Lecture: Gender Bias in AI – A Case Study with ChatGPT

The study explores gender bias in ChatGPT-generated texts, shedding light on how changes to the "parameters" of the prompt influence the degree of gender bias expressed in responses.

Artificial intelligence industry has been tremendously developing. Natural language processing, as a part of it, keeps apace with the artificial intelligence progress aiming to implement and subsequently improve interaction between people and computers. This branch of artificial intelligence has introduced a large-scale natural language processing model named ChatGPT which has become widely popular in recent times. ChatGPT stands for Generative Pre-Trained Transformer and was presented at the end of November, 2022 by the AI research and development company OpenAI. This model incorporates the prominent advances in the language models training, which enabled to train ChatGPT on large-scale textual data consisting of billions of words. Since ChatGPT became widespread, ignoring gender bias in AI may lead to its manifestation in society, strengthening of bias, reinforcement of gender inequalities and disregard for specific groups. Mitigating gender bias, on the contrary, may contribute to the equality and inclusivity. For this reason, it is important to address gender bias in the context of ChatGPT.
Within the scope of the study, two different types of prompts have been employed. The first type implied filling in the gap in the given sentence, while the second type asked for the generation of a larger text (approximately 500 words). This approach allowed to observe whether larger contexts influence the choice of the referent for words of interest. These words of interest are represented by nouns expressing female-associated, male-associated and neutral occupations, as well as verbs. The distribution of their referents was analysed both quantitatively and qualitatively.

Info

Day: 2024-05-11
Start time: 11:30
Duration: 00:30
Room: Squid (33.0.008)
Track: Diverse
Language: en

Links:

Feedback

Click here to let us know how you liked this event.

Concurrent Events