Stereotype Detection in LLM-Generated Text in Portuguese

• Fernanda Malheiros Assi UFSCar
• Helena de Medeiros Caseli UFSCar

DOI: https://doi.org/10.5753/eniac.2026.30810

Resumo

The widespread adoption of large language models (LLMs) has intensified concerns about the generation of socially biased content, particularly in the form of stereotypes. However, there is still limited work focusing on stereotype detection in languages other than English, such as Portuguese. This work presents a human-annotated dataset in Brazilian Portuguese for detecting stereotype alignment in sentence completions output by LLMs. We design a set of 164 templates in a neutral form, which are later instantiated with social markers representing gender, race, and their intersectional combinations, which are used to generate sentence completions from 37 LLMs. A subset of the resulting sentences was annotated by humans to capture the degree of alignment with socially recognized stereotypes. Using this dataset, we trained a supervised classification model that achieves a macro F1-score of 0.68 in distinguishing between stereotype-aligned and non-aligned content, a result consistent with the difficulty and subjectivity of the task, as reflected by the inter-annotator disagreement observed during annotation.

Referências

Assi, F. M. and Caseli, H. (2024). Biases in GPT-3.5 turbo model: a case study regarding gender and language. In Claro, D. B. and Pagano, A., editors, Proceedings of the 15th Brazilian Symposium in Information and Human Language Technology, pages 32–43, Belém do Pará, Brazil. ACL.

Fleiss, J. L. (1971). Measuring nominal scale agreement among many raters. Psychological Bulletin, 76(5):378–382.

Gemini Team (2023). Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.

Martinková, S., Stanczak, K., and Augenstein, I. (2023). Measuring gender bias in West Slavic language models. In Piskorski, J., Marcińczuk, M., Nakov, P., Ogrodniczuk, M., Pollak, S., Přibáň, P., Rybak, P., Steinberger, J., and Yangarber, R., editors, Proceedings of the 9th Workshop on Slavic Natural Language Processing 2023 (SlavicNLP 2023), pages 146–154, Dubrovnik, Croatia. ACL.

Melo, J. L. L. d. and Souza, M. (2026). Levados em consideração: Uma avaliação de vieses de estima por raça, gênero e região em grandes modelos de linguagem em português brasileiro. In Souza, M., de Dios-Flores, I., Santos, D., Freitas, L., Souza, J. W. d. C., and Ribeiro, E., editors, Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1, pages 516–528, Salvador, Brazil. ACL.

Nozza, D., Bianchi, F., and Hovy, D. (2021). HONEST: Measuring hurtful sentence completion in language models. In Toutanova, K., Rumshisky, A., Zettlemoyer, L., Hakkani-Tur, D., Beltagy, I., Bethard, S., Cotterell, R., Chakraborty, T., and Zhou, Y., editors, Proceedings of the NAACL-HLT 2021, pages 2398–2406, Online. ACL.

Puttick, A., Ikae, C., Rigotti, C., Fosch-Villaronga, E., Kharas, M. W., Søraa, R. A., and Kurpicz-Briki, M. (2025). A systematic review of bias detection methods for non-english word embeddings and language models. Artificial Intelligence Review, 58(12):370.

Qi, J. (2024). The impact of large language models on social media communication. In Proceedings of the 2024 7th International Conference on Software Engineering and Information Management, ICSIM ’24, page 165–170, New York, NY, USA. Association for Computing Machinery.

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI Technical Report.

Roumeliotis, K. I., Tselikas, N. D., and Nasiopoulos, D. K. (2024). Llms in e-commerce: A comparative analysis of gpt and llama models in product review evaluation. Natural Language Processing Journal, 6:100056.

Sena, R., Souza, M., Santana, A., and Melo, J. (2025). Me deixe pensar sobre isso! uma análise do uso de cot para identificar vieses nas respostas de llm para o português brasileiro. In Anais do VI Workshop sobre as Implicações da Computação na Sociedade, pages 105–120, Porto Alegre, RS, Brasil. SBC.

Sheng, E., Chang, K.-W., Natarajan, P., and Peng, N. (2019). The woman worked as a babysitter: On biases in language generation. In Inui, K., Jiang, J., Ng, V., and Wan, X., editors, Proceedings of the EMNLP-IJCNLP 2019, pages 3407–3412, Hong Kong, China. ACL.

Shin, J., Song, H., Lee, H., Jeong, S., and Park, J. (2024). Ask LLMs directly, “what shapes your bias?”: Measuring social bias in large language models. In Ku, L.-W., Martins, A., and Srikumar, V., editors, Findings of the Association for Computational Linguistics: ACL 2024, pages 16122–16143, Bangkok, Thailand. ACL.

Shrawgi, H., Rath, P., Singhal, T., and Dandapat, S. (2024). Uncovering stereotypes in large language models: A task complexity-based approach. In Graham, Y. and Purver, M., editors, Proceedings of the 18th EACL (Volume 1: Long Papers), pages 1841–1857, St. Julian’s, Malta. ACL.

Souza, F., Nogueira, R., and Lotufo, R. (2020). Bertimbau: Pretrained bert models for brazilian portuguese. In Intelligent Systems: 9th Brazilian Conference, BRACIS 2020, Rio Grande, Brazil, October 20–23, 2020, Proceedings, Part I, page 403–417, Berlin, Heidelberg. Springer-Verlag.

Taso, F., Reis, V., and Martinez, F. (2023). Discriminação algorítmica de gênero: Estudo de caso e análise no contexto brasileiro. In Anais do IV Workshop sobre as Implicações da Computação na Sociedade, pages 13–25, Porto Alegre, RS, Brasil. SBC.

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., and Lample, G. (2023). LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.

Zhang, Q., Naradowsky, J., and Miyao, Y. (2023). Ask an expert: Leveraging language models to improve strategic reasoning in goal-oriented dialogue models. In Rogers, A., Boyd-Graber, J., and Okazaki, N., editors, Findings of the Association for Computational Linguistics: ACL 2023, pages 6665–6694, Toronto, Canada. ACL.