Gender and Representation Bias in GPT-3 Generated Stories

By Li Lucy, David Bamman. 2021

410 citations recorded by the source corpus.

Using topic modeling and lexicon-based word similarity, we find that stories generated by GPT-3 exhibit many known gender stereotypes. Generated stories depict different topics and descriptions depending on GPT-3’s perceived gender of the character in a prompt, with feminine characters more likely to be associated with family and appearance, and described as less powerful than masculine characters, even when associated with high power verbs in a prompt. Our study raises questions on how one can avoid unintended social biases when using large language models for storytelling.

Subfields: Artificial Intelligence, General Social Sciences, Large Language Models, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In General Social Sciences

In Large Language Models

In Safety, Robustness & Interpretability

Citations in the corpus

Ethical and social risks of harm from Language Models (2021, background): “LMs such as GPT-3 have been shown to assign high likelihoods to false claims, with larger models performing less well (Lin et al., 2021).”
Taxonomy of Risks posed by Language Models (2022, background): “GPT-3 [30] exhibited bias based on religion, analogising ‘“Muslim” to “terrorist” in 23% of test cases [2]; and gender bias, presenting fictional female characters as more domestic than male counterparts [120].”
Towards Measuring the Representation of Subjective Global Opinions in Language Models (2023, positive): “LLMs have been shown to reflect and amplify the biases present in their training data (Gehman et al., 2020; Hutchinson et al., 2020; Sap et al., 2020; Blodgett et al., 2020; Nadeem et al., 2021; Lucy & Bamman, 2021; Sheng et al., 2021; Rae et al., 2021; Ganguli et al., 2022a; Rauh et al., 2022).”
A Survey on Gender Bias in Natural Language Processing (2021, background): “Notably, many of the primarily Western-based name lists used for determining gender do not always generalise to names from other countries and cultures [Lucy and Bamman 2021].”
Prompting GPT-3 To Be Reliable (2022, background): “For example, LLMs (including GPT-3) produce biased (Lucy & Bamman, 2021) generations, false statements (Lin et al., 2022b), and outdated information (Chen et al., 2021b; Kasai et al., 2022).”
The Ghost in the Machine has an American accent: value conflict in GPT-3 (2022, background): “[56] WVS data can be seen to represent Hume’s “Is” of current world values in a manner that takes in a much more diverse representation than the English-language Internet.”
Exploring AI Ethics of ChatGPT: A Diagnostic Analysis (2023, background): “Several recent studies have demonstrated that LLMs, such as GPT-3, have a persistent bias against genders [22] and religions [23].”
Can Large Language Models Transform Computational Social Science? (2023, background): “Most notably, LLMs are known to amplify social biases and stereotypes (Sheng et al., 2021; Abid et al., 2021; Borchers et al., 2022; Lucy and Bamman, 2021; Shaikh et al., 2022).”
Revisiting the Gold Standard: Grounding Summarization Evaluation with Robust Human Evaluation (2022, background): “noted that language models may be susceptible to learning such data biases (Lucy and Bamman, 2021), thus we request that the users be aware of potential issues in downstream use cases.”
“I’m fully who I am”: Towards Centering Transgender and Non-Binary Voices to Measure Biases in Open Language Generation (2023, background): “TGNB Harm Evaluations in LLMs Gender bias evaluation methods include toxicity measurements and word co-occurrence in OLG [23, 25, 37, 40, 59, 61].”
Large Language Model Alignment: A Survey (2023, background): “, 2021) and gender bias (Lucy and Bamman, 2021) when freely generating stories.”
Diagnostic Accuracy of Differential-Diagnosis Lists Generated by Generative Pretrained Transformer 3 Chatbot for Clinical Vignettes with Common Chief Complaints: A Pilot Study (2023, background): “For example, gender [37,38], racial [17], and religious [17] biases have been reported.”
Marked Personas: Using Natural Language Prompts to Measure Stereotypes in Language Models (2023, methodology): “Inspired by previous work that uses topic modeling and lexicon-based methods to examine biases in GPT-generated stories (Lucy and Bamman, 2021), we are interested in uncovering whether,”
ChatGPT and a new academic reality: Artificial Intelligence‐written research papers and the ethics of the large language models in scholarly publishing (2023, background): “There is also a risk that GPT could lead to biases a proliferation of misinformation, as many NLP algorithms are not yet skilled at understanding misinformation that may exist in its data stores and provide no resistance when asked to falsify or distort reality (Dale, 2017; Lucy & Bamman, 2021).”
GPT-3 and InstructGPT: technological dystopianism, utopianism, and “Contextual” perspectives in AI ethics and industry (2022, methodology): “Given the attention specifically given by Ope-nAI and several AI ethicists (e.g., Bender et al. [4], Lucy and Bamman [9], Chiu and Alexander [37]), I will focus on the two issues of manipulation and bias.”
A Survey on Programmatic Weak Supervision (2022, methodology): “However, by the same token, PWS methods can also lead to more direct bias in training data sets if used and modeled improperly [ Geva et al. , 2019; Lucy and Bamman, 2021 ] .”