Towards Socially Responsible AI: Cognitive Bias-Aware Multi-Objective Learning

AI systems inherit historical biases from training data: The paper demonstrates that AI models trained on real-world data automatically learn and reproduce cognitive biases present in human society, such as associating specific emotions with particular genders (e.g., predicting women are more likely to be afraid and men more likely to be angry).

Stereotypical predictions can perpetuate social injustice: The research identifies a concrete risk where AI systems may reinforce harmful stereotypes—such as predicting that certain racial groups use abusive language—which could amplify existing prejudices at scale when deployed in real applications.

Multi-objective learning can reduce bias in AI predictions: The paper proposes a technical framework that simultaneously optimizes for prediction accuracy while actively reducing biased associations between identity attributes (gender, ethnicity) and sensitive outcome categories, showing measurable improvements over standard approaches.

Bias mitigation requires explicit design choices: Rather than assuming AI systems will be fair by default, the research shows that developers must intentionally identify sensitive identity attributes and problematic stereotype combinations, then build these considerations into the model's learning process from the start.

Progress is measurable but requires ongoing attention: While the proposed bias-aware approach successfully reduced gender-based emotional stereotypes compared to baseline models, the paper suggests this is an ongoing challenge requiring continued vigilance as AI systems are deployed across different contexts and datasets.