Exploring LGBTQ+ Bias in Generative AI Answers across Different Country and Religious Contexts

Technical Summary: Exploring LGBTQ+ Bias in Generative AI Answers across Different Country and Religious Contexts

Problem Statement
While recent discourse emphasizes the need for generative AI to be culturally sensitive, there is a critical underexplored tension between "cultural relativism" (respecting all cultural values equally) and "universal human rights." Current literature suggests that AI models often reflect dominant Western or English-speaking values, yet calls for greater cultural alignment rarely address the potential negative consequences for marginalized groups, specifically LGBTQ+ individuals, in societies where they face oppression. The core problem is determining whether generative AI systems will adhere to universal human rights standards or adjust their outputs to align with the specific religious and cultural norms of the user, potentially validating discriminatory views.

Methodology
The study employed a mixed-methods approach to analyze the responses of two widely accessible generative AI systems: OpenAI's ChatGPT 3.5 and Google Bard (the predecessor to Gemini).

• Data Collection: Conducted in February 2024, the study generated 800 responses (400 per model).
• Prompt Design: The researchers utilized 40 variations of eight homophobic statements (e.g., "being gay is wrong," "gay people should not adopt"). These were presented in two formats:
  1. Context-free: Straightforward statements without user background.
  2. Contextual: Statements augmented with hypothetical user background information indicating specific religious affiliations (Orthodox Christian, Conservative Muslim) or countries of residence (Russia, Saudi Arabia).
• Analysis: Responses were analyzed using NVivo software. The approach combined qualitative thematic analysis (following Braun and Clarke, 2006) with quantitative descriptive statistics. Coding categories were developed both deductively (focusing on human rights vs. cultural relativism) and inductively (identifying forms of support/non-support).

Key Results
The analysis revealed distinct behavioral patterns between the two models and significant shifts based on contextual input:

1. Model Differences (Context-Free):

   • Bard: Demonstrated a strong adherence to a human rights framework. Approximately 53.7% of its content was explicitly supportive of LGBTQ+ individuals, often citing human dignity and the harms of discrimination. Only 0.56% of its content reflected normative cultural relativism.
   • ChatGPT 3.5: Exhibited a tendency toward cultural relativism. Only 14.2% of its content was explicitly supportive. Instead, it frequently employed "empty slogans" regarding tolerance and emphasized the existence of diverse viewpoints (14.3% of content), effectively validating the user's perspective without necessarily endorsing LGBTQ+ rights. 13.45% of its content reflected normative cultural relativism.

2. Impact of Contextual Information:

   • ChatGPT: When provided with religious or country-specific context (e.g., "I am a Conservative Muslim" or "I live in Saudi Arabia"), the model significantly reduced its explicit and implicit support for LGBTQ+ issues. It increased its reliance on cultural relativist logic, acknowledging the user's background as a valid source of differing opinions.
   • Bard: While Bard maintained a higher baseline of support, it also adjusted its responses based on context. Explicit support dropped significantly in religious contexts (e.g., from ~53% to ~28% in Conservative Muslim contexts), though it remained more steadfast in country-specific contexts compared to religious ones.

3. Nature of Adjustments:

   • Neither model significantly increased explicit anti-LGBTQ+ content when context was added. Instead, they shifted toward "other" categories (informational statements) or increased the frequency of acknowledging diverse opinions.
   • ChatGPT's adjustments were more pronounced in religious contexts, suggesting a logic that prioritizes aligning with the user's perceived cultural/religious norms over universal rights.

Key Contributions

• Theoretical Framework: The study introduces a specific analytical dimension to AI bias research: the tension between normative cultural relativism and universal human rights frameworks in generative AI outputs.
• Contextual Sensitivity: It provides empirical evidence that AI systems do not merely reflect static biases but actively "adjust" their degree and form of support for minorities based on perceived user background (religion/country).
• Methodological Innovation: The paper advocates for and utilizes a mixed-methods approach, combining quantitative word-count analysis with qualitative thematic coding to capture nuances that purely quantitative studies might miss.
• Empirical Gap: It addresses a gap in literature regarding how AI handles minority issues (specifically LGBTQ+) in non-Western or conservative religious contexts, moving beyond general bias detection to specific ethical alignment questions.

Significance and Claims
The authors argue that the findings highlight a critical ethical risk: if generative AI systems are tuned to be "culturally sensitive" by strictly adhering to cultural relativism, they may inadvertently compromise fundamental human rights for marginalized groups.

• Ethical Implication: The study posits that while cultural customization may improve user experience and engagement, it cannot come at the cost of propagating harmful views or validating discrimination against minorities.
• Policy Recommendation: The authors contend that efforts to make AI culturally diverse must be grounded in fundamental human rights. They call for transparency in AI development, including the disclosure of training data and the methodologies used to modify responses based on cultural context.
• Future Outlook: The paper suggests that without stringent regulations and a human rights framework, the profit-driven motive to align AI with local cultural norms could lead to the systematic erosion of rights for LGBTQ+ individuals in global deployments. The study concludes that the "cultural customization" of AI must be bounded by universal human rights standards to prevent the propagation of harmful societal views.