AI evidence extraction

The Attention Mismatch: Mapping the Structural Academic Governance Deficit in the Age of Generative AI

Zhenning Guo, Haoran Mao · Publications · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.3390/publications14020027

Methodology & findings

Study design

Multi-scalar integrative framework combining three complementary approaches: (1) Retraction database analysis from Retraction Watch Database using frequency statistics, Mann-Whitney U tests, Kolmogorov-Smirnov tests, and Cohen's d effect size calculations; (2) Large-scale corpus analysis of Common Crawl web archives (2013-2026) using ensemble machine learning with TF-IDF features and logistic regression classifiers to detect AI-likeness in text; (3) Bibliometric analysis of governance literature from Web of Science Core Collection, Scopus, and Google Scholar (2020-2026) using keyword co-occurrence network analysis via VOSviewer, burst detection using Kleinberg's algorithm, and creation of a Normalized Coverage Index (NCI) to quantify disciplinary misalignment between governance attention and retraction pressure..

Sample

> 1000, 4 groups

Primary method

Gaussian kernel density estimation (Gaussian KDE, bandwidth = 0.3) for visualizing retraction lag distributions; Mann-Whitney U test (nonparametric) for comparing median retraction lag differences; Kolmogorov-Smirnov (KS) test for comparing overall distribution functions; Cohen's d calculation for effect size measurement; Five-fold cross-validation for model performance evaluation; ROC-AUC, accuracy, precision, recall, and F1 score metrics for ensemble model assessment; TF-IDF logistic regression (word-level and character-level) with class weight balancing; Ensemble learning with weighted average (0.6 word-level, 0.4 character-level) probability combination; VOSviewer co-occurrence network analysis and clustering with minimum threshold ≥5; Kleinberg's burst detection algorithm with hidden-state automaton (state transition base s=2.0, penalty coefficient γ=1.0); Bootstrap resampling (2000 iterations) for NCI confidence interval estimation; Rule-based keyword matching algorithm for cross-dataset subject harmonization; Cohen's κ for classification reliability assessment; Spearman's ρ for sensitivity analysis robustness; Python 3.11 for all data processing and modeling

Main result

The study reveals a systematic mismatch between governance research attention and empirically observed misconduct pressure. Specifically, "Since 2022, retractions related to AIGC-generated content have exhibited a sudden surge, with a growth rate significantly exceeds the historical average and clearly diverges from the patterns observed in traditional retraction categories." Furthermore, "the mean retraction lag for AI-related papers is 2.77 years, with a median of 1.62 years, both of which are higher than those observed for traditional misconduct (mean = 1.91 years; median = 0.24 years)." Most critically, "disciplines with the lowest NCIs—namely Chemistry; Physics, Mathematics & Statistics; Economics, Business & Management; Environmental, Earth & Agricultural Sciences; and Life Sciences & Biology—represent the most critical governance gaps."

Reports effect sizes and confidence intervals.

Research paradigm

Positivist/empiricist with mixed methods (quantitative bibliometric and corpus analysis combined with qualitative thematic analysis)

Author conclusions

The authors conclude: "This study provides the first empirical demonstration that the governance of AI-related academic misconduct is systematically misaligned with the actual distribution of retraction risk across disciplines. By integrating open web corpus analysis, retraction dynamics, and bibliometric evidence, we reveal a clear underrepresentation of governance research attention in the basic sciences relative to their risk exposure (NCI < 0.5), whereas fields such as education receive disproportionate attention (NCI > 29)." They further state: "Ultimately, safeguarding scientific knowledge in the age of generative AI requires not more generic ethics declarations, but a fundamental reallocation of our scarcest resources—scholarly attention. By diagnosing its current misalignment, this study provides an empirical basis to inform future efforts toward this critical task."

Risk of bias

Detection bias: AI-likeness model may misclassify text due to ongoing model evolution; Selection bias: Retraction Watch Database captures only detected and reported cases, not all misconduct; Geographic bias: Absolute retraction counts are not normalized by publication volume; countries with high research output (e.g., China) show higher absolute numbers; Disciplinary bias: Fields with weaker post-publication scrutiny may exhibit systematically lower observed retraction rates; Detection infrastructure circularity: Observed misalignment between governance and retraction is partly shaped by the detection infrastructure that governance research aims to strengthen; Visibility bias: ChatGPT discussions concentrated in educational domains creates path dependency in governance literature; Sampling bias: Common Crawl provides non-random web sample with heterogeneous domain and structural characteristics; Detection bias: AIGC detection model trained on existing corpora may misclassify as AI technologies evolve; Selection bias: Common Crawl sample is limited relative to full web corpus; heterogeneous web content may affect generalizability; Reporting bias: Retraction data reflects detection capacity and reporting mechanisms that vary by discipline, not true misconduct rates; Geographic/disciplinary concentration bias: Absolute retraction counts not normalized by publication volume; China's high absolute numbers reflect high research output rather than necessarily higher misconduct rates; Circular dependency: Observed retraction patterns are shaped by detection infrastructure that governance research aims to strengthen; Post-publication scrutiny bias: Weak post-publication scrutiny in certain experimental or foundational sciences may lead to systematically lower observed retraction rates; Publication lag bias: Governance research lags behind technological development; some emerging discussions may not yet be indexed in databases; Classification bias in AI-likeness detection model due to ongoing evolution of AI systems; Detection bias in retraction data reflecting disciplinary differences in detection capacity and reporting mechanisms; Geographic concentration bias in retraction data (China accounts for 11,952 of retractions) influenced by publication volume rather than misconduct rates; Sampling bias in Common Crawl corpus analysis despite standardized random selection procedures; Path dependency bias in governance discourse concentrated in education and social science domains; Publication bias in governance literature potentially favoring visible issues over methodological challenges in basic sciences

Limitations

The authors state: "First, with regard to the detection of AIGC, the 'AI-likeness' model employed in this study is trained on existing corpora and may be influenced by ongoing model evolution, thereby introducing a risk of misclassification
Accordingly, the results should be interpreted as indicative of trends rather than precise estimates of the true proportion of generated [content]." Additionally, "retraction data, used here as a proxy for misconduct, may be affected by disciplinary differences in detection capacity, reporting mechanisms, and temporal windows, thereby introducing bias
As such, the analysis reflects 'observable risk' rather than the full extent of misconduct." Furthermore, "the Common Crawl sample, while substantial, remains limited relative to the full scale of the underlying corpus, and the heterogeneity of web content in terms of domain and structure may affect the generalizability of the findings."

Open questions raised

Need for causal analyses beyond association-based findings on AI technology integration and misconduct detection
Lack of proactive detection mechanisms and continuous monitoring systems for AI-assisted misconduct
Insufficient research on hallucinated data and AI-assisted fabrication in experimentally intensive disciplines
Limited qualitative analysis of disciplinary cultures and publication practices affecting detection
Need for cross-national comparative studies examining variations in governance models across research ecosystems
Absence of longitudinal tracking of governance mechanism evolution across disciplines

Data: Retraction Watch Database (proprietary, access required); Common Crawl web archive (2013-2026) (publicly available at https://commoncrawl.org/); HC3 dataset (for AI-generated text samples); Web of Science Core Collection (subscription required); Scopus (subscription required); Google Scholar (publicly available); Retraction Watch Database (accessible online, not explicitly provided with paper); Common Crawl web archive (https://commoncrawl.org/, 2013-2026 snapshots analyzed); HC3 dataset (used for AI-generated text positive samples, source not specified in detail); Retraction Watch Database (publicly accessible); Common Crawl web archive (publicly accessible, data extracted as of 30 January 2026); HC3 dataset (used for positive samples in AI-likeness model training); Web of Science Core Collection (access-restricted academic database); Scopus (access-restricted academic database); Google Scholar (publicly accessible)Extracted from: pdfAgreement 56%

Explore related topics

Institutional Policy AI Governance Mixed Methods