X will open-source the recommendation algorithm and reveal the labels used by the ranking systems
The social network X is expanding access to the open-source code of its recommendation systems, including the “For You” feed, and is gradually rolling out a transparency tool. This will allow users to check whether the ranking systems have applied labels to their accounts or posts, according to TechCrunch.
The source code for the “For You” feed—which users see by default when they open the app—will be available on GitHub under the Apache v2 license. The company will also provide more detailed information about model configurations, filters, and the core ranking system. In particular, the code will include parameters that determine the weight of various signals during post selection. As a result, the codebase will grow by approximately 10–15 times.
X Vice President of Product Kit Coleman stated that users and researchers will be able to view the code that selects posts, ranks them for a specific user, and generates the feed. Systems will also be available to filter out potentially problematic content or material that may violate the platform’s rules. Some components, including the ranking algorithm and the post evaluation system, will be able to run outside the company.
For more breaking news, follow the UA.News Telegram channel.
A new tool in the “Under the Hood” settings section will allow users who have published at least 10 posts in the last month to download aggregated statistics in JSON format. The file will indicate whether certain tags were applied to their account or posts during the previous calendar month.
Initially, the feature will be tested on a limited group of accounts created at least one year ago. Following the pilot launch, X plans to expand access. Developers will also be able to submit code change proposals via pull requests, which X engineers will review for potential implementation.
At the same time, the company is not making all components of the system publicly available. In particular, tools that use Grok to predict potential rule violations in posts were not included in the release. X explained this by the need to prevent malicious actors from circumventing the platform’s rules and spreading spam.
