The Surface Consistency Trap: Evaluating LLM-Generated Deepfake Explanations
DOI:
https://doi.org/10.54691/3x9yrf49Keywords:
Deepfake Detection; Large Language Models; Faithfulness Evaluation; Numerical Hallucination; Surface Consistency Trap.Abstract
Automatic metrics can reward surface agreement while missing factual and logical errors. We define the Surface Consistency Trap as the systematic overestimation of explanation quality when surface-level signal consistency is treated as evidence of deep logical correctness. We introduce the Dual-Layer CFG framework for evaluating LLM-generated deepfake explanations from structured forensic inputs. Layer 1 (Auto-CFG) checks surface consistency, whereas Layer 2 (Human-CFG) assesses D1 (Numerical Correctness), D2 (Threshold Judgment), and D3 (Causal Reasoning). In a retrospective audit of 5,000 legacy explanations, the mean graded CSF coverage score is 93.5%. On the matched N = 101 subset, the binary Auto-CFG signal-consistency pass rate is 91.1%, while both human raters assign a strict Human-CFG pass rate of 0.0% (95% CI, 0.0–3.7%). A separate API-based GLM-4 cross-system validation (N = 300) reaches 99.7% on D1 and D2 but only 1.0% on D3 and strict pass. A post-repair audit (N = 300) uses validated visual TFL evidence with a constrained local FLAN-T5-base+LoRA generator; it reaches 100.0% on D1/D2 and D3=n for all items because the schema explicitly abstains from unsupported mechanism claims. A human baseline (N = 45) achieves 100% strict pass. These distinct arms show that input validity, surface fidelity, and mechanism-level reasoning require separate evaluation.
Downloads
References
[1] Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., & Ortega-Garcia, J. (2020). Deepfakes and beyond: A survey of face manipulation and fake detection. Information Fusion, 64, 131–148. https://doi.org/10.1016/j.inffus.2020.06.014
[2] Chen, Y., Yan, Z., Cheng, G., Zhao, K., Lyu, S., & Wu, B. (2024). X²-DFD: A framework for explainable and extendable deepfake detection. arXiv preprint arXiv:2410.06126. https://doi.org/10.48550/arXiv.2410.06126
[3] Xu, Z., Zhang, X., Li, R., Tang, Z., Huang, Q., & Zhang, J. (2025). FakeShield: Explainable image forgery detection and localization via multi-modal large language models. In International Conference on Learning Representations. https://openreview.net/forum?id=pAQzEY7M03
[4] Zhang, Y., Colman, B., Guo, X., Shahriyari, A., & Bharaj, G. (2024). Common sense reasoning for deepfake detection. arXiv preprint arXiv:2402.00126. https://doi.org/10.48550/arXiv.2402.00126
[5] Kuckreja, K., Gupta, P., Khan, M. H., & Dhall, A. (2026). Pixels don't lie (but your detector might): Bootstrapping MLLM-as-a-judge for trustworthy deepfake detection and reasoning supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 25871–25881). https://openaccess.thecvf.com/content/CVPR2026/html/Kuckreja_Pixels_Dont_Lie_But_Your_Detector_Might_Bootstrapping_MLLM-as-a-Judge_for_CVPR_2026_paper.html
[6] Maynez, J., Narayan, S., Bohnet, B., & McDonald, R. (2020). On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 1906–1919). https://doi.org/10.18653/v1/2020.acl-main.173
[7] Wiegreffe, S., & Marasović, A. (2021). Teach me to explain: A review of datasets for explainable natural language processing. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 1. https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/698d51a19d8a121ce581499d7b701668-Abstract-round1.html
[8] Singh, A. K., & Strouse, D. J. (2024). Tokenization counts: The impact of tokenization on arithmetic in frontier LLMs. arXiv preprint arXiv:2402.14903. https://doi.org/10.48550/arXiv.2402.14903
[9] Wallace, E., Wang, Y., Li, S., Singh, S., & Gardner, M. (2019). Do NLP models know numbers? Probing numeracy in embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) (pp. 5307–5315). https://doi.org/10.18653/v1/D19-1534
[10] Cai, Z., Ghosh, S., Dhall, A., Gedeon, T., Stefanov, K., & Hayat, M. (2023). Glitch in the matrix: A large-scale benchmark for content-driven audio-visual forgery detection and localization. Computer Vision and Image Understanding, 236, 103818. https://doi.org/10.1016/j.cviu.2023.103818
[11] Carreira, J., & Zisserman, A. (2017). Quo vadis, action recognition? A new model and the Kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 6299–6308). https://doi.org/10.1109/CVPR.2017.502
[12] Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., & Paluri, M. (2018). A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 6450–6459). https://doi.org/10.1109/CVPR.2018.00675
[13] Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140), 1–67. https://jmlr.org/papers/v21/20-074.html
[14] Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S. S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., … Wei, J. (2024). Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70), 1–53. https://www.jmlr.org/papers/v25/23-0870.html
[15] Team GLM. (2024). ChatGLM: A family of large language models from GLM-130B to GLM-4 all tools. arXiv preprint arXiv:2406.12793. https://doi.org/10.48550/arXiv.2406.12793
[16] Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2022). LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations. https://openreview.net/forum?id=nZeVKeeFYf9
[17] Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., & Nießner, M. (2019). FaceForensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 1–11). https://doi.org/10.1109/ICCV.2019.00009
[18] Li, Y., Yang, X., Sun, P., Qi, H., & Lyu, S. (2020). Celeb-DF: A large-scale challenging dataset for DeepFake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 3207–3216). https://doi.org/10.1109/CVPR42600.2020.00327
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Scientific Journal of Intelligent Systems Research

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.




