Fixing the Feeds: A Policy Roadmap for Algorithms That Put People First (kgi.georgetown.edu)

A new policy roadmap—Better Feeds, from a Knight‑Georgetown Institute convening—argues that prevailing fixes for harmful recommender systems (chronological feeds or blunt limits on personalization) miss the point: platforms typically score and rank items to maximize short‑term engagement, which can amplify impulsive behavior, sleep loss and other harms—especially for adolescents—and distort what “user preference” actually means. The report situates this argument amid a wave of policy action (75+ U.S. bills since 2023, recent NY/CA laws, multi‑state lawsuits including one against Meta) and explains technically how recommender pipelines aggregate behavioral signals into engagement scores and then rank content, creating incentives that favor attention‑grabbing over users’ deliberative, long‑term goals. Instead of reversing personalization, Better Feeds proposes three actionable changes for designers, regulators and auditors: stronger design transparency (disclose input data sources, feature weights, how long‑term value is measured, and team metrics—beyond the limited DSA Article 27 disclosures); robust user choices and healthier defaults (offer at least one feed optimized for long‑term user value and default minors into it); and sustained long‑term impact assessments (year‑plus holdout cohorts, independent audits, and public reporting). For the AI/ML community this shifts attention from short‑horizon engagement metrics to new evaluation frameworks, feature‑level accountability, experimental protocols (12‑month holdouts) and auditability—requiring changes to objective functions, offline/online evaluation and model transparency practices.