Evidence (13661 claims)
Adoption
8339 claims
Productivity
7479 claims
Governance
6715 claims
Human-AI Collaboration
6267 claims
Org Design
4098 claims
Innovation
3987 claims
Labor Markets
3488 claims
Skills & Training
2888 claims
Inequality
2016 claims
Evidence Matrix
Claim counts by outcome category and direction of finding.
| Outcome | Positive | Negative | Mixed | Null | Total |
|---|---|---|---|---|---|
| Other | 740 | 192 | 95 | 871 | 1945 |
| Governance & Regulation | 796 | 388 | 185 | 119 | 1512 |
| Organizational Efficiency | 765 | 186 | 123 | 82 | 1166 |
| Technology Adoption Rate | 610 | 227 | 121 | 95 | 1061 |
| Research Productivity | 409 | 121 | 56 | 331 | 928 |
| Output Quality | 464 | 174 | 58 | 47 | 743 |
| Decision Quality | 318 | 173 | 75 | 42 | 615 |
| Firm Productivity | 432 | 55 | 88 | 20 | 601 |
| AI Safety & Ethics | 214 | 273 | 65 | 33 | 589 |
| Market Structure | 175 | 165 | 120 | 24 | 489 |
| Task Allocation | 206 | 64 | 70 | 31 | 376 |
| Skill Acquisition | 161 | 57 | 57 | 16 | 291 |
| Innovation Output | 201 | 27 | 41 | 18 | 288 |
| Fiscal & Macroeconomic | 130 | 69 | 43 | 26 | 275 |
| Employment Level | 104 | 50 | 105 | 13 | 274 |
| Consumer Welfare | 116 | 62 | 42 | 11 | 231 |
| Firm Revenue | 149 | 45 | 26 | 3 | 223 |
| Inequality Measures | 43 | 120 | 49 | 6 | 218 |
| Task Completion Time | 164 | 29 | 8 | 12 | 214 |
| Worker Satisfaction | 89 | 60 | 20 | 12 | 181 |
| Error Rate | 69 | 89 | 9 | 2 | 169 |
| Regulatory Compliance | 74 | 67 | 14 | 4 | 159 |
| Training Effectiveness | 91 | 19 | 13 | 19 | 144 |
| Wages & Compensation | 77 | 33 | 25 | 6 | 141 |
| Team Performance | 86 | 17 | 27 | 9 | 140 |
| Automation Exposure | 49 | 50 | 22 | 12 | 136 |
| Developer Productivity | 91 | 17 | 14 | 5 | 128 |
| Job Displacement | 12 | 80 | 19 | 1 | 112 |
| Hiring & Recruitment | 51 | 7 | 8 | 3 | 69 |
| Creative Output | 31 | 16 | 7 | 2 | 57 |
| Skill Obsolescence | 5 | 43 | 6 | 1 | 55 |
| Social Protection | 27 | 16 | 8 | 2 | 53 |
| Labor Share of Income | 17 | 17 | 17 | — | 51 |
| Worker Turnover | 11 | 12 | — | 3 | 26 |
| Industry | — | — | — | 1 | 1 |
HAI-C task complexity increases employees' HAI-C tech-learning anxiety.
Longitudinal survey data (n=497) analyzed with hierarchical regression; reported as a finding in the Results that task complexity amplifies tech-learning anxiety.
When models err, their incorrect predictions disproportionately lean intervention-oriented.
Error analysis of model predictions showing that among incorrect predictions, a larger share favor intervention-oriented causal signs than market-oriented ones (directional skew in errors).
Across 18 of 20 models, accuracy is systematically higher when the empirically verified causal sign aligns with intervention-oriented expectations than with market-oriented ones.
Model-by-model accuracy comparison broken down by whether the empirically verified causal sign aligns with intervention-oriented vs market-oriented expectations; observed higher accuracy for intervention-aligned cases in 18/20 models.
GenAI-related benefits are likely to materialize only when AI capabilities are embedded in standardized routines, integrated data infrastructures, and cross-functional governance arrangements (organizational embedding).
Paper's synthesized process model and interpretive case evidence from the three firms indicating organizational conditions required for observed/documented AI effects.
GenAI-related capabilities enhance analysis by translating complex data into more interpretable, scenario-sensitive, and action-oriented outputs (analytical augmentation).
Interpretive finding from analysis of disclosures and literature; presented as a second linked mechanism through which GenAI may influence management accounting.
GenAI-related capabilities broaden the informational basis of management accounting by making operational, service, quality, and ecosystem data more usable in planning and control (information enrichment).
Interpretive inference from corporate disclosures of the three firms and review of AI-and-accounting literature; described as a primary mechanism in the paper.
Meaningful human oversight of AI agents in knowledge work requires not improved post-hoc review mechanisms, but active participation in decisions as they are made.
Authors' conclusion drawn from the formative (N=8) and summative (N=16) studies and associated observations.
Users reported a sense of co-ownership over the resulting output.
Participant self-reports from the formative and/or summative studies (authors report users expressed co-ownership of outputs when participating in execution).
Users detected errors that post-hoc review would have failed to surface.
Empirical observation reported from the studies (authors report that active participation allowed users to detect errors that would be missed by post-hoc review).
Users identified their own intent reflected in the agent's actions.
Reported participant observations/self-reports from the formative (N=8) and/or summative (N=16) studies; claim presented as a finding of the evaluations.
A formative study (N = 8) and a within-subjects summative evaluation (N = 16) comparing Pista to a baseline agent demonstrated that active participation in execution influenced not only task outcomes but also users' comprehension of the task, their perception of the agent, and their sense of role within the workflow.
Empirical evaluation consisting of a formative study with N=8 and a within-subjects summative evaluation with N=16 comparing Pista to a baseline agent (authors report influence on task outcomes, comprehension, perception, and role).
We introduce Pista, a spreadsheet AI agent that decomposes execution into auditable, controllable actions, providing users with visibility into the agent's decision-making process and the capacity to intervene at each step.
System description / design contribution presented by the authors (implementation description rather than empirical evidence).
Effective governance of AI as a dual-use technology will likely require a multilateral institutional architecture functionally analogous (though not identical) to the role performed by the IAEA in the nuclear domain, with explicit safeguards against co-option of hardware controls for domestic repression.
Normative institutional design argument and analogy to the IAEA presented in the paper (policy proposal; comparative institutional analysis).
Hardware-layer governance, including chip-level attestation mechanisms such as FlexHEG, trusted execution environments, confidential computing, and complementary software-layer safeguards, offers a defense-in-depth alternative to the current binary framing of openness vs restriction.
Proposed governance architecture and technical discussion in the paper citing concrete mechanisms (technical-proposal and conceptual analysis; no experimental or deployment data reported in the summary).
The global concentration of compute infrastructure makes open-weight models one of the most viable pathways to sovereign AI capacity in the Global South.
Analysis of global compute infrastructure concentration and pathway mapping in the paper (conceptual/structural analysis; no numerical sample provided in the summary).
Selective forgetting should be considered a fundamental capability for next-generation LLM agents operating in real-world, resource-constrained scenarios.
Conclusion/argument in paper based on conceptual analysis and reported empirical benefits.
The work bridges cognitive neuroscience (hippocampal indexing/consolidation theory and Ebbinghaus forgetting curve) and AI systems to inform forgetting mechanisms.
Claimed theoretical grounding and cross-disciplinary framing in paper (stated in abstract).
Empirical results show security performance with 100% elimination of security risks.
Reported experimental result in abstract claiming full elimination of security risks.
Empirical results show content quality improved by +29.2% signal-to-noise ratio.
Reported experimental result in abstract (signal-to-noise ratio improvement).
Empirical results show access efficiency improved by +8.49%.
Reported experimental result in abstract.
Building on advances in LLM agent architectures and vector databases, the paper presents detailed specifications, implementation strategies, and empirical validation from controlled experiments.
Methodological claim in abstract indicating implementation and controlled experiments (no experimental details in abstract).
Selective forgetting improves security through active forgetting of malicious inputs, sensitive data, and privacy-compromising content.
Authors' taxonomy and safety-triggered forgetting mechanism; abstract reports empirical security performance (100% elimination of security risks).
Selective forgetting improves content quality by dynamically updating outdated preferences and context.
Conceptual claim supported by authors' implementation and empirical validation; abstract reports content quality improvement (signal-to-noise ratio).
A well-designed forgetting mechanism improves efficiency via intelligent memory pruning.
Claim supported by authors' framework and controlled experiments reported in the paper (abstract references empirical results for access efficiency).
In resource-constrained environments, a well-designed forgetting mechanism is as crucial as remembering.
Argument and conceptual analysis in paper; motivated by theoretical considerations and (claimed) empirical validation.
The findings point to a staged progression of AI utility from low-consequence assistance toward higher-order automation, as trust, infrastructure, and verification mature.
Synthesis of interview responses (over 30) indicating current use cases are lower-risk assistance and that stakeholders expect (or prefer) gradual progression toward automation contingent on trust/infrastructure/verification improvements.
Reliability, verification, and auditability are central requirements for adoption, driving human-in-the-loop frameworks and governance aligned with existing engineering reviews.
Consistent themes from interviews (over 30) indicating stakeholders prioritize reliability, verifiability, and audit trails, leading to preference for human-in-the-loop designs integrated with current review processes.
Higher-value agentic gains come from orchestrating multi-step workflows across tools.
Observed and reported in interviews (over 30) with stakeholders in engineering and manufacturing workflows describing value from agentic orchestration across tools.
Near-term AI gains cluster around structured, repetitive work and data-intensive synthesis.
Qualitative findings from an exploratory state-of-practice study based on over 30 semi-structured interviews across four stakeholder groups (large enterprises, small/medium firms, AI developers, and CAD/CAM/CAE vendors).
‘Smarter’ AI agents are more profitable.
Measured profits earned by agents of different capability levels in the trading experiment and observed higher profits for higher-capability ('smarter') agents.
‘Smarter’ AI agents perform better at information aggregation.
Experimental comparison of AI agents with different capability levels ('smarter' vs. less smart) in the trading experiment; measured aggregation via log error of last price and found better performance for higher-capability agents.
Prediction markets are robust to cheap talk, market duration, initial price, and strategic prompting.
Synthesis of experimental results showing no change in aggregation performance across manipulations (cheap talk, duration, initial price, strategic prompting).
The median market is effective at aggregating information in the easy information structures.
Controlled laboratory experiment in which AI agents traded in prediction markets after receiving private signals; information aggregation measured by the log error of the last price; comparison across 'easy' information structures using median-market outcomes.
Because misalignment can occur along each axis -- and affect stakeholders differently -- alignment cannot be 'solved' through technical design alone, but must be managed through ongoing institutional processes that determine how objectives are set, how systems are evaluated, and how affected communities can contest or reshape those decisions.
Normative conclusion drawn from the three-axis framework and discussion of stakeholder impacts (conceptual policy prescription; no empirical testing reported).
Alignment is inherently pluralistic and context-dependent, and resolving misalignment involves trade-offs among competing values.
Theoretical and normative argument in the paper about pluralism and context-dependence of values (conceptual discussion; no empirical quantification).
The three-axis decomposition implies that alignment is fundamentally a problem of governance rather than engineering alone.
Logical inference from the proposed decomposition and normative argument in the paper (theoretical reasoning; no empirical evidence).
The three-axis framework provides a systematic way of diagnosing why misalignment arises in real-world systems and clarifies that alignment cannot be treated as a single technical property of models but an outcome shaped by how objectives are specified, how information is distributed, and whose interests count in practice.
Conceptual argument and analytic claim about the explanatory utility of the proposed framework (theoretical demonstration; no empirical tests reported).
Misalignment can be reconceptualised as arising along three interacting axes: objectives, information, and principals (drawing on the principal–agent framework).
Theoretical framing using the principal–agent framework; conceptual decomposition proposed in the paper (no empirical validation reported).
The alignment problem is better understood as a structural question about governance: not whether an AI system is aligned in the abstract, but whether it is aligned enough, for whom, and at what cost.
Normative and conceptual argument presented by the author proposing a governance-focused reconceptualization (theoretical analysis; no empirical data).
SWE-chat is a living dataset; our collection pipeline automatically and continually discovers and processes sessions from public repositories.
Description of the dataset collection infrastructure and pipeline provided in the paper; operational behavior asserted by authors.
The dataset currently contains 6,000 sessions, comprising more than 63,000 user prompts and 355,000 agent tool calls.
Descriptive statistics reported by the authors based on their dataset collection pipeline (dataset metadata).
We present SWE-chat, the first large-scale dataset of real coding agent sessions collected from open-source developers in the wild.
Paper authorship / dataset description; dataset curated and presented by the paper as a contribution. No external validation provided in excerpt.
Statelessness is the load-bearing property explaining enterprises' preference for weaker but replayable retrieval pipelines, and DPM demonstrates this property is attainable without the decisioning penalty retrieval pays.
Synthesis/conclusion based on theoretical argument and empirical results presented (architectural analysis + experiments showing DPM performance and auditability).
The audit surface follows the same one-versus-N pattern: DPM logs two LLM calls per decision while summarization logs 83-97 on LongHorizon-Bench.
Empirical measurement on LongHorizon-Bench reported in the paper: logged LLM calls per decision are 2 for DPM vs 83-97 for summarization.
DPM is additionally 7-15x faster at binding budgets, making one LLM call at decision time instead of N.
Empirical runtime/efficiency measurement reported in the paper (range 7-15x speedup) comparing number of LLM calls and latency under tight memory budgets.
At a 20x compression ratio, DPM improves reasoning coherence by +0.53 (Cohen's h=1.13, p=0.0034) compared to summarization-based memory (paired permutation, n=10).
Paired permutation test over 10 cases at a 20x compression ratio; reported effect +0.53 with Cohen's h=1.13 and p=0.0034.
At a 20x compression ratio, DPM improves factual precision by +0.52 (Cohen's h=1.17, p=0.0014) compared to summarization-based memory (paired permutation, n=10).
Paired permutation test over 10 cases at a 20x compression ratio; reported effect +0.52 with Cohen's h=1.17 and p=0.0014.
On ten regulated decisioning cases at three memory budgets, DPM matches summarization-based memory at generous budgets and substantially outperforms it when the budget binds.
Empirical evaluation on 10 decisioning cases across three memory budgets; comparison between DPM and summarization-based memory as reported in the paper (n=10).
We propose Deterministic Projection Memory (DPM): an append-only event log plus one task-conditioned projection at decision time.
Method/architectural proposal described in the paper.
Presumptuousness in legal AI is systematic but addressable, and addressing it is a necessary step towards systems that reliably support, rather than supplant, human judgment wherever decisions must await sufficient evidence.
Synthesis conclusion in paper based on the benchmark experiments, comparisons across prompting methods, and SPEC results.