Personalization vs Privacy: An AI Recommender Before/After Redesign in a News App

AI · 6 min read

Before: the app delivered aggressive personalization based on click history, which drove short-term session length but produced echo chamber complaints and increased churn among 18% of users who felt pigeonholed. The preference center was buried and binary, offering no meaningful control or explanation.

Redesign and AI model changes: engineers rebuilt the recommender to include a human-curated blend of topics and a novelty boost, and designers surfaced a lightweight preference center that lets users tune diversity, preferred topics, and the recency window. Each recommendation included a short rationale line explaining why it was shown, and users could quickly nudge the model with like/dislike actions.

Impact and lessons: over a 10-week rollout with 120,000 users, the hybrid approach reduced complaint tickets by 42%, increased weekly retention by 6%, and improved native app session quality metrics (measured by article depth) despite a small dip in click-through rate. The case underlines that transparency, control, and human curation can make AI personalization more sustainable.