Meta rolls out LLM-powered feed controls on Facebook

Facebook users can now type natural-language instructions to shape their algorithmic feeds, a shift that moves Meta away from relying solely on passive behavioral signals to determine what appears in recommendations.

The new personalization interface, accessible through a “your algorithm” menu attached to individual feed posts and Reels, allows people to add or block specific topics. A text field invites users to describe their preferences conversationally rather than toggling preset categories. Meta confirmed the system processes these inputs using large language models integrated directly into its ranking infrastructure.

Unlike traditional recommendation engines that infer interest primarily from clicks, watch time, and dwell duration, the updated architecture parses intent from written prompts. This means a user typing “stop showing me cast iron cookware videos” provides a semantic signal the model can map to content clusters, bypassing the ambiguity of engagement metrics where lingering on a post might previously have been misread as interest.

The deployment follows an identical rollout on Instagram earlier this year. Meta has spent the last two years aggressively unifying the underlying codebases for Facebook and Instagram recommendations, meaning features tested on one platform routinely migrate to the other once scaled. Both networks now run on a shared AI discovery engine designed to surface content from accounts users do not follow, a strategy that significantly increased the volume of algorithmically injected posts and made manual feed curation more difficult.

Practical limitations observed during the Instagram launch raise questions about how much control Facebook users will actually wield. The text prompt does not currently accept requests to suppress advertising inventory or prioritize chronological posts from followed accounts. Because ad delivery operates on a separate auction-based system governed by Meta’s business tools rather than its consumer recommendation models, user-generated text commands cannot override paid placements.

The update arrives as Meta faces mounting pressure over the saturation of AI-generated imagery and low-effort engagement bait across its platforms. In 2024, the company introduced specific policies targeting synthetic media spam and manipulated content farms, but enforcement relies heavily on automated detection systems that frequently lag behind new generation tools. Whether the new text-prompt feature will recognize and act on broad categorical requests—such as filtering out all AI-generated images or suppressing cross-posted forum threads—remains unconfirmed. The system’s effectiveness depends entirely on how Meta’s engineering teams have bounded the LLM’s interpretation scope within the ranking pipeline.

The feature is rolling out gradually and is not yet visible to all Facebook accounts globally.