Machine learning (ML) offers powerful tools for extracting insights from large-scale census data, enabling data-driven decision-making in public policy. However, the integration of ML into socio-economic analysis raises critical concerns about fairness, particularly when models are used to inform resource allocation and policy development. This study investigates the fairness of six widely used ML algorithms applied to the 2021 Australian Census data, focusing on income classification across demographic groups defined by gender and race. Using a rigorous evaluation framework based on k-fold cross-validation and statistical testing, we assess three key fairness metrics: Equalised Odds, Equal Opportunity, and Treatment Equality. Our findings reveal that several models exhibit significant biases, potentially disadvantaging historically marginalised communities. These findings highlight the need for fairness-aware methodologies and ethical safeguards in deploying ML models for policy applications. By identifying disparities in algorithmic outcomes, this research contributes to the broader discourse on equitable AI and responsible data use in public governance. Given the influential role of census data in shaping public policy, ensuring fairness in predictive models is essential to prevent the reinforcement of existing social inequities.