Algorithmic bias does not stay in the backend: it emerges in the interface. A recommendation system that favours certain groups is visible in what appears on the homepage. A risk model that discriminates translates into different messages, limits and options for different users — often without explanation.
"Smart" filters in hiring, dating, credit or healthcare can exclude opaquely. The designer who creates these flows — what to show, how to phrase a rejection, whether to offer appeal — determines whether exclusion is visible and contestable or silent.
AI-optimised onboarding can seem like a UX improvement: fewer steps, personalisation, higher completion rate. But if it optimises for majority users (young, urban, tech-savvy), those who deviate from the "average" profile hit invisible obstacles — forms that do not understand their context, impossible data requests, exclusive language.
Recognition interfaces — face, voice, documents — have different error rates by ethnicity, gender, age, accent. Designing human fallback, non-blaming error messages and alternative paths is not optional: it is baseline inclusivity.
The designer can ask for equity metrics by segment: onboarding completion rates by age/disability/language, recognition errors by demographics, recommendation distribution. Without metrics, bias stays invisible until it explodes in a media scandal.
Case study: Snapchat filters that altered African-American features ("blackface filters"), or CV screening systems that penalised women candidates. In both cases, the interface normalised algorithmic discrimination. The designer could — should — have flagged and redesigned.