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
Low-information AI neither improves immediate performance nor preserves performance after AI assistance is removed, and is linked to weaker learning overall.
Within-study comparison of low-information AI assistance versus other conditions in the controlled logical reasoning task; immediate and post-AI performance measured (sample size not reported in abstract).
Greater AI usage is associated with weaker skill development: heavy AI users underperform relative to comparable peers, whereas light AI users perform similarly to matched users who do not use AI.
Controlled experiment using a logical reasoning task with on-demand AI assistance; comparison between heavy users, light users, and matched non-users reported in the study (sample size not stated in abstract).
Studies finding true synergy are scarce.
Authors' literature synthesis / meta-analytic overview claiming that few studies report combined human-AI performance exceeding both parties alone (no numerical count provided).
Genuine human-AI synergy—combined performance that exceeds what either party achieves alone—is uncommon.
Authors' synthesis of the literature and meta-analytic findings referenced in the paper indicating scarcity of studies showing combined performance > either alone (no specific counts or sample sizes given in the excerpt).
Agentic systems show persistent failures in repository setup, dependency handling, permission gating, and hardware verification.
Issue-resolution benchmarks and hardware/RTL verification research synthesized in the paper (specific failure rates or sample sizes not provided in abstract).
Controlled studies report slowdowns in mature open-source work when using agentic/code-generation systems.
Controlled studies and trials cited in the paper (no sample sizes given in abstract).
Low-wage workers on platforms perform supporting tasks—such as data annotation and content moderation—that underpin technological infrastructures.
Empirical grounding drawn from cited ethnographic, sociological and anthropological studies and mapping exercises discussed in the paper documenting the kinds of work performed on microtask platforms.
Artificial intelligence (AI) systems depend on invisible labor performed on microtask platforms.
Claim based on synthesis of sociological and anthropological studies cited in the paper mapping production networks and documenting microtask platform work (e.g., data labeling, content moderation) that supports AI.
Socio-technical imaginaries that forecast the displacement of humans from production accompany the technological developments of the Fourth Industrial Revolution.
Conceptual claim supported by literature review and theoretical framing in the paper describing historical and contemporary narratives around automation and the Fourth Industrial Revolution.
Emerging evidence indicates that algorithms often inherit and amplify the historical biases present in training data.
Literature claim in paper referencing 'emerging evidence' and empirical studies (2024–2026) — specific studies, methods, and sample sizes not included in excerpt.
Current agents are not yet able to reliably produce professional-quality spreadsheets at the level of complexity real-world workflows demand.
Conclusion/interpretation based on the benchmark results and qualitative review reported in the paper (supporting quantitative details not shown in the excerpt).
Even the strongest agents frequently fall short of professional finance standards and degrade sharply as the difficulty increases beyond a few chained calculations.
Empirical observation from the paper's benchmark showing performance declines with task difficulty (no numeric breakdown or sample sizes provided in excerpt).
Existing spreadsheet benchmarks do not measure this advanced capability, focusing instead on question-answering or single-formula edits.
Paper's review/assessment of prior benchmarks (stated in introduction); no comprehensive benchmark list or counts provided in excerpt.
In the absence of communicative and institutional safeguards, individually adaptive delegation aggregates into a systemic collective action problem (modeled as a prisoner's dilemma), producing a sociotechnical lock-in that degrades shared epistemic standards.
Game-theoretic analysis in the paper demonstrating aggregation effects and mapping them to a prisoner's-dilemma–style collective action problem (theoretical modeling, no empirical sample).
The complementarity thesis is an over-simplification of the modalities of human-AI interaction and the possibility-space for both individual and collective action that human-AI interaction potentiates.
Theoretical argumentation and conceptual analysis presented in the paper (no empirical data reported).
Seventy-four percent of task misalignments could be attributed to developers who tended to overfocus on efficiency and speed, especially for systems performing tasks in people-facing occupations such as those in the human resources sector.
Result from comparing traits causing incidents to developers' stated preferences (sample of 197 developers) and computing the proportion of misalignments where developer-desired traits matched the traits causing incidents; noted sectoral concentration in people-facing occupations (e.g., HR).
In most cases, workers wanted systems that are precise, insightful, or personal, but instead received systems that are basic, simple, or general.
Qualitative/quantitative comparison of preferred traits (from 202 workers) versus traits observed in AI systems in incident reports (LLM-coded); reported dominant preference traits versus dominant delivered traits.
As many as 83% of workplace incidents stem from worker-AI misalignments.
Result from comparing LLM-extracted traits of AI systems (from 1,524 incident reports) to the traits preferred by workers (sample of 202); counted incidents where traits did not match worker preferences and reported proportion.
4.7% of modified files introduce new Bandit findings (security issues).
Static security analysis using Bandit run before and after each refactoring change on the AIDev Python PRs; reported proportion of modified files that gained new Bandit findings.
24.17% of modified files introduce new Pylint issues, predominantly convention-level violations such as long lines.
Domain-independent static analysis using Pylint applied before and after refactoring commits in the AIDev Python PRs; proportion of modified files with new Pylint findings reported.
So far, we lack a sound conceptual basis for categorizing and comparing these arrangements across organizations.
Statement of a gap in the literature based on the authors' literature review (no quantitative measure of literature coverage provided in abstract).
However, models remain limited in long-horizon reliability and domain-specific planning.
Evaluation results and analysis in paper highlighting failures in maintaining reliability over long-horizon tasks and in planning for domain-specific workflows.
Extensive evaluations reveal that existing agents achieve only 36.0% task success on realistic media editing tasks.
Empirical evaluation reported in paper measuring task success rates of existing GUI agents on the Cutverse benchmark (benchmark size: 186 tasks across 7 apps implied).
There is a "Sim-to-Real" gap: synthetic tests maintain constant memory usage but realistic workflows exhibit linear memory growth of about 3 tokens per message, with consolidation quality emerging as the primary scalability bottleneck.
Empirical comparison between synthetic tests and realistic workflows over the reported message corpus (15,000 messages); reported growth rate (~3 tokens/message) and qualitative identification of consolidation quality as bottleneck.
Full-context models fail at 10,000 messages due to context overflow.
Empirical comparison reported in the paper (within the large-scale evaluation), statement that full-context models fail at ~10,000 messages.
Monolithic approaches suffer from quadratic cost scaling and cognitive degradation when used for long scientific workflows.
Author statement contrasting monolithic approaches with the proposed architecture; conceptual/architectural claim rather than a single quantified experiment.
Context window saturation is a critical bottleneck as LLMs evolve into persistent scientific collaborators, because iterative data analysis and hypothesis refinement rapidly saturate even extended contexts with dense technical content.
Author claim based on observed behavior of scientific workflows; contextual motivation in paper (no specific experiment cited for this general statement).
In strategic decision scenarios, individuals may modify their features after deployment, inducing a post-deployment distribution shift; this strategic manipulation creates a mismatch between the non-strategic prior learned during pretraining and the post-manipulation strategic prior, which leads to systematic prediction bias.
Conceptual/theoretical claim stated in the paper that strategic feature manipulation causes distribution shift and mismatch between learned prior and strategic prior; the abstract asserts this as a cause of systematic prediction bias. No empirical sample sizes given in the abstract.
Underspecified prompts can lead to low-quality answers and additional interaction.
Motivating claim in the paper; presented as the problem the study addresses (no sample size or statistical test reported in the abstract).
Content filtering (blocking searches for Gaza War and Tulsa race massacre).
Documented cases of content filtering cited/synthesized in the paper (specific blocked search topics reported).
AI cataloguing failures (26% F1 accuracy for subject headings).
Empirical studies of AI accuracy in cataloguing synthesized by the paper (reported F1 accuracy for subject heading assignment).
Prompting and AI literacy alone may be insufficient to ensure epistemically independent AI support; system-level approaches are needed to better promote critical engagement in human–AI collaboration.
Authors' interpretation and implication drawn from the experiment's preliminary results: partial improvement from prompt-based training but persistent propagation of contextual errors.
An intervention (prompting training—either general or sycophancy-focused) did not eliminate the propagation of contextual errors from users into AI responses.
Participants received either general or sycophancy-focused prompting training in a pre/post within-subject design; authors report that the propagation of contextual errors persisted after the intervention.
The propagation of user errors into AI responses significantly reduced final user task performance.
Same experiment (n=60); authors report that user errors propagated into AI responses and that this propagation was associated with lower final participant performance on the analytical survival ranking tasks.
The propagation of user errors into AI responses significantly reduced the quality of AI feedback.
Same controlled mixed-design experiment (n=60) with multi-turn human–AI interactions; authors report that when users supplied lower-quality initial responses, those errors were propagated into the AI's responses, reducing AI feedback quality.
LLMs are highly sensitive to user input: lower-quality initial responses lead to poorer AI advice, suggesting that the model mirrors or incorporates user reasoning rather than correcting it or offering better alternatives.
Controlled mixed-design experiment with 60 participants performing multi-turn analytical survival ranking tasks; participants generated individual rankings and then made final decisions after collaborating with an AI assistant. Reported as a preliminary result in the paper.
Twin agents dissolve that boundary, raising a class of trust calibration challenge these frameworks were not designed to handle.
Argument and design observations from the authors' ongoing project presented in the paper; conceptual claim explaining why existing frameworks may be insufficient for twin agents.
When a human colleague doubts a twin agent's output, they face three failure modes (a schema gap, an epistemic gap, and a model artifact) with no reliable attribution path between them.
Conceptual taxonomy derived from the authors' early design observations; presented as an identified set of failure modes in the paper (qualitative, no numeric sample reported in abstract).
Drawing on early design work in an ongoing project, we identify a trust calibration problem specific to this approach.
Based on the authors' early design work (qualitative/design research) described in the paper; no sample size or quantitative metrics reported in the abstract.
Frontier agents struggle with end-to-end completion despite partial progress.
Evaluation experiments reported in the paper showing frontier (state-of-the-art) agents achieving partial progress but failing to reliably complete end-to-end tasks in the OpenComputer benchmark.
Major open challenges for responsible adoption include reliability, bias, privacy, automation bias, transparency, and evaluation.
Authors' identification of risks and open research challenges based on their review/analysis (conceptual synthesis).
Current AI support for code review remains fragmented, with tools focusing on isolated tasks such as reviewer recommendation, PR description generation, or comment suggestion rather than the end-to-end PR review workflow.
Authors' survey/overview of existing AI tooling for code review described in the paper (conceptual / review-based evidence). No quantitative counts provided in the abstract.
AI coding assistants expand the volume of code requiring review, turning code review into a growing bottleneck.
Authors' analytical claim linking increased code production from AI assistants to increased review workload; presented as an observed/trend claim in the paper rather than supported by a quantified study in the abstract.
Code review has evolved for decades, from informal peer checking to today's pull request (PR) workflows, yet it remains a largely manual, uneven, and cognitively demanding process.
Authors' literature review and historical synthesis of code review practices presented in the paper (conceptual / review-based evidence). No empirical sample or experiment reported in the abstract.
Challenges including algorithmic bias, data privacy concerns, high costs, and skill gaps persist across contexts.
Cross-study synthesis of barriers and challenges reported in the 21 included studies spanning multiple contexts.
SMEs face unique resource constraints yet lag in AI-HRM adoption.
Synthesis conclusion from the systematic review of 21 included studies (published 2019–2026) comparing adoption patterns and barriers for SMEs.
Greater automation can obscure rather than eliminate failure modes.
Analytical claim in paper arguing that increased automation hides failures; presented as an interpretive finding rather than a quantified experimental result in the excerpt.
End-to-end autonomous systems have not yet consistently reached major-venue acceptance standards.
Paper's statement based on review of acceptance/peer-review outcomes and standards as of April 2026; no numeric acceptance-rate data presented in the excerpt.
Research code lags far behind pattern-matching benchmarks.
Paper's evaluative claim from its experiments/coding analysis indicating code produced for research tasks is weaker than benchmark performance on pattern-matching tasks; excerpt contains no numerical comparison.
Generated ideas often degrade after implementation.
Paper statement about the gap between idea generation and implemented results reported in the Creation-phase analysis; no quantified follow-up study reported in the excerpt.