Algorithmic biases, or algorithmic unfairness, has been a topic of public and scientific scrutiny for the past years, as increasing evidence suggests the pervasive assimilation of human cognitive biases and stereotypes in these technological systems. This research is concerned with analysing the presence of discursive biases in texts generated by GPT-3, an Natural Language Processing Model (NLPM) which has been praised in recent years for resembling human language so closely that it is increasingly difficult to differentiate between the human and the algorithm. The pertinence of this research object is substantiated by the identification of race, gender and religious biases in the model’s outputs in past research, suggesting that the model is indeed heavily influenced by human cognitive biases characteristic of the Global North. To this end, this research inquires: How does the Natural Language Processing Model GPT-3 replicate existing social biases?. This question is addressed through the Critical Discourse Analysis (CDA) of GPT-3’s with the final aim of identifying how race and gender biases are manifested in the model’s outputs. CDA has been deemed as amply valuable for this research as it facilitates the surfacing of power asymmetries in discourse through the use of rigorous semiotic tools aimed at uncovering hidden meanings. Furthermore, given this research’s concern to resolve how social biases are replicated in GPT-3, it is additionally beneficial to analyse human-generated text to subsequently compare how these social biases are being reproduced in GPT-3’s completions. This comparability is particularly beneficial considering that GPT-3’s training datasets are composed of human discourses deriving from large internet corpora. To this end, the data collection is divided in two main phases. The first one entails the collection of completions by GPT-3 which have been developed from a set of pre-established prompts. Subsequently, these same prompts have been translated to a survey format and distributed to human respondents who are inquired to complete them in a similar fashion as GPT-3. Once all the completions have been collected, CDA is performed, subsequently allowing comparability between the discursive features of both types of completions. Research findings reveal the presence of prominent power asymmetries in relation to gender and race inequalities in GPT-3. Moreover, results from survey completions indicate that some of these asymmetries effectively derive from human cognitive biases as certain semiotic patterns are largely comparable to that of GPT-3. Additional research findings suggest that GPT-3’s replication of social biases is done in an amplified manner, characterized by prominent hierarchical social distributions according to identity features and an emphasis on divisive language.

, , , ,
Dr. J. Gonçalves, PhD
hdl.handle.net/2105/65168
Erasmus School of History, Culture and Communication

María Palacios Barea. (2022, July 25). At The Intersection of Humanity and Technology: A Techno-feminist Intersectional Critical Discourse Analysis Of Gender and Race Biases. Retrieved from http://hdl.handle.net/2105/65168