X opens algorithm components, offers users insight into visibility labels

Social media platform X has made more of the code behind its recommendation and visibility systems publicly available. In parallel, the company is giving a pilot group of users access to information about labels that can impact their posts' reach, offering greater transparency into the platform's content distribution.

The expanded release includes configuration parameters, ranking weights, and visibility-filtering code, along with the code used to train and operate its "Phoenix" recommendation model. This update significantly broadens the scope compared to a similar open-source release in 2023. The new "Under the Hood" feature, located in account settings, allows eligible users to download aggregate data on labels applied to their accounts and posts that may restrict visibility.

X distinguishes between content ranking and visibility filtering. The "Phoenix" model determines the order of posts in the "For You" feed, while separate visibility systems decide whether a post is shown, suppressed, or placed behind an interstitial. The published code details these filtering systems and the labels they utilize. While users often refer to unexplained reach decreases as "shadowbans," the "Under the Hood" tool addresses specific visibility labels rather than acting as a direct shadowban detector.

The "For You" feed mixes content from followed accounts and new accounts. "Phoenix" predicts user engagement probabilities based on past interactions and combines these with other signals using weights defined in the released code. These weights cover actions like likes and shares, as well as negative signals such as reporting or blocking. The published code reveals a mathematical component of the ranking system, not a simple formula for reach changes, as weights apply to predicted probabilities, not raw engagement counts.

While the release of code and label information provides researchers and users with more data for analysis, it does not offer a complete explanation for individual post visibility fluctuations. Researchers can now inspect the system's structure, but cannot necessarily reproduce precisely why a specific user saw a particular post at a certain moment. The "Under the Hood" tool provides aggregate label data but doesn't fully explain label triggers or their precise impact on reach.