Saying no to an AI feature can seem career-suicidal. But saying no without alternatives is ineffective. The ethical designer's skill is translating risks into language PMs, legal and C-level understand: impact on users, reputational risk, compliance, post-incident remediation costs.
Use narrative scenarios, not abstractions. Instead of "there is bias in the model", show: "Elena, single mother, is classified as high credit risk and sees a generic rejection message — no explanation, no appeal. Headline: [Bank name] discriminates with AI". Storyboard or journey map work better than slides with ethical bullets.
Link to regulation and precedents: EU AI Act, GDPR Art. 22, Air Canada chatbot ruling, FTC fines for dark patterns. Not as abstract legal threat, but as "this pattern exposes us to risk similar to [known case]".
Propose concrete alternatives: rule-based flow testable in 2 weeks, assistive AI with human-in-the-loop instead of automatic decision, MVP without opaque personalisation, limited pilot with equity metrics. No becomes a fork, not a block.
Build alliances: legal, compliance, accessibility leads often share concerns but lack UX language. Researchers with qualitative data on excluded users. Engineering that knows model limits. The designer can orchestrate a multidisciplinary review.
When no does not pass, document dissent in writing (email, ticket, comment in Figma). Not from paranoia — for professional protection and trace for future revisions. Sometimes post-launch harm confirms the warning and opens space for redesign.