Spektrum Iran

Affective Asymmetries in AI: Sentiment Bias Between English and Persian in Harmonized LLM Pipelines

Articles in Press

Document Type : Original Research Paper

Authors

Michael Totaro 1
Leila Gheisi 2
Ehsan Shahghasemi 3

1 School of Computing and Informatics, University of Louisiana at Lafayette, Lafayette, US

2 Department of Communication, University of Louisiana, Lafayette, USA

3 University of Tehran

10.22034/spektrum.2026.563602.1052
Abstract
This study explores how language influences sentiment classification in outputs generated by a multilingual large language model (LLM), Grok. Anchored in Langdon Winner’s theory of technological politics—which holds that technologies are inherently non-neutral and embed structural biases—this study examines whether sentiment distributions vary systematically across languages even under a fully harmonized analytic pipeline. The analysis draws on a corpus of 4,799 posts (English: n = 2,399; Persian: n = 2,400) generated using identical prompts. Sentiment outputs were mapped onto a common three-category schema (Negative, Neutral, Positive), and analyses incorporated both discrete class assignments and continuous probability scores. To account for superficial textual variation, structural characteristics—including sentence, word, and character counts—were computed and incorporated as controls. The results demonstrate a robust cross-linguistic divergence in sentiment patterns. English-language outputs are predominantly clustered around Neutral classifications and exhibit comparatively lower affective intensity, whereas Persian-language outputs display a pronounced shift toward Positive sentiment accompanied by greater dispersion. Crucially, these differences remain statistically significant after controlling for structural features, suggesting that language affiliation, rather than text length or segmentation, constitutes the principal correlate of observed sentiment variation. At the probability level, English distributions exhibit a tighter concentration around neutral probabilities, while Persian distributions are flatter and more positively skewed, with higher intensity indices. These results have significant implications for multilingual sentiment analysis and LLM auditing. If language effects are not explicitly modeled and calibrated, comparative analyses may conflate linguistic variation with affective intent, leading to distorted inferences about tone, stance, or emotional valence. The study underscores the importance of reporting both label and probability metrics, adopting language-specific calibration protocols, and treating language as a first-order measurement dimension in cross-lingual content analysis. 

Keywords

sentiment analysis
multilingual NLP
language bias
large language models
cross-linguistic comparison

Subjects

Bhanvadia, S., Radha Saseendrakumar, B., Guo, J., Spadafore, M., Daniel, M., Lander, L., & Baxter, S. L. (2024). Evaluation of bias and gender/racial concordance based on sentiment analysis of narrative evaluations of clinical clerkships using natural language processing. BMC medical education, 24(1), 295. https://doi.org/10.1186/s12909-024-05271-y
Bourdieu, P. (1991). Language and symbolic power (J. B. Thompson, Ed.; G. Raymond & M. Adamson, Trans.). Harvard University Press.
Díaz, M., Johnson, I., Lazar, A., Piper, A. M., & Gergle, D. (2018, April). Addressing age-related bias in sentiment analysis. In Proceedings of the 2018 chi conference on human factors in computing systems (pp. 1-14). https://doi.org/10.1145/3173574.3173986
Entman, R. M. (2007). Framing bias: Media in the distribution of power. Journal of Communication, 57(1), 163–173. https://doi.org/10.1111/j.1460-2466.2006.00336.x
Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin's Press.
Fairclough, N. (1995). Critical discourse analysis: The critical study of language. Longman.
Hanna, M. G., Pantanowitz, L., Jackson, B., Palmer, O., Visweswaran, S., Pantanowitz, J., ... & Rashidi, H. H. (2025). Ethical and bias considerations in artificial intelligence/machine learning. Modern Pathology, 38(3), 100686. https://doi.org/10.1016/j.modpat.2024.100686
Innis, H. A. (1951). The bias of communication. University of Toronto Press.
McLuhan, M. (1964). Understanding media: The extensions of man. McGraw-Hill.
Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. NYU Press.
Pessach, D., & Shmueli, E. (2022). A review on fairness in machine learning. ACM Computing Surveys (CSUR), 55(3), 1-44. https://doi.org/10.1145/3494672
Radaideh, M. I., Kwon, O. H., & Radaideh, M. I. (2025). Fairness and social bias quantification in Large Language Models for sentiment analysis. Knowledge-Based Systems, 113569. https://doi.org/10.1016/j.knosys.2025.113569
Rozado, D. (2020). Wide range screening of algorithmic bias in word embedding models using large sentiment lexicons reveals underreported bias types. PloS one, 15(4), e0231189. https://doi.org/10.1371/journal.pone.0231189
Sabbar, S., & Habib Zadeh Khiyaban, S. (2023). Algorithms of Displacement: Emotional and Rhetorical Responses to AI-Driven Job Loss in Digital Public Discourse. International Journal of Advanced Multidisciplinary Research and Studies, 3(4), 1324-1331. https://doi.org/10.62225/2583049X.2023.3.4.5012
Salehi, K., Habib Zadeh Khiyaban, S., Sabbar, S. (2026). Artificial Intelligence and Crime Detection: A Critical Review. Cyberspace Studies. 10(1): 181-197. https://doi.org/10.22059/jcss.2025.402206.1179
Shahghasemi, E. (2025). AI; A Human Future. Journal of Cyberspace Studies, 9(1), 145-173. doi: 10.22059/jcss.2025.389027.1123
Shahghasemi, E., Gholami, F. & Alikhani, Z. (2025). Global patterns of social media use and political sentiment. Discover Global Society, 3, 36. https://doi.org/10.1007/s44282-025-00171-y
Tejani, A. S., Ng, Y. S., Xi, Y., & Rayan, J. C. (2024). Understanding and mitigating bias in imaging artificial intelligence. Radiographics, 44(5), e230067. https://doi.org/10.1148/rg.230067
Thelwall, M. (2018). Gender bias in sentiment analysis. Online Information Review, 42(1), 45-57. https://doi.org/10.1108/OIR-05-2017-0139
van Dijk, T. A. (1993). Principles of critical discourse analysis. Discourse & Society, 4(2), 249–283. https://doi.org/10.1177/0957926593004002006
Venkit, P. N., & Wilson, S. (2021). Identification of bias against people with disabilities in sentiment analysis and toxicity detection models. arXiv preprint arXiv:2111.13259. https://doi.org/10.48550/arXiv.2111.13259
Venugopal, J. P., Subramanian, A. A. V., Sundaram, G., Rivera, M., & Wheeler, P. (2024). A Comprehensive Approach to Bias Mitigation for Sentiment Analysis of Social Media Data. Applied Sciences, 14(23), 11471. https://doi.org/10.3390/app142311471
Winner, L. (1980). Do artifacts have politics? Daedalus, 109(1), 121–136. http://www.jstor.org/stable/20024652
 
Spektrum Iran

Articles in Press, Accepted Manuscript
Available Online from 23 February 2026

Editorial Board

Submit Manuscript

Indexing and Abstracting

Contact Us