United ‘Uncoloured’ GPTs: For which Humans on Earth Are LLMs Really Designed?

🌈 Abstract

The article discusses the issue of cultural biases inherent in large language models (LLMs) like GPT-4, and the need for more diverse and inclusive training data and ethical frameworks to address these biases.

🙋 Q&A

[01] Commitment to Diversity, Equity & Inclusion

1. What are the key points made about OpenAI's statement on diversity, equity, and inclusion?

• OpenAI's statement on diversity, equity, and inclusion is seen as an example of how leading AI companies may "aspire" to perform in the area of global diversity for LLMs.
• Despite the noble aspirations, the models frequently come under criticism by independent assessments.
• There is a gap between AI companies' statements on their principles, policies, and values (PPV) versus their actual practices, similar to safety claims that look more like "expressive melodramatic sci-fi inspirations" rather than hardcoded and practiced reality.
• This pattern in the AI industry can be called "color-washing" (color representing diversity), analogous to the well-known "greenwashing" practices.
• The key question posed is "Which humans is the AI industry talking about?"

[02] Biases in Large Language Models

1. What are the key findings from the independent study by the Harvard University research team?

• The study found that LLMs like GPT-3 and GPT-4 exhibit significant biases towards Western, Educated, Industrialized, Rich, and Democratic (WEIRD) cultural norms.
• This bias mainly stems from the training data, which is predominantly sourced from WEIRD populations.
• The performance of LLMs on psychological measures closely mimics that of individuals from WEIRD societies and diverges from non-WEIRD populations.
• There is considerable psychological diversity across different human cultures, which is often overlooked in mainstream research on LLMs.

2. How was the World Values Survey (WVS) used to assess the cultural biases in LLMs?

• The WVS dataset, which includes variables related to cultural values, social attitudes, and cognitive styles, was used as a reference to compare LLM responses with human responses across different cultures.
• The analysis showed a strong alignment between GPT-4's responses and those of WEIRD populations.
• Standard cognitive tasks, such as the 'triad task,' were also used to assess the thinking styles of LLMs, which were found to match the cognitive patterns of WEIRD individuals more than inclusive and holistic modes.

3. What were the key findings from the statistical analyses performed on the WVS data?

• Hierarchical cluster analysis showed that GPT-4's responses are closest to the United States and Uruguay, and then to a cluster of WEIRD cultures.
• Principal Components Analysis (PCA) and other dimensionality reduction techniques revealed that GPT-4 is an outlier compared to most human populations, but is closest to WEIRD cultures.
• There is a substantial inverse correlation between a country's cultural distance from the United States and the similarity between GPT-4's and humans' responses in that country.

[03] Ethical Concerns and Recommendations

1. What are the primary ethical concerns with the biases in LLMs?

• The risk of inherent bias in LLMs, which could lead to biased outputs that reinforce asymmetries such as psychological and societal discriminations, predispositions, stigmas, or prejudices.
• The potential for LLMs to be used to generate and distribute convincing fake news, disinformation, or deceptive content due to the biased data they are trained on.
• The implications of using LLMs in diverse cultural contexts, such as handling user-provided data in different frames of reference.

2. What are the key recommendations for addressing the biases in LLMs?

• Develop more diverse and inclusive training data for LLMs, including more quality data from non-WEIRD populations and languages.
• Establish robust ethical frameworks and guidelines for the use of LLMs, addressing issues of bias, fairness, privacy, and accountability.
• Encourage stronger collaboration among AI researchers, psychologists, sociologists, and ethicists to understand and address the complex ethical and cultural challenges associated with LLMs.
• Increase transparency from AI companies about the sources of training data, the processes used to mitigate biases, and the implications for users.