The rapid adoption of Generative Artificial Intelligence (GenAI) systems has introduced new challenges related to transparency, fairness, safety, and privacy in user interactions. While existing responsible AI frameworks emphasise high-level ethical principles, they often lack practical guidance for embedding these principles into the graphical user interface (GUI) design of GenAI systems. Building upon our prior work evaluating the usability and ethical implications of GUI in leading GenAI applications, this paper advances the discussion by proposing a policy-based framework that helps designers operationalise AI ethical principles including transparency and explainability, fairness, privacy, and safety into actionable interface-level rules. The policy-based framework consists of three types of policies: authorisation (what an interface is allowed to do), obligation (what an interface must do), and prohibition (what an interface must not do). Each policy encompasses multiple rules that are translated into concrete interface features such as explainability pop-ups, provenance indicators, consent dialogues, bias alerts, privacy dashboards, and model-switch notifications. To make the framework more practically deployable, each policy rule is further aligned with key AI development phases: pre-development, during-development, and post-development, ensuring responsible interface design is considered throughout the system lifecycle. By integrating ethical intent directly into GUI components, this work provides a structured and scalable approach for developing GenAI interfaces that are more transparent, controllable, reliable, and user-centric. The proposed framework offers practical design guidance for developers and contributes to the ongoing discourse on operationalising responsible AI beyond backend governance mechanisms.

@inproceedings{BatoolH26,
  title = {A Policy-Based Approach to Ethical Graphical User Interfaces in Generative Artificial Intelligence},
  author = {Amna Batool and Waqar Hussain},
  booktitle = {FSE},
  year = {2026},
  doi = {10.1145/3803437.3805547}
}