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
There exists a simple pricing strategy for the platform that covers all M tasks with wait time O(M) while paying only an O(log(M)/M) fraction of the total cost of labor.
Theoretical result from the paper's posted-price procurement model under stated assumptions on workers' estimated costs; formal analysis/proof showing existence of such a pricing strategy for general M (no empirical sample).
Because the technical threshold for this transition is already crossed at modest engineering effort, the window for protective frameworks covering disclosure, consent, compensation and deployment restriction is the present, while deployment remains optional rather than infrastructural.
Authors' normative claim based on their implementation (distillation and deployment) and interpretation that modest engineering sufficed; used to argue policy urgency for disclosure/consent/compensation frameworks.
We term this the Relic condition: when publication systems make stable reasoning architectures legible, extractable and cheaply deployable, the public record of intellectual labor becomes raw material for its own functional replacement.
Conceptual framing introduced by the authors as an interpretation of the observed results and their implications; not an empirical measurement but a named condition/argument.
Agency in software engineering is primarily constrained by organizational policies rather than individual preferences.
Authors' synthesis of qualitative results across the ACTA/Delphi and task/review phases indicating organizational policy factors were cited as primary constraints.
Existing evaluations of large language models remain limited to judgmental tasks in simple formats, such as binary or multiple-choice questions, and do not capture forecasting over continuous quantities.
Literature/benchmark critique asserted in the paper (argument that current benchmarks focus on simple judgmental formats and miss continuous numerical forecasting capabilities).
Calibration degrades sharply at extreme magnitudes, revealing systematic overconfidence across all evaluated models.
Empirical observations from QuantSightBench evaluation showing model calibration performance as a function of magnitude (paper statement noting sharp degradation and overconfidence at extremes).
The top performers Gemini 3.1 Pro (79.1%), Grok 4 (76.4%), and GPT-5.4 (75.3%) all fall at least 10 percentage points short of the 90% coverage target.
Reported empirical coverage percentages from evaluation on QuantSightBench for the listed models (paper provides these percentage values).
None of the 11 evaluated frontier and open-weight models achieves the 90% coverage target.
Empirical evaluation on the newly introduced QuantSightBench benchmark across 11 frontier and open-weight models; models were assessed on empirical coverage of prediction intervals versus a 90% target (paper statement).
The study identified significant implementation challenges including algorithmic bias, digital divide concerns, data privacy risks, and low technology readiness among HR teams in Tier 2 cities.
Synthesis of qualitative case study findings from 4 organizations plus survey responses (N=150) reporting barriers and risks encountered during adoption.
Current attack policies do not saturate LinuxArena (human-crafted attacks evade monitors at substantially higher rates than model-generated attacks, indicating headroom for attackers).
Empirical observation comparing human-crafted attacks (LaStraj) and elicited model-generated attacks; authors interpret higher human evasion rates as evidence that current automated attack policies have not saturated the challenge posed by LinuxArena.
LaStraj is a dataset of human-crafted attack trajectories that evade monitors at substantially higher rates than any model-generated attacks we elicited.
Authors release LaStraj and report empirical comparisons showing human-crafted trajectories evade monitors at higher rates than the model-generated attacks they tested (exact evasion rates and sample sizes not provided in the excerpt).
Against a GPT-5-nano trusted monitor at a 1% step-wise false positive rate, Claude Opus 4.6 achieves roughly a 23% undetected sabotage success rate.
Empirical sabotage evaluation reported by the authors: monitoring a trusted monitor (GPT-5-nano) at a specified step-wise false positive rate and reporting attacking model (Claude Opus 4.6) undetected success rate. (Sample size / number of evaluated runs not provided in the excerpt.)
Current LLMs are unreliable delegates: they introduce sparse but severe errors that silently corrupt documents, compounding over long interaction.
Qualitative and quantitative analysis of errors observed across the DELEGATE-52 experiments (19 LLMs) showing sparse, high-severity, and silently introduced errors that accumulate over long workflows.
Degradation severity is exacerbated by document size, length of interaction, or presence of distractor files.
Additional experiments and analyses varying document size, interaction length, and presence of distractor files reported in the paper showing increased degradation under these conditions.
Agentic tool use does not improve performance on DELEGATE-52.
Additional experiments reported in the paper that compare plain LLM delegation vs. agentic tool-using configurations on DELEGATE-52 and find no performance improvement from agentic tool use.
Even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows.
Reported results from the experiment evaluating 19 LLMs on DELEGATE-52; these named models are highlighted and an average corruption fraction (25%) is reported at the end of long workflows.
Our large-scale experiment with 19 LLMs reveals that current models degrade documents during delegation.
Large-scale experiment reported in the paper evaluating 19 LLMs on DELEGATE-52 long delegated workflows; observed document degradation across models.
Prior research has focused mainly on functional or behavioral alignment rather than moral alignment.
Asserted as a characterization of the literature in the paper (literature-review / conceptual claim; no empirical sampling or quantitative synthesis reported in the supplied text).
Underreliance on AI might deprive software developers of potential gains in productivity and quality.
Stated in the paper and motivated by themes from twenty-two developer interviews indicating missed benefits when developers underuse LLM tools.
Overreliance on AI may lead to long-term negative consequences (e.g., atrophy of critical thinking skills).
Paper explicitly states this risk and grounds the discussion in findings from twenty-two developer interviews (qualitative evidence and participant-reported concerns).
Even explicitly aligned agents exhibit intrinsic biases toward certain ethical frameworks, consistent with known left-leaning tendencies in large language models.
Empirical observation in the alignment-conditioned agents' choices and reasoning frameworks in the triage experiments; authors relate these observations to prior literature on LLM political/ideological tendencies.
This condition of authorship uncertainty reshapes how teams attribute ideas, negotiate accountability, and coordinate collective reasoning.
Theoretical claim based on conceptual analysis in the paper; no empirical method or sample described in the abstract.
As generative AI becomes an ambient presence in collaborative work, a new social ambiguity emerges around authorship and responsibility.
Conceptual argument presented in the paper (theoretical/observational claim). No empirical method or sample size reported in the abstract.
Effective clarification remains challenging in software engineering tasks as not all missing information is equally valuable, and questions must target information users can realistically provide.
Analytic claim supported by the paper's empirical study of clarification in real software engineering tasks (methods mentioned: quantifying types of information affecting task success and simulated-user question-answering; no sample size given in the abstract).
Large language models remain confined to linguistic simulation rather than grounded understanding.
Conceptual assertion in the paper arguing limits of current models; no empirical tests or measurements reported.
Fluency is not reliability: without structures that stabilise both human and model reasoning, AI cannot be trusted or governed where it matters most.
Central thesis/claim of the paper; normative argument synthesising the paper's observations and proposals rather than an empirically tested finding provided here.
Humans often mistake fluency for reliability: when a model responds smoothly, users tend to trust it, even when both model and user are drifting together.
Behavioral/psychological assertion in the paper referencing human interaction patterns with fluent outputs; no experimental data or sample size reported in this paper excerpt.
LLMs produce fluent outputs even when their internal reasoning has drifted; a confident answer can conceal uncertainty, speculation, or inconsistency, and small changes in phrasing can lead to different conclusions.
Conceptual/observational claim presented in the paper; no original empirical test or sample size reported here.
Human decision makers may fail to execute optimal follow-up actions, potentially reducing overall performance.
Motivating argument in the paper (conceptual observation about human suboptimal policies in sequential decision-making).
Stronger reasoning capabilities do not prevent LLMs from defecting in single-shot social dilemmas (i.e., models defect with or without reasoning enabled).
Authors' experiments that explicitly compared model behavior with reasoning enabled vs disabled in single-shot social dilemmas; details not provided in the excerpt.
Repetition-induced cooperation deteriorates drastically when co-players vary.
Authors' experimental observation comparing repeated-game cooperation under fixed vs varying co-players in their study; no quantitative metrics or sample sizes provided in the excerpt.
Our experiments show that recent models — with or without reasoning enabled — consistently defect in single-shot social dilemmas.
Authors' experimental results comparing recent LLMs in single-shot social dilemma games, with reasoning enabled vs disabled; specific models, number of games, and sample sizes are not provided in the excerpt.
Recent works report that LLMs with stronger reasoning capabilities behave less cooperatively in mixed-motive games such as the prisoner's dilemma and public goods settings.
Statement referencing prior literature (recent works) summarized in the paper's introduction/background; no specific dataset or sample size given in the excerpt.
The opacity, fluency, and low-friction interaction patterns of LLMs obscure the boundary between human and machine contribution, leading users to infer competence from outputs rather than from the processes that generate them.
Theoretical argument grounded in prior literature on automation bias and cognitive offloading; presented as explanatory mechanism in the paper rather than an empirically tested causal estimate.
The paper introduces the 'LLM fallacy,' a cognitive attribution error in which individuals misinterpret LLM-assisted outputs as evidence of their own independent competence, producing a systematic divergence between perceived and actual capability.
Conceptual/theoretical claim and formal definition offered in the paper; no empirical validation reported in the abstract.
Efficiency (e.g., minimizing time and cost with AI-only planning) does not equal effectiveness: optimizing for efficiency can erode team cognition and reduce decision quality.
Synthesis of experimental quantitative results (time/cost vs. risk capture and rework) and qualitative assessment indicating that AI-driven efficiency can come at the expense of risk awareness and planning robustness.
Human-only planning incurs substantial overhead.
Same controlled experiment reporting that human-only planning produced higher time and cost overheads relative to AI-assisted approaches.
AI-only planning increases rework due to unstated assumptions.
Experiment measured rework rates and accompanying qualitative analysis attributing increased rework in the AI-only condition to unstated assumptions made by algorithmic planning.
AI-only planning significantly degrades risk capture rates.
Same controlled three-condition experiment on a live client deliverable; paper reports measures/qualitative indicators of risk capture rates and states degradation for AI-only condition.
Two wrong-entity mutations escaped all consumer-contributed layers; only disambiguation and confirmation mechanisms intercept this class.
Empirical observation during the 25 scenario trials spanning seven failure families in the deployed multi-tenant evaluation; the paper reports two instances of wrong-entity mutations that were not blocked by consumer-contributed protections.
The unconstrained AI configuration completed only 17 of 25 tasks.
Same evaluation described above: deployed multi-tenant enterprise application, 25 scenario trials comparing unconstrained AI (safety layers disabled) against bounded autonomy and manual operation.
Conventional methods that use AI predictions as direct proxies for true labels can be inefficient or unreliable when the relationship between AI outputs and human labels is weak or misspecified.
The paper's motivation and critique of standard proxy-using approaches; asserted in the abstract as background rationale for the proposed method.
Asymptomatic effects of AI use evolved into chronic harms such as skill atrophy and identity commoditization among workers.
Reported longitudinal findings from the study indicating progression from asymptomatic (subtle) effects to chronic harms; abstract lists harms but provides no quantification or sample details.
Initial operational gains from AI use masked a phenomenon called 'intuition rust' — a gradual dulling of expert judgment.
Empirical observation reported from the year-long longitudinal study of cancer specialists (phenomenon named and described; abstract provides no quantitative measures or sample size).
Human review remains necessary for maintainability and correct domain interpretation of generated scripts.
Qualitative finding from the mixed-method case study indicating limitations and the need for human oversight.
Validated test specifications accumulate faster than they are automated in many teams, limiting regression coverage and increasing manual work.
Observational claim stated in the paper as a motivating problem; likely based on industry experience and the Hacon case study context.
Existing AutoML methods partially alleviate this burden but remain limited to narrow aspects such as hyperparameter optimization and model selection within predefined search spaces, leaving the full development lifecycle largely dependent on human expertise.
Statement summarizing limitations of prior work (literature review/background in the paper).
Developing high-performing AI models remains a labor-intensive process that requires expert practitioners to iteratively design architectures, engineer representations, implement training pipelines and refine approaches through empirical evaluation.
Background statement in paper's introduction; general literature context rather than a specific empirical test within this paper.
Most AI tooling targets that fraction [the ~10% of the workday spent writing code].
Assertion made in the paper (abstract) as an observed mismatch between where AI tooling focuses and overall developer work activities.
Failure analysis reveals key obstacles (such as breakdowns in cross-artifact consistency) and improvement directions for agentic AI in high-stakes professional workflows.
Qualitative/quantitative failure analysis reported in abstract identifying obstacle categories (example given: cross-artifact consistency breakdowns).