Recommendation settings become conversational
Meta is rolling out new Facebook controls that let people describe what they want to see in Feed and Reels using ordinary language. Announced on 13 October, the changes add a dedicated Your algorithm area for managing topics and expand the feedback available after selecting Interested or Not interested. Meta says recommendations can respond in real time, including during the same session.
Traditional preference controls often reduce a complex interest to a single tap. Freeform input can express a more specific intention: a person might want more practical gardening advice but fewer promotional videos, or local football analysis without transfer rumours. A large language model can interpret those distinctions alongside existing engagement signals, giving the ranking system a richer instruction than a binary reaction.
Your algorithm exposes the topic layer
The new Your algorithm space is available from the three-dot menu and allows people to add or manage topics they want represented in Feed and Reels. This creates a visible preference layer over systems that have historically inferred interests from behaviour such as likes, comments and watch time. Direct input should make it easier to correct an inference without waiting for months of changed behaviour.
Visibility does not necessarily mean complete control. Meta describes the feature as fine-tuning recommendations, not replacing every ranking signal or guaranteeing a particular item. Users will need clear feedback about whether a topic was accepted, how long it remains influential and whether the setting applies across Feed and Reels. A control is most trustworthy when its effect can be observed and reversed.
Follow-up questions add nuance
Facebook may now ask for more detail after a user marks a post as interesting or uninteresting. The person can explain what they liked or disliked in their own words, and the recommendations update accordingly. This is a substantial interface change: feedback becomes a short conversation with the ranking system rather than a hidden signal inferred from the initial click.
The quality of that experience will depend on interpretation. Sarcasm, negation and mixed preferences can confuse language systems. Meta should avoid presenting an inferred preference as a certainty and provide an easy route to correct it. People will also want assurance that sensitive statements entered as ranking feedback are handled under clear privacy and retention rules.
LLMs join existing ranking signals
Meta says large language models help the system understand the interests, context and intent contained in freeform text. That information is combined with what Facebook already knows about the person’s engagement. The result is not an LLM choosing every post by itself, but language understanding feeding a broader ranking system with thousands of existing signals.
This hybrid approach can be more practical than replacing a mature recommender. Existing systems remain good at predicting engagement from behaviour, while language adds an explicit channel for goals those systems cannot safely infer. The risk is that an eloquent preference box may imply more agency than the ranking pipeline actually grants. Transparent explanations and measurable changes will be important.
A test of meaningful user agency
The most useful evaluation is whether people can reliably steer their experience. Meta could measure whether stated preferences appear in recommendations, how often users revise them and whether unwanted topics decline without collateral changes. Independent researchers will also be interested in whether controls work consistently across languages, regions and sensitive subject areas.
For advertisers and creators, more explicit preference signals may shift which audiences see content, but the release is primarily a consumer-control story. Facebook is acknowledging that inferred behaviour is not always enough to represent intent. If Your algorithm remains understandable, reversible and responsive, natural-language feedback could become a more humane interface to recommendation systems that have long felt opaque.
The wording of control matters
Natural-language settings can lower the barrier to participation for people who never visit a conventional preferences page. They can also create accessibility challenges if the product depends on long explanations or subtle wording. Facebook should support concise instructions, examples and multilingual interpretation, while ensuring that an accidental sentence does not permanently reshape a feed without a visible record.
A useful history could show the topics a person added, the feedback the system inferred and a simple option to reset recent changes. That would turn personalisation from an invisible accumulation of behavioural data into something closer to an editable profile. The feature’s success will be defined less by the sophistication of its language model than by whether ordinary people can predict and control the outcome.
Meta will also need to manage adversarial feedback. Creators may encourage followers to paste carefully worded instructions designed to increase a topic’s distribution, while coordinated campaigns could try to manipulate stated interests at scale. Account integrity controls and rate limits remain relevant even when input appears to be a personal preference. The product should protect an individual’s control without turning that control channel into a new ranking exploit.