Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations

Meta's engineering team detailed Your algorithm, a set of tools that shows Facebook users an editable view of the interests driving their Feed and Reels. There is a one-line summary of recent viewing, an Interested In list prefilled from past behavior, and a Not Interested In list. People can add, move or remove topics in natural language, and changes apply across both surfaces.

For years, recommendation systems have inferred what people want from what they click. This flips part of that around. Your stated wishes become a direct input, enforced at several stages of the ranking system.

We think this is one of the more genuinely user-respecting uses of large language models we've seen from a big platform. Meta is using LLMs to do housekeeping that makes control practical: turning activity into readable topic names, checking that a typed phrase is a coherent topic, catching duplicates, and running a final pass to strip out content that matches something the person said they don't want. That is AI used to serve a stated preference rather than to guess around it.

The details suggest Meta took the 'not interested' side seriously. Blocks are enforced during ranking, again by a fine-tuned reranking model, and again on the device by dropping cached items that were queued before the change. Requests like seeing less AI-generated or sexually suggestive content map to dedicated demotion signals. When a person names a creator, that creator is filtered directly. Those are the requests users have been making for years, and it is good to see them treated as hard rules rather than hints.

We do have reservations. The Interested In side is deliberately throttled: new interests get extra weight for a limited period, then the system goes back to learning from engagement. That is a reasonable way to avoid flooding a feed, and also a reminder that behavior still wins in the long run. Meta also notes it evaluates these LLM calls with another LLM acting as judge, which is useful for iteration but not the same as independent measurement of whether people actually feel more in control.

And there's a trade-off nobody should ignore. An editable, plain-language interest profile is also a very clean record of what a person cares about. Meta's post is about ranking, not ads, and doesn't say how these stated interests are used elsewhere. We'd like that answered plainly. Steerable feeds are a real improvement. They become a better deal if the steering data stays in the feed.

The system has four parts: profile generation, preference management that writes to a single store read by both Feed and Reels, ranking that applies both lists across the funnel, and delivery tweaks such as an extra fetch when an interest is added and cache invalidation when one is blocked. Meta says next steps include better handling of ambiguous input and more places to shape preferences.