AI evidence extraction

Sem-Detect: Semantic Level Detection of AI Generated Peer-Reviews

André V. Duarte, Brian Tufts, Aditya Oke, Fei Fang, Arlindo L. Oliveira, Lei Li · ArXiv.org · 2026

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

10/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Supervised machine learning classification.

Sample

N = 20165, 3 groups

Primary method

Gradient-boosted decision trees using LightGBM framework. Hyperparameter selection via randomized search with five-fold stratified cross-validation. Macro-F1 used as optimization metric for balanced performance across three classes. Bootstrap resampling (1,000 iterations) for uncertainty estimation. ROC curves, AUC, TPR at fixed FPR thresholds (0.1% and 1%), and macro F1 reported as evaluation metrics. Cosine similarity computed for claim embeddings.

Main result

The study found that "Sem-Detect improves over the strongest prior detector by 25.5% in TPR@0.1% FPR in binary detection, with fewer than 3.5% of LLM-refined human reviews misclassified as AI-generated." Additionally, "when reviewing the same paper, AI-generated reviews exhibit higher claim-level overlap with one another than human-written reviews, including those refined using LLMs."

Reports effect sizes and confidence intervals.

Research paradigm

Positivist/Empiricist

Author conclusions

The authors conclude: "By distinguishing fully AI-generated reviews from those where humans used an LLM only to improve clarity, our framework supports policies that can detect problematic content without penalizing responsible AI assistance." They further state that "Sem-Detect also shows reasonable behavior under distribution shift. The method generalizes to unseen models, transfers to medical imaging reviews without retraining, and produces plausible predictions on recent ICLR 2026 data. This shows that effective detection and fairness to legitimate LLM use can coexist."

Risk of bias

Class imbalance: LLM-refined reviews (12,332) are more prevalent than human reviews (3,065), which the authors acknowledge influences error patterns; Temporal bias: All human review baselines are from 2022 or earlier; evaluation on ICLR 2026 lacks ground truth; Model-specific bias: Training on only four LLM families (Gemini-2.5-Flash, Gemini-2.5-Pro, DeepSeek-V3.1, Qwen3-235B) may limit generalization; Domain bias: Evaluation primarily on ML/AI conference reviews; limited cross-domain validation (only one medical imaging venue tested); Selection bias in reference review pairing: Reference reviews selected must match evaluation score of target review; Conservative bias in classification: Model tends to classify uncertain samples as LLM-refined rather than fully AI-generated; Class imbalance in training data: LLM-refined reviews (12,332) far outnumber human reviews (3,065), potentially biasing the model toward predicting LLM-refined; Selection bias: Only papers from ICLR and NeurIPS conferences used; limited to pre-2022 human reviews; Generation bias: AI reviews generated using only four specific LLM models; may not represent all possible generation approaches; Conservative classification bias: Model tends to classify ambiguous cases as LLM-refined rather than fully AI-generated, introducing systematic conservative bias; Cross-domain limitation: Medical imaging evaluation (MIDL) still focuses on deep learning, limiting generalizability assessment; Temporal bias: ICLR 2026 evaluation uses different detection method (EditLens) for comparison, making direct validation difficult; Class imbalance in training data (LLM-refined reviews more prevalent than other classes); Temporal bias: human reviews collected only up to 2022, potentially not representative of current review practices; Venue selection bias: limited to computer science conferences (ICLR, NeurIPS, MIDL); Model selection bias: training on only four specific LLM families; Selection bias in paper sampling: 200 papers randomly selected per conference per year

Limitations

The authors state: "Our method assumes that the originality of ideas can help distinguish human from AI authorship
As models continue to improve, they may eventually produce reviews with novel, high-quality insights that are indistinguishable from, or even better than, those of human experts
If that happens, the line between human and AI authorship may blur." Additionally, "MIDL, while medically oriented, still centers on deep learning methods
Evaluating on more distant fields would be ideal, but open peer-review data remains limited outside of computer science." They also note that "any detection system risks false accusations, which can harm reviewers' reputations."

Open questions raised

Evaluation on more distant fields beyond deep learning and computer science (authors note open peer-review data remains limited)
Study of how improved AI models may produce reviews with novel insights indistinguishable from human experts
More advanced verification pipelines for validating factual claims using external document sources
Investigation of whether broader generator pools during training could improve robustness when combined with precision-preserving mechanisms
Need for evaluation on more distant academic fields beyond computer science (limited by availability of open peer-review data)
Future work could explore more advanced verification pipelines for factual accuracy assessment, leveraging external document sources to validate factual claims

Data: Complete set of reviews for 800 papers from ICLR and NeurIPS 2021-2022 (stated to be released by authors); MIDL 2022 medical imaging venue reviews (approximately 100 papers sampled for cross-domain evaluation); ICLR 2026 reviews (approximately 600 papers sampled for temporal evaluation); Complete set of reviews for 800 papers from ICLR and NeurIPS 2021-2022 (human-written, AI-generated, and LLM-refined variants) - to be released by authors; Medical imaging venue data from MIDL 2022 (approximately 100 random papers); ICLR 2026 reviews (approximately 600 papers sampled randomly); Over 20,000 peer reviews from ICLR and NeurIPS (2021-2022); Medical imaging venue reviews (MIDL 2022); ICLR 2026 submissions (approximately 600 papers); Authors state: "we release the following artifacts: Code: full pipeline, two pre-trained classifiers, and a self-hosted Flask web demo. Data: complete set of reviews for the 800 papers from ICLR and NeurIPS 2021-2022."Code: Full pipeline and pre-trained classifiers to be released; Self-hosted Flask web demo to be released; GitHub repository not explicitly provided in paper but authors state they will release code; Full pipeline code and pre-trained classifiers to be released; Full pipeline code; Pre-trained classifiers; Self-hosted Flask web demoExtracted from: pdfAgreement 56%

Explore related topics

Authorship & Attribution Peer Review Empirical Study