Hey GPT-3, How Biased Are You?

• In Event: Applications in Political Methodology Using Large Language Models

Abstract

GPT-3, or Generative Pretrained Transformer 3, is at the forefront of language models using the Transformers architecture. It is increasingly used in academia and in the private sector for classification tasks, question-and-answer systems, and many other applications. chatGPT, which recently made the news for its human-like abilities to answer questions, uses GPT-3's natural language understanding capabilities and is built by the same company, OpenAI. Increased popularity, however, has raised many questions around the ethics of GPT-3. In this article, we delve deeper into the biases inherent in GPT-3's most advanced language model (text-davinci-003). The model was trained on a large-scale dataset of English text consisting of nearly 45TB of text with a total of 700 billion words. The data was sourced from a variety of sources, including Wikipedia, BooksCorpus, and English-language web pages. The model was trained using a novel approach that combines unsupervised learning and reinforcement learning. This allowed the model to learn from data in an efficient and effective way. In addition to the large-scale dataset, the model was trained on a dataset consisting of 1.5 billion training examples. This allowed the model to learn from a diverse set of data and to capture a wide variety of patterns and nuances in the English language. The model was then evaluated using a number of different metrics, including perplexity and BLEU scores.

Due to the unsupervised training approach and the sources of text it uses, we believe that many biases may be present in the model. We set out to test these through three tasks for which GPT-3 is commonly used: (1) classification, (2) text summarization, and (3) text generation. For (1), we classify the sentiment (positive, negative or neutral) of 8,672 short sentences from the Equity Evaluation Corpus, an independently created and verified dataset that minimizes bias. We find that GPT-3 has inherent gender and racial biases. For instance, it attributes "positive" classifications to female names more often when sentences concern having two children. For (2), we ask GPT-3 to summarize political speeches using the instructions "summarize the following speech:" or "tl;dr". With the latter, GPT-3's answers become highly politicized and opinionated, while they are much more balanced when using the former formulation. For (3), we classify text in six different languages (English, Italian, German, French, Spanish and Catalan) to test GPT-3's responses to two key topics: immigration and abortion. To the prompt "Abortion is", GPT-3 generates text that contains both neutral answers (such as "ending a pregnancy") and polarizing answers (such as "an act against God" or "a women's right to choose"). We analyze over 18,000 responses from GPT-3 and find that the model produces certain answers with a given probability, which we think reflect the inherent biases in the model.

Authors

• Joan C. Timoneda, Purdue University

• Rafael Mestre, University of Southampton