Inside X’s newly published visibility filtering code

X announced both on Thursday 13 August 2026 through its @XOpenSource account. The code sits in the same public repository the company has been adding to since January, github.com/xai-org/x-algorithm, under the Apache 2.0 licence. The xai-org address is there because xAI bought X in 2025. The repository had passed 27,000 stars and 4,600 forks by Friday morning.

Previous releases explained how posts get scored and sorted. This one covers the filters that run afterwards, the ones that decide a post should not be recommended at all, and the labels that trigger them.

What “shadowban” actually means here

X has never used the word shadowban in its own rules. The internal term is visibility filtering, and the released code now shows what it does. The system attaches labels to posts and to accounts, and each label carries a defined effect.

The most common effect is that a post is hidden from recommendations to people who do not follow you. Your followers still see it. It stops reaching strangers through the For You feed. That is the mechanism most people mean when they say their reach collapsed for no visible reason.

A second group restricts a post so it can only be found on the author’s profile. Posts labelled for abuse and harassment, hateful conduct, violent speech or civic integrity fall here. X’s file says all users can see a notice explaining that the post has limited visibility, so this one is not silent.

A third group removes the post from X entirely, either because it was flagged as spam, or because it is pending review, or because X wants the author to delete it.

Adult content labels work differently again. They put the post behind a content warning and hide it from users who are underage, users who have not stated an age, and anyone browsing while logged out.

Account-level labels apply the same treatment to everything you post. The published list includes labels for accounts detected as compromised, as likely spam sources, as impersonating someone, or as primarily posting adult content.

The full list is in a single file, underTheHoodLabels.strato, which we read directly. It carries 18 post labels and 12 account labels, each with a plain-English description of what it means and what it does. One caveat sits in the file itself. It describes its contents as a public allowlist, which is the set of labels X will display in your report. The file does not claim these are the only labels the platform applies.

The Under the Hood page

The companion tool lives at x.com/i/under_the_hood. You need to have posted at least 10 times in the previous month to see anything. X says the pilot is limited to a randomised group of accounts that are at least a year old.

What you download is a JSON file covering the past calendar month. X’s own description says it gives aggregate statistics on the labels applied to your account and posts, rather than a line-by-line audit of every post you published. Karissa Bell at Engadget flagged this gap on the day of launch, and also noted that the raw file is hard to read for anyone who does not work with code.

X’s suggested workaround is to feed the file and the GitHub repository into an AI chatbot and ask it to explain the result. Keith Coleman, X’s VP of Product, told TechCrunch that outside researchers were given the code in advance and managed to get the scoring system running outside the company.

What X did not release

Coleman told TechCrunch’s Sarah Perez that the systems using Grok to predict whether a post breaks a rule are held back, to stop people reverse-engineering their way past spam controls. The repository also says X withholds detail that could be used to game the algorithm.

That falls short of what Elon Musk promised. On 24 July he posted that within a month, every line of code touching X would be open source and third-party audited. This release is a large step, not that.

Why X keeps doing this

The company has been under regulatory pressure on exactly this point. On 5 December 2025 the European Commission fined X 120 million euros, about KES 17.9 billion at mid-August rates, in the first non-compliance decision issued under the Digital Services Act. The breaches concerned the paid blue checkmark, the advertising repository, and blocked access to public data for researchers.

The older argument still runs underneath it. American politicians spent years accusing Twitter of suppressing conservative accounts, the company denied it, and Musk bought the platform partly on that complaint. Publishing the filter code is the strongest answer he can give to it.

Meta, TikTok and YouTube have not published their ranking code.

What this changes for Kenyan users

Reach on X is now money for a growing number of Kenyan accounts. We covered X scrapping Creator Revenue Sharing on 7 September, with everyone required to reapply and a threshold of 500,000 verified Home Timeline impressions over 90 days. Those impressions come from the same For You feed these labels control. A single account-level label that hides your posts from non-followers takes a direct bite out of the number that decides whether you get paid.

The people most affected are also the ones already paying. X Premium in Kenya runs from KES 259 a month to KES 5,160 a month depending on tier. Until now, a subscriber whose reach dropped had no way to tell whether it was a label, a change in the ranking model, or ordinary variation. The labels file at least names the possibilities.

One label deserves a closer read here. FOSNR_CIVIC_INTEGRITY restricts a post so it can only be found on the author’s profile, and X’s file says it is applied in response to a user report. Kenya’s election-period speech online is already contested ground, and this is now a named, documented mechanism rather than a rumour.

One older complaint goes unanswered. We have written about X pushing region-locked posts onto Kenyan timelines that users cannot watch or reply to. Nothing in this release addresses geography.

What to do with this

If your reach has fallen, open x.com/i/under_the_hood. You are either in the test group or you are not, and there is no way to request access yet.

If you are in it, the entry to look for is any label whose effect reads “hidden from recommendations to non-followers.” That is the one that decides whether strangers ever see you. The rest either show you a visible notice already or remove the post outright.

X said it will keep updating the repository as the algorithm changes, so the label list can move. Musk’s promise of a fully audited codebase still carries an August deadline, with about two weeks left in the month.