Capacitismo Automatizado: Uma Auditoria de Modelos de Linguagem em Português nas Tarefas de Sentimento e Toxicidade

• Janaína N. S. Lopes UFMS https://orcid.org/0009-0000-2408-1018
• Vitória P. Firmino UFMS https://orcid.org/0009-0001-9267-5624
• Valéria Q. dos Reis UFMS / Leuphana Universität Lüneburg https://orcid.org/0000-0002-2581-6562
• Anderson C. de Lima UFMS https://orcid.org/0000-0003-4826-688X
• Bruno M. Nogueira UFMS https://orcid.org/0000-0002-4703-1379

DOI: https://doi.org/10.5753/wics.2026.23048

Resumo

Este trabalho investiga a presença de viés capacitista em modelos de Processamento de Linguagem Natural para o português brasileiro. Propomos um conjunto de auditoria, chamado BITS-PTBR, composto por sentenças com termos relacionados à deficiência e versões neutras equivalentes. Avaliamos modelos em tarefas de análise de sentimento e detecção de toxicidade, comparando respostas entre pares de frases que diferem apenas pela presença de marcadores de deficiência. Os resultados indicam que alguns modelos tendem a atribuir avaliações mais negativas ou tóxicas a menções de deficiência mesmo em contextos neutros, sugerindo evidências de capacitismo automatizado e a necessidade de auditorias éticas em sistemas de IA para o português.

Referências

Bariffi, F. J. (2021). Artificial intelligence, human rights and disability. Pensar: Revista de Ciências Jurídicas, 26(2).

Barocas, S. and Selbst, A. D. (2016). Big data’s disparate impact. California Law Review, 104(3):671–732.

Bennett, C. L. and Keyes, O. (2019). What is the Point of Fairness? Disability, AI and The Complexity of Justice.

Crochík, J. A. L. (1996). Preconceito, individuo e sociedade. Temas em Psicologia, 4:47 – 70.

Desvelar, S. (2026). Danos e discriminação algorítmica: Mapeamento. Desvelar Justiça racial, IA e tecnologias digitais. Acesso em: 06/01/2026.

Ferrara, E. (2024). Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1).

Gesser, A. (2008). Do patológico ao cultural na surdez: para além de um e de outro ou para uma reflexão crítica dos paradigmas. Trabalhos em Linguística Aplicada, 47(1):223–239.

Glazko, K., Mohammed, Y., Kosa, B., Potluri, V., and Mankoff, J. (2024). Identifying and Improving Disability Bias in GPT-Based Resume Screening. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’24, page 687–700, New York, NY, USA. Association for Computing Machinery.

Griffin, P., Peters, M. L., and Smith, R. M. (2007). Ableism curriculum design. In Adams, M., Bell, L. A., and Griffin, P., editors, Teaching for Diversity and Social Justice, page 24. Routledge, 2nd edition.

Herold, B., Waller, J., and Kushalnagar, R. (2022). Applying the stereotype content model to assess disability bias in popular pre-trained NLP models underlying AI-based assistive technologies. In Ebling, S., Prud’hommeaux, E., and Vaidyanathan, P., editors, Ninth Workshop on Speech and Language Processing for Assistive Technologies (SLPAT-2022), pages 58–65, Dublin, Ireland. Association for Computational Linguistics.

Hutchinson, B., Prabhakaran, V., Denton, E., Webster, K., Zhong, Y., and Denuyl, S. (2020). Social biases in NLP models as barriers for persons with disabilities. In Jurafsky, D., Chai, J., Schluter, N., and Tetreault, J., editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5491–5501, Online. Association for Computational Linguistics.

Krupiy, T. T. and Scheinin, M. (2023). Disability Discrimination in the Digital Realm: How the ICRPD Applies to Artificial Intelligence Decision-Making Processes and Helps in Determining the State of International Human Rights Law. Human Rights Law Review, 23(3):ngad019.

Li, R., Kamaraj, A., Ma, J., and Ebling, S. (2024). Decoding ableism in large language models: An intersectional approach. In Dementieva, D., Ignat, O., Jin, Z., Mihalcea, R., Piatti, G., Tetreault, J., Wilson, S., and Zhao, J., editors, Proceedings of the Third Workshop on NLP for Positive Impact, pages 232–249, Miami, Florida, USA. Association for Computational Linguistics.

Mondal, I., Kaur, S., Bali, K., Vashistha, A., and Swaminathan, M. (2022). “DisabledOnIndianTwitter” : A dataset towards understanding the expression of people with disabilities on Indian Twitter. In He, Y., Ji, H., Li, S., Liu, Y., and Chang, C.-H., editors, Findings of the Association for Computational Linguistics: AACL-IJCNLP 2022, pages 375–386, Online only. Association for Computational Linguistics.

Pacheco, K. M. D. B. and Alves, V. L. R. (2007). A história da deficiência, da marginalização à inclusão social: uma mudança de paradigma. Acta Fisiátrica, 14(4):242–248.

Urbina, J. T., Vu, P. D., and Nguyen, M. V. (2025). Disability Ethics and Education in the Age of Artificial Intelligence: Identifying Ability Bias in ChatGPT and Gemini. Archives of Physical Medicine and Rehabilitation, 106(1):14–19. Epub 2024-08-30.

Venkit, P. N., Srinath, M., and Wilson, S. (2022). A study of implicit bias in pre-trained language models against people with disabilities. In Calzolari, N., Huang, C.-R., Kim, H., Pustejovsky, J., Wanner, L., Choi, K.-S., Ryu, P.-M., Chen, H.-H., Donatelli, L., Ji, H., Kurohashi, S., Paggio, P., Xue, N., Kim, S., Hahm, Y., He, Z., Lee, T. K., Santus, E., Bond, F., and Na, S.-H., editors, Proceedings of the 29th International Conference on Computational Linguistics, pages 1324–1332, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.

Venkit, P. N., Srinath, M., and Wilson, S. (2023). Automated ableism: An exploration of explicit disability biases in sentiment and toxicity analysis models. In Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023), pages 26–34, Toronto, Canada. Association for Computational Linguistics.