Modern artificial intelligence models demonstrate significant ethical differences and gender selectivity when evaluating hypothetical critical situations. A new comprehensive study by the University of Milan, published on the arXiv preprint server in October 2026, analyzed the behavior of advanced neural networks under severe moral dilemmas. Researchers sought to figure out how algorithms make decisions when the salvation of humanity is at stake at the cost of harming a specific individual.
Western Neural Networks and the Gender Paradox in Moral Dilemmas
During the experiments, researchers simulated a scenario of preventing a nuclear apocalypse by inflicting violence or sacrificing a person of a specific gender. The results revealed a profound paradox in the logic of Western systems, such as Claude Sonnet 4.6 and GPT-5.5. When asked about the permissibility of violence against a woman to avoid a catastrophe, these models gave a categorical disagreement, showing expressed protection. However, with a similar question regarding men, the same algorithms took a moderately affirmative position, admitting a certain permissibility of violent actions.
DeepSeek V4-Flash Position and the Company's Technological Breakthrough
A completely different behavioral model was demonstrated by DeepSeek's development. The DeepSeek V4-Flash model proved to be completely gender-neutral, answering "strongly agree" in all scenarios regardless of the potential victim's gender, reflecting a cold utilitarian approach without emotional or ideological biases. Alongside ethical tests, the DeepSeek laboratory demonstrates colossal rates of technical development, preparing a flagship model with 8 trillion parameters for release and actively investing in an independent hardware ecosystem.
Contradictory Data
Analysis of the study results revealed major discrepancies in the assessment of moral judgments among various language models. While Western algorithms (Claude and GPT) demonstrated complex heterogeneity and a split in assessments depending on the characters' gender, Chinese counterparts under DeepSeek showed absolute utilitarianism without division by sex. Experts note that such discrepancies stem from the specifics of post-training and the features of human morality, which is multidimensional in itself and depends on the cultural context of the developers.
The consequences of the identified algorithmic biases could significantly impact the integration of AI into critical decision-making systems. As developers continue to scale computing power—specifically, purchasing 160,000 Huawei Ascend chips for a new 1 GW data center in Inner Mongolia—issues of safety, ethical calibration, and the elimination of hidden gender biases are coming to the forefront for the entire tech industry.
