TLDR: This paper introduces a human-AI collaborative interface for movie recommender systems, allowing users to edit AI-generated summaries of their interests. An eight-week study on MovieLens found that while AI summaries weren’t always perfect, the act of reviewing and editing them significantly increased user engagement, reflection, and self-awareness, especially for disliked interests. The research suggests that “productively wrong” AI summaries can drive user intervention and lead to more transparent and trustworthy recommendation experiences.
Recommender systems are everywhere, from suggesting movies to products, but how well do they truly understand us? Often, the profiles these systems build are static and lack transparency, making it hard for users to see, understand, or even correct what the system thinks their interests are. This can lead to recommendations that miss the mark and a feeling of disconnect for the user.
A new research paper, “Co-Authoring the Self: A Human-AI Interface for Interest Reflection in Recommenders,” explores a novel approach to this challenge. Authored by Ruixuan Sun, Junyuan Wang, Sanjali Roy, and Joseph A. Konstan from the University of Minnesota, this study introduces a human-AI collaborative profile for a movie recommender system. The core idea is to give users an editable, personalized summary of their movie interests, allowing them to directly inspect, modify, and reflect on the system’s inferences. You can read the full paper here: Co-Authoring the Self: A Human-AI Interface for Interest Reflection in Recommenders.
A New Way to See Yourself Through Movies
The researchers designed an interactive user profile page on the popular MovieLens platform. This “self-portrait” interface, powered by large language models (LLMs), dynamically generates textual summaries of a user’s movie interests based on their historical ratings. It’s not a one-time thing; the interface supports user editing and regenerates summaries as new ratings are added, fostering continuous interest representation and user reflection.
The self-portrait is divided into three editable text fields: recent (one-year) interests, long-term enjoyed movies, and long-term disliked movies. The AI generates these summaries using sophisticated techniques like clustering community-contributed movie tags and applying contrastive filtering to avoid contradictions. For recent interests, it focuses on specific movie-level details like top genres, actors, and directors from highly-rated films in the past year. What’s crucial is that any user edits are fed back into the AI’s prompt, creating a co-creative loop where both human and AI contribute to the profile.
Understanding User Behavior and System Impact
The study involved an eight-week online field deployment with 1,775 active MovieLens users. It adopted a mixed-methods design, starting with a formative survey to understand user needs, followed by the online experiment, and concluding with log and post-survey analyses to evaluate the interface’s utility and long-term behavioral effects.
One of the key findings was how users engaged with the editing feature. Out of the active participants, 216 edited the AI-generated summary at least once. Users were more likely to “prune” (significantly change or delete) disliked long-term interest summaries, while “liked” long-term interests were often “retained” (kept with over 95% semantic similarity). Interestingly, editing activity peaked in the second week and then gradually declined, suggesting that users’ perceived interests converged with the AI-generated summaries over time.
Beyond direct editing, the interface had a significant impact on how users interacted with the recommender system. Participants who edited their interest summaries (categorized as “Interacted” or “Collaborated” groups) showed substantially higher engagement. They viewed and rated more movies, logged in more frequently, and spent longer sessions on the platform compared to users who only read the summaries. These interacting users also showed increased “reflection,” re-rating more movies they had previously scored differently. This suggests that empowering users to co-author their profiles leads to deeper and more varied interaction with the system.
Also Read:
- AI’s Role in Peer Review: A Framework for Systematic Evaluation
- Decoding Robot Mistakes: How Human Reactions Can Guide AI
The Value of “Productively Wrong” AI
The study also revealed a persistent gap between users’ self-perceived interests and the AI’s interpretation. Many participants felt the initial AI summaries weren’t perfectly accurate, sometimes “overfitting” to sparse data or missing nuanced preferences like co-viewing habits. However, the researchers argue that this isn’t necessarily a failure. Instead, a “productively wrong” starting point can serve as a catalyst for self-reflection and engagement. When the AI’s summary is slightly off, it prompts users to intervene, correct, and clarify their preferences, providing valuable explicit feedback that is often scarce in recommender systems.
This approach fosters greater algorithmic agency, allowing users to actively manage their digital identities rather than passively accepting AI-generated labels. It moves beyond AI as a mere task-execution tool to one that supports mutual understanding and co-management of goals, paving the way for more transparent and trustworthy recommender experiences.
In conclusion, this research demonstrates that a human-AI collaborative interface for user profiles can significantly enhance user self-awareness and lead to more engaged and reflective interaction with recommender systems. It highlights the potential of designing AI as a collaborative partner that scaffolds user reflection and control, rather than a perfect expert.
