The Pulse
Meta details AI controls for Facebook recommendations
Meta's Facebook tool lets people add, remove and move interests used to shape Feed and Reels recommendations. Language models help interpret preferences and check recommendations, while Meta notes that handling ambiguous requests remains an
AI.info Team · Written with AI · fact-checked ·
Editable interests for Feed and Reels
Meta’s engineering blog published details on Nov. 10 of an AI-powered system that gives people ways to shape recommendations on Facebook. The company calls the system Your algorithm and says it is designed to make recommendations more transparent and steerable.
The tools let people describe what they want to see in Feed and Reels using their own words. The post contrasts that approach with recommendation systems that have traditionally learned from implicit engagement signals, which Meta says can adapt slowly and capture little nuance.
The interface begins with a short “Lately Summary” of recent viewing activity, followed by “Interested In” and “Not Interested In” lists. The first list is prefilled with interests inferred from past activity, while the second includes topics people have marked as unwanted, hidden or removed.
People can add interests in natural language, shift them between lists or remove them. The changes apply to both Feed and Reels and can be updated further through feedback on individual posts.
How preferences affect recommendations
The system has several components handling a preference from its initial creation through ranking and delivery. Profile generation creates an initial interest profile, preference management checks edits and stores them, ranking applies the two lists, and delivery optimization, which Meta says is intended to make changes take effect more quickly, handles delivery.
For a new interest, the system uses four retrieval sources to find potential recommendations. Two look for matching content directly, while two seek related material through content connections and item-to-item similarity; candidates that pass a relevance check can then receive a boost later in ranking.
The system aims to give a new interest visibility without letting it dominate recommendations, Meta says. The system spaces out items about the same interest, limits each interest’s daily contribution and retrieves candidates for only one interest per request; new interests receive extra weight for a limited period.
For topics on the “Not Interested In” list, the system filters or demotes matches at multiple points in the recommendation process. Requests about qualities such as AI-generated or sexually suggestive material use dedicated demotion signals, while content from a named creator is filtered directly.
Language models and delivery checks
Large language models help create the initial profile, validate changes and check recommendations against unwanted interests. For edits, a model checks whether a phrase is a coherent topic, spots duplicates and maps it to a standardized topic name; the system also runs separate integrity and safety checks.
A fine-tuned language model reviews ranked recommendations as a final check for content related to recently added unwanted interests. Meta says it found that a small model trained to identify items for removal was faster and more reliable than a general-purpose model judging each item individually.
The system also includes ways to make preference changes affect content already being fetched or stored on a device. Adding an interest can trigger an extra request for recommendations outside the normal fetch cycle; adding an unwanted interest prompts the client to remove matching items already in its cache before they appear.
A separate model grades outputs against expected behavior, including whether summaries are accurate and unwanted recommendations are removed. Meta says it is still working on interpreting ambiguous input more effectively and is exploring additional places where people could adjust preferences.