational study reported in the paper covering interactions across four AI tools; the paper reports the 72% figure.
Job insecurity emerges as a critical mediating factor influencing employee attitudes and behavioural responses to generative AI, including upskilling intentions and resistance to technological change.
Review-level synthesis identifying job insecurity reported in included studies as mediating relationships between AI adoption and employee attitudes/behaviours (e.g., upskilling, resistance).
Employees express concerns about role displacement (job loss or role changes) associated with generative AI adoption.
Reported across multiple studies included in the review; the review summarises these concerns as part of mixed employee perceptions.
These positive perceptions coexist with employee concerns about skill obsolescence related to generative AI.
Synthesis of studies included in the review documenting worker concerns about skills becoming obsolete due to AI-driven changes.
The explanatory interface suppresses the natural development of both cognitive trust and emotional trust.
Longitudinal/within-experiment measures of cognitive and emotional trust reported in the RCT; authors state that explanatory interface suppressed the natural development of these trust dimensions in the 120-participant experiment.
The explanatory interface exerts a negative effect on learned trust.
Randomized controlled experiment measuring learned trust; authors report a negative (statistically significant) effect of explanatory interface on learned trust in their sample of 120 pre-service teachers.
The improvement in task performance due to the explanatory interface is confined to the task execution stage and does not transfer to subsequent independent tasks.
Experimental measurement of immediate (during-assisted) task performance and subsequent independent task performance; authors report improvement only during task execution and no transfer effect to later independent tasks in their RCT with 120 participants.
This combination (rapid but uneven capability advance and lagging knowledge about harms/safeguards) creates a difficult policy condition: governments must decide under uncertainty across multiple plausible technological trajectories through 2030.
Reasoned argument in the article synthesizing foresight scenarios and the literature on uncertainty in AI progress (references to documents like OECD foresight and the International AI Safety Report 2026).
Knowledge about harms, safeguards, and effective interventions remains partial and lagged relative to capability advances.
Analytic claim in the article, supported by cited reports and literature that document gaps in understanding of harms and safeguards.
Result 2: When managers are short-termist or worker skill has external value, the decision-maker's optimal policy can produce the augmentation trap, leaving the worker worse off than if AI had never been adopted.
Analytical result from the dynamic model comparing planner/objective variations (short-termist manager or externalities) and showing an outcome labeled the 'augmentation trap'.
Result 1: Even a decision-maker who fully anticipates skill erosion rationally adopts AI when front-loaded productivity gains outweigh long-run skill costs, producing steady-state loss: the worker ends up less productive than before adoption.
Analytical result from the dynamic model showing optimal adoption choice can lead to a steady-state where worker productivity is lower than pre-adoption (model-based comparative statics).
Experimental evidence shows that sustained use of AI tools can erode the expertise on which productivity gains depend (deskilling).
Statement in paper referencing experimental studies (no specific study, method, or sample size reported in the excerpt).
Claude Sonnet 4.6 achieves only 33.3% (completion rate) on ClawBench.
Paper gives a concrete example performance result for Claude Sonnet 4.6 (reported completion percentage on the benchmark).
The authors evaluated 7 frontier models on ClawBench and found that both proprietary and open-source models can complete only a small portion of these tasks.
Paper reports evaluations of 7 models on the ClawBench tasks (empirical evaluation across the benchmark).
Absent a validation sample, researchers cannot assess possible errors in LLM outputs, and consequently seemingly innocuous choices (which model, which prompt) can produce dramatically different parameter estimates.
Warning/claim in the abstract that without validation samples researchers lack a way to assess LLM output errors and that modeling/prompting choices can materially affect parameter estimates; no empirical example or quantified effect reported in the excerpt.
For prediction problems—forecasting outcomes from text—valid conclusions require 'no training leakage' between the LLM's training data and the researcher's sample, which can be enforced through careful model choice and research design.
Stated methodological requirement in the abstract arguing that prediction validity depends on preventing overlap/leakage between model training data and the evaluation sample; no empirical test or sample size given in the excerpt.
There is a 'capability-demand inversion' where skills most demanded in AI-exposed jobs are those LLMs perform least well at in our benchmark.
Cross-referencing SAFI performance with Anthropic Economic Index demand data (reported in paper); described as an observed inversion pattern.
Aggressive compression increased total session cost by 67% despite reducing input tokens by 17%, because it shifted interpretive burden to the model's reasoning phase.
Result reported from the controlled experiment comparing log-format conditions; four conditions described but specific number of sessions/replications not provided in the abstract.
Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall.
Paper reports an evaluation across 17 models and states the maximum overall score observed was below 66%.
Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval.
Statement in paper introduction contrasting prior benchmarks' focus on explicit recall with a claimed gap in evaluating implicit (non-declarative) memory; no systematic literature review or quantitative survey reported in the excerpt.
The impossibility is structural: transparency, audits, and oversight cannot resolve it without reducing autonomy.
Logical consequence derived from the Accountability Incompleteness Theorem and the formal model; stated directly in the paper.
Accountability Incompleteness Theorem: for any collective whose compound autonomy exceeds the Accountability Horizon and whose interaction graph contains a human-AI feedback cycle, no framework can satisfy all four accountability properties simultaneously.
Central theoretical result stated in the paper; supported by a formal impossibility proof based on the model and axioms.
Agentic AI systems violate the above shared accountability assumption not as an engineering limitation but as a mathematical necessity once autonomy exceeds a computable threshold.
Formal theoretical development in the paper culminating in the Accountability Incompleteness Theorem (mathematical proof based on the introduced formal model and axioms).
OpenAI o3 achieves only 17% of optimal collective performance.
Experimental measurement of collective performance for OpenAI o3 in the paper's multi-agent setup (value reported in abstract; no sample size provided there).
We term this the Logic Monopoly -- the agent society's unchecked monopoly over the entire logic chain from planning through execution to evaluation.
Terminology/definition introduced by the authors to describe the conceptual governance problem; definitional claim rather than empirical finding.
When agents from different human principals collaborate at scale, the collective becomes opaque: no single human can observe, audit, or govern the emergent behavior.
Conceptual/analytical claim presented as a security/governance risk in the paper; no empirical study or quantified measurement given in the excerpt.
Participants incentivized for originality incorporate fewer AI suggestions verbatim.
Usage and output-analysis from the pre-registered RCT comparing verbatim incorporation rates of AI suggestions across incentive conditions (no numeric rates provided in excerpt).
Early evidence suggests generative AI increases productivity but does so at the cost of collective diversity, potentially narrowing the set of ideas and perspectives produced.
Statement refers to prior literature/early studies (no specific study, sample size, or method reported in the excerpt).
The study observed errors and limitations in both phases (test generation and refactoring), and manual intervention was necessary at times.
Case study observations reported in the paper describing observed model errors/limitations and instances requiring manual developer intervention.
Current AI coding assistants, such as GitHub Copilot and Amazon CodeWhisperer, emphasize developer speed and convenience, with energy impact not yet a primary focus.
Stated as an observation in the paper; no specific empirical comparison or quantification provided in this excerpt.
Frontend code, replicated across millions of page views, consumes significant energy and contributes directly to digital emissions.
Asserted in paper's introduction; no specific empirical data or sample reported in this excerpt.
We posit that persistence is reduced because AI conditions people to expect immediate answers, denying them the experience of working through challenges on their own.
Authors' proposed psychological mechanism / explanation inferred from observed behavior; presented as a hypothesis rather than directly proven causal mediator.