Co-Authoring the Self: A Human-AI Interface for Interest Reflection in Recommenders

TL;DR

The paper tackles the lack of interactive, transparent user profiles in recommender systems by introducing a human–AI collaborative self-portrait interface that generates editable natural language summaries of a user's movie interests. It evaluates this interface through an eight‑week field trial on MovieLens with 1,775 active users, complemented by formative surveys, log analysis, and post‑experiment feedback. The study demonstrates that while AI summaries are imperfect and misalign with user perceptions, the critique and co-creation process increases engagement, exploration, and reflection, and provides design directions for trustworthy, user-driven personalization. Overall, the work argues for designing recommender interfaces as collaborative partners that invite dialogue and correction, enabling users to manage their digital identities and reduce opacity in personalization.

Abstract

Natural language-based user profiles in recommender systems have been explored for their interpretability and potential to help users scrutinize and refine their interests, thereby improving recommendation quality. Building on this foundation, we introduce a human-AI collaborative profile for a movie recommender system that presents editable personalized interest summaries of a user's movie history. Unlike static profiles, this design invites users to directly inspect, modify, and reflect on the system's inferences. In an eight-week online field deployment with 1775 active movie recommender users, we find persistent gaps between user-perceived and system-inferred interests, show how the profile encourages engagement and reflection, and identify design directions for leveraging imperfect AI-powered user profiles to stimulate more user intervention and build more transparent and trustworthy recommender experiences.

Figures (10)

• Figure 1: The default MovieLens "About your ratings" page. We asked users to evaluate their experience with the interface during the formative survey.
• Figure 2: Example interface for the movie interest self-portrait and categorical treemap visualization. Users can toggle between the two views using the tab button in the top-left corner. Top: The self-portrait interface, allowing users to view, edit, and save their current interest summaries. Bottom: The movie taste treemap that users could use to explore their rated movies in six categories, with zoom-in and zoom-out interactions available for each block.
• Figure 3: User interest summary generation pipeline. The dotted box indicates the optional user-edited summary (not available in initial session), which is incorporated as context in subsequent regeneration. The pipeline runs for each qualified user at experiment start, and loops after rating activity exceeds the defined threshold (added 10% of total ratings or 10 more new ratings after each generation).
• Figure 4: Distribution of user editing count and chronological change by different interest types.
• Figure 5: Distribution of user editing level by different interest types.