How should generative AI design embed responsible AI principles?

AWS's Cloudscape Design System states that "Responsible AI at AWS relies on the following core dimensions: fairness, explainability, privacy and security, robustness, governance and transparency," and that its generative AI design guidance is built "upon these dimensions." From them it derives a set of design principles for teams building generative AI experiences: evaluate, before adopting generative AI, "if the use case would benefit from automating generation of custom outputs," starting with "basic text or image generation" and expanding "to support more complex use cases gradually"; ensure outputs meet high accuracy, quality and safety standards before deployment and are "directly useful for the user's goals and context"; and "maintain security by using systems to detect and filter harmful or abusive content generated by AI."

Each responsible-AI dimension maps to a specific interface obligation. Governance and transparency becomes a disclosure rule: "be transparent about capabilities and limitations of generative AI," make it "clear when content is AI-generated," and communicate "when errors occur or confidence in the output generated by AI is low." Privacy and security becomes a consent rule: don't "collect or use customer data to improve services without user consent." Fairness becomes an ongoing testing obligation: "take steps to mitigate potential biases in training data," "continue testing for unfair outputs," and "expose settings to encourage diverse, inclusive output from the AI when possible." A fifth, unlabeled thread of user agency runs through the set as its own principle: "enable users to control how generative AI is employed in a task," letting them "guide the experience, set parameters, and override AI if needed."

This sits alongside but differs in kind from How should users control systems with unpredictable outputs?, which addresses the interaction problem that generative AI creates — stochastic outputs that vary even on identical inputs, and the mental-model and trust problems that follow. Cloudscape's principles don't name generative variability as a design problem at all; they sit upstream of it, as a responsible-AI checklist about whether and how to deploy generative AI, and what to disclose once deployed. The user-control principle does overlap with one axis of How should generative UI be organized as a design space?, which treats who controls change as a structural variable of generative UI design; Cloudscape states the same control concern as a design obligation ("override AI if needed") rather than a dimension to be varied by use case.

The excerpt is undated and offers no study, usage data, or product example backing any principle — it is AWS's own stated design-system guidance, not an evaluated finding about which practices actually improve generative AI experiences or reduce harm. It should be read as one organization's codification of its responsible-AI commitments into design language, useful as a comparison point for how other guideline sets (How should users control systems with unpredictable outputs? among them) carve up the same problem space differently, rather than as validated best practice.