AI evidence extraction

TemporalXAI-Det: Temporal-Aware Explainable Detection of Multi-Model AI-Generated Academic Text via Continual Learning and Cross-Lingual Transfer

Imeldawaty Gultom, Ratih Puspadini, Fauzi Erwis, Elyandri Prasiwiningrum, Ridwan · JOURNAL OF ICT APLICATIONS AND SYSTEM · 2026

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

9/10

Relevance

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.56313/jictas.v5i1.531

Methodology & findings

Study design

Empirical evaluation study using a multi-stage framework with six integrated modules: (1) stylometric feature extraction, (2) cross-lingual semantic encoding, (3) hybrid feature fusion, (4) deep learning classification, (5) continual learning adaptation, and (6) explainability suite.

Sample

N = 72000, 7 groups

Primary method

McNemar's test (α = 0.05) with Bonferroni correction for pairwise statistical significance assessment. Stratified sampling for train/validation/test partitioning. AdamW optimizer with cosine annealing for model training (initial learning rate 3×10⁻⁵, batch size 64, 60 epochs, early stopping patience 10). Focal loss (γ = 2) for class-level imbalance handling within adversarial subsets. k-means clustering for experience replay buffer selection. Fisher information matrix for Elastic Weight Consolidation (EWC) regularization coefficient λ = 4,000.

Main result

TemporalXAI-Det achieves "97.2% accuracy and a macro F1 of 0.941" on clean test sets and demonstrates "a 78.4% reduction in forgetting achieved by EWC+Replay relative to standard fine-tuning" in continual learning scenarios. Under adversarial attacks, the system "exhibits a mean performance degradation of Δ = 2.9 pp across all attack conditions, compared to a mean of 24.6 pp across baselines," and the "LAPT mechanism achieves a mean adversarial macro F1 of 0.887 across eleven non-English languages using only 5% language-specific parameters."

Reports effect sizes and confidence intervals.

Research paradigm

Empiricist (computational experiments with quantitative measurement)

Author conclusions

The authors conclude that "meaningful LLM-family attribution is achievable at high accuracy (97.2% clean, 94.1% adversarial) even under sophisticated paraphrasing attacks" and that "the continual learning results demonstrate, for the first time in the AI-text detection literature, that catastrophic forgetting represents a quantifiable and addressable threat to deployed detection systems." They further note that "the LAPT mechanism achieves a mean adversarial macro F1 of 0.887 across eleven non-English languages using only 5% language-specific parameters," carrying "substantial equity implications" for global deployment.

Risk of bias

Potential selection bias in human-authored text sources (ArXiv, SSRN, PubMed, ERIC) which may not represent all academic disciplines equally; model selection bias toward five major LLM families; potential translation bias in multilingual extension; interannotator agreement at κ = 0.84 suggests moderate but not perfect annotation reliability.; Potential selection bias: AI text generation controlled via structured prompts derived from human texts; Temporal bias: Evaluation simulated temporal drift using predetermined phases rather than actual real-world model releases; Commercial detector evaluation: GPTZero and Turnitin APIs evaluated in March 2026; performance may reflect API-specific configurations; Class imbalance handling: Focal loss used to address imbalance within adversarial subsets, but equal class sizes may not reflect real-world distribution

Open questions raised

The authors identify that binary human-vs-AI classifiers are inadequate for multi-source detection
temporal model drift has been neglected by existing literature
multilingual detection infrastructure is absent for non-English languages
and future work on LLM attribution should focus on sub-lexical stylometric profiling rather than purely semantic approaches.
Authors identify three compounding inadequacies in prior work:
model multiplicity—binary classifiers fail across multiple LLM families with cross-model generalization rates below 60%

Data: MTA-72K: a multilingual temporal adversarial benchmark corpus comprising 72,000 samples, released under CC BY 4.0 license. The corpus includes an English partition (60,000 samples) and a multilingual partition (MTA-72K-ML) with translations into eleven additional languages: Bahasa Indonesia, Arabic (Modern Standard), Mandarin Chinese, Spanish, French, Portuguese (Brazilian), German, Japanese, Korean, Hindi, and Swahili.; 60,000 English partition and 6,000 multilingual partition across twelve languages. Released under CC BY 4.0. English partition: 42,000 training, 9,000 validation, 9,000 test samples. Comprises six balanced source classes with four adversarial attack variants applied to AI-generated classes.; MTA-72K: "the first large-scale multilingual temporal adversarial benchmark corpus for academic AI-text detection, released under CC BY 4.0." The corpus comprises 72,000 samples (60,000 English partition for training/evaluation; 6,000 multilingual partition for cross-lingual transfer). Specific download URL or repository location not provided in the paper.Extracted from: pdf

Explore related topics

Hallucination Control Research Integrity Benchmarking