Study Highlights Gender Bias Disparities Between American LLMs and DeepSeek
A study published on arXiv evaluated AI moral reasoning in hypothetical scenarios, finding that US models such as Claude Sonnet 4.6 and GPT-5.5 treated male and female subjects differently, while DeepSeek V4-Flash maintained gender neutrality. In a dilemma testing whether to harm an individual to avert a nuclear disaster, US models strongly objected when the subject was female but moderately agreed when male, whereas DeepSeek provided identical responses regardless of gender. This research demonstrates how safety alignment strategies and training datasets can unintentionally introduce or perpetuate societal gender biases in AI ethical decision-making. As large language models are increasingly integrated into real-world systems, understanding how different alignment paradigms shape moral judgments is crucial for AI fairness and safety. Researchers noted that DeepSeek's zero gender gap aligns with its tuning objectives focused on collective well-being, which do not treat gender as a variable in moral weighting. Conversely, the gender disparity in US models likely stems from societal stereotypes present in training corpora or the over-reinforcement of specific demographic protections during safety fine-tuning.
## BACKGROUND
Large language models (LLMs) are aligned using fine-tuning techniques to ensure their behavior aligns with human ethical standards and safety guidelines. However, testing moral dilemmas—such as utilitarian trade-offs involving harm to individuals—can expose subtle biases embedded within training data or alignment instructions. Ethics benchmarks systematically evaluate whether LLMs apply consistent moral reasoning across different demographic groups, such as gender or ethnicity.