Bias in Decision-Making for AI's Ethical Dilemmas: A Comparative Study of ChatGPT and Claude

All major AI language models exhibit measurable biases related to protected attributes (such as race, gender, and other demographic characteristics), with the study testing nine popular models across 50,400 trials and finding significant biases in every single one, indicating this is a widespread issue rather than isolated to specific systems.

Different types of AI models show different bias patterns: Open-source models tend to favor marginalized groups and show stronger ethical preferences, while closed-source models (like ChatGPT and Claude) are more selective and tend to favor mainstream groups, suggesting that model design and training approaches significantly influence ethical outcomes.

AI models behave inconsistently depending on the type of ethical scenario, maintaining relatively consistent patterns when deciding protective actions but showing more varied and unpredictable responses in harmful scenarios, which raises concerns about reliability in high-stakes decision-making contexts.

Biases become more pronounced and complex when multiple demographic characteristics are considered together (intersectional conditions), meaning that AI systems may exhibit deeper or different biases when making decisions about people with multiple marginalized identities compared to single-attribute evaluations.

Current evaluation methods may be insufficient for assessing AI ethics, as the study demonstrates the need for multi-dimensional, context-aware approaches rather than simple one-off tests, suggesting that organizations deploying these systems should conduct more rigorous and comprehensive ethical audits.