PREVENTING FALSE CATEGORICAL CONCLUSIONS WHEN USING LARGE LANGUAGE MODELS IN FORENSIC EXPERT PRACTICE
DOI:
https://doi.org/10.32782/forensic.science.2026.4.11Keywords:
large language model, categorical conclusion, verification, professional error, forensic examination, quality controlAbstract
The use of large language models in forensic expert practice creates a specific reliability risk: a plausible but unverified model output may be inherited by subsequent stages of an examination and influence a categorical expert conclusion. An error may arise during source selection and interpretation, content generation, revision, verification, technical transformation, or release of the final digital document. The problem is therefore not limited to isolated hallucinations or incorrect answers; it concerns the conditions under which unverified model-derived information may acquire professional significance. The purpose of this article is to substantiate a process-based approach to preventing false categorical conclusions when large language models are used in forensic expert activity. The study combines qualitative synthesis of practical experience with complex analytical and document-oriented tasks, retrospective analysis of recurrent operational errors and control failures, process modelling, and comparison with research published in 2025–2026. Professional work with an LLM is treated as a sequence of professionally significant states. Transition between states is permitted only when predefined criteria are confirmed. Machine traceability and professional verification are distinguished: execution logs provide a basis for reconstructing the trajectory and diagnosing failures, whereas a control point determines whether a result may influence the next stage of the expert examination. Six practically significant risk classes are identified: reliance on an unsupported object or source; incorrect application of valid information; contamination between sources or contexts; loss of previously verified content during revision; a false claim that verification has been completed; and inconsistency between verified content and the digital object actually released. The article argues that an LLM output cannot serve as an independent basis for a categorical expert conclusion. If information obtained or transformed using an LLM is materially significant to the answer and cannot be properly verified independently of the model, it should not enter the logical basis of a categorical conclusion. The proposed process is not intended to guarantee an error-free language model. Its purpose is to prevent unverified model-derived information from being promoted to a categorical forensic conclusion without an adequate evidentiary and methodological basis.
References
1. Про судову експертизу : Закон України від 25.02.1994 р. № 4038–XII. База даних «Законодавство України» / Верховна Рада України. URL: https://zakon.rada.gov.ua/go/4038-12 (дата звернення: 09.09.2026).
2. Про затвердження Інструкції про призначення та проведення судових експертиз та експертних досліджень та Науково-методичних рекомендацій з питань підготовки та призначення судових експертиз та експертних досліджень : наказ Міністерства юстиції України від 08.10.1998 р. № 53/5. База даних «Законодавство України». URL: https://zakon.rada.gov.ua/go/z0705-98 (дата звернення: 09.09.2026).
3. Кривоножко Г., Собін О. Штучний інтелект і спеціалізовані цифрові інструменти в експертному забезпеченні розслідування кіберзлочинів та цифрових правопорушень. Криміналістика майбутнього: цифрові інновації, штучний інтелект, глобальні виклики та загрози : збірник матеріалів Міжнародної науково-практичної конференції, м. Одеса, 5 червня 2026 р. / заг. ред. С. Ківалова, М. Аракеляна. Одеса : НУ «ОЮА», 2026. С. 184–188. DOI: https://doi.org/10.32837/11300.33533
4. Свінцицький А. Функціонал замовлення судової експертизи за допомогою інтегрованого чат-бота ChatGPT. Судова експертиза. 2025. № 2. С. 4–9. DOI: https://doi.org/10.32782/forensic.science.2025.2.1
5. Chernyshev M., Baig Z., Syed N., Doss R., Shore M. Large language models in digital forensics: capabilities, challenges and future directions. Forensic Science International: Digital Investigation. 2026. Vol. 56. Article 302043. DOI: https://doi.org/10.1016/j.fsidi.2025.302043
6. Metropolitansky D., Larson J. VeriTrail: Closed-Domain Hallucination Detection with Traceability. International Conference on Learning Representations (ICLR 2026). 2026. URL: https://proceedings.iclr.cc/paper_files/paper/2026/hash/42ec51d1031d25eca238ccc67559fe84-Abstract-Conference.html (дата звернення: 09.09.2026).
7. Zhang S. et al. Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems. Proceedings of the 42nd International Conference on Machine Learning. PMLR. 2025. Vol. 267. P. 76583–76599. URL: https://proceedings.mlr.press/v267/zhang25cq.html (дата звернення: 09.09.2026).
8. Rosset C., Sharma P., Zhao A., González-Fernández M., Awadallah A. The Art of Building Verifiers for Computer Use Agents. Conference on Language Modeling (COLM 2026). 2026. URL: https://www.microsoft.com/en-us/research/publication/the-art-of-building-verifiers-for-computer-use-agents/ (дата звернення: 09.09.2026).
9. Kim E. M., Garg A., Peng K., Garg N. Correlated Errors in Large Language Models. Proceedings of the 42nd International Conference on Machine Learning. PMLR. 2025. Vol. 267. P. 30038–30066. URL: https://proceedings.mlr.press/v267/kim25e.html (дата звернення: 09.09.2026).
10. Laban P., Schnabel T., Neville J. LLMs Corrupt Your Documents When You Delegate. arXiv:2604.15597. 2026. DOI: https://doi.org/10.48550/arXiv.2604.15597
11. Srinivasan A., Shamdasani S., Araujo A. F., de Wasseige J. Building self-improving tax agents with Codex. OpenAI. 27.05.2026. URL: https://openai.com/index/building-self-improving-tax-agents-with-codex/ (дата звернення: 09.09.2026).
12. National Institute of Standards and Technology. The TEVV-Athlon Framework for Evaluating AI Systems. Initial Public Draft NIST AI 200-2. 07.08.2026. DOI: https://doi.org/10.6028/NIST.AI.200-2.ipd



