Russian Federation
Russian Federation
Russian Federation
Voronezh, Russian Federation
The problem of bias in large language models (LLM), such as GPT, is one of the most relevant in the field of artificial intelligence. Biases inherited from unrepresentative and unfair training data can lead to discriminatory and harmful results that increase social inequality. The purpose of this article is to systematize and investigate the sources of bias, methods for detecting it, and ways to mitigate it. The paper formalizes the problem of displacement as a multi-criteria optimization problem, and also suggests a mathematical modeling methodology for evaluating the effectiveness of various methods of reducing displacement. The methodology considers approaches such as data preprocessing, fine tuning, and controlled output generation (RLHF). The simulation results demonstrate that the combined use of these methods can significantly reduce the level of bias, but does not eliminate it completely. The key conclusion is the need for a comprehensive, multi-step approach to managing bias throughout the LLM lifecycle, as well as the importance of developing better detection methods and ethical standards.
large language models, bias, AI ethics, machine learning, RLHF, justice
1. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? DOI: https://doi.org/10.1145/3442188.3445922
2. Bommasani, R., et al. (2022). On the Opportunities and Risks of Foundation Models.
3. Weidinger, L., et al. (2021). Ethical and social risks of harm from language models.
4. Cimmerling, A. V. Modal'nye operatory v avtomaticheskih sistemah perevoda i bol'shih yazykovyh modelyah / A. V. Cimmerling, A. M. Bayuk // Komp'yuternaya lingvistika i intellektual'nye tehnologii : Doklady po materialam ezhegodnoy mezhdunarodnoy konferencii "Dialog" (2025), Moskva, 23 aprelya 2025 goda. Tom 23. – Moskva, 2025. – S. 471-480. – EDN VHOYVM.
5. Mohammad, Zh. H. Izvlechenie klyuchevyh fraz na osnove bol'shih yazykovyh modeley / Zh. H. Mohammad // Izvestiya YuFU. Tehnicheskie nauki. – 2024. – № 5(241). – S. 143-151. – DOIhttps://doi.org/10.18522/2311-3103-2024-5-143-151. – EDN SFYQUK.
6. Czyyue, V. Optimizaciya i potencial razvitiya mashinnogo perevoda na primere bol'shih yazykovyh modeley tipa ChatGPT / V. Czyyue // Aktual'nye voprosy perevodovedeniya i regionovedeniya : sbornik statey Mezhdunarodnogo foruma, Nizhniy Novgorod, 20–22 oktyabrya 2023 goda. – Nizhniy Novgorod: Nizhegorodskiy gosudarstvennyy lingvisticheskiy universitet im. N.A. Dobrolyubova, 2024. – S. 67-75. – EDN VSPADK.
