X AI moderation: the part the author never sees
X previewed labels that explain why its AI flagged a post. The open-source visibility rules show why nobody could see it before: 66 of 117 post rules skip the author.
On this page · 10 sections
- Quick answer
- What was actually previewed, and when
- One widely-read report described a different product
- Why the author is the last person to find out
- The labels already have names
- X already built the explainer. You cannot run it.
- Today, an explanation is a config switch
- What X's own policy pages say about labels
- What to do before the labels ship
- The short version
The short version
- ▸On 10 November 2026 X's new head of safety, Michael O'Herlihy, previewed label designs that would sit under a flagged post and say what the AI detected, how exactly the reach was limited, and how to appeal. Adweek saw the designs. Nothing has shipped, and X has published no date.
- ▸The reason no one could see the problem before is in X's own published code. Of the 117 `audience:` declarations in the post rules of `xai-org/x-algorithm`, 66 are `NonAuthors`. When a rule blurs or drops your post, you are explicitly outside the audience it applies to, so your own view stays clean.
- ▸The labels already have names. `tweet_rules.rs` names 32 distinct safety label types, including `NSFW_HIGH_PRECISION` and `GORE_AND_VIOLENCE_HIGH_PRECISION`. Those two are exactly the categories behind the rocket-launch misfire Adweek opened with.
- ▸X already built the explainer. `visibility-filtering/why.rs` and the `vf-why` binary take a post id and print the deciding rule for every safety level. It is a staff command line tool that needs a datacenter, not a user feature.
- ▸Whether you get any explanation at all is a config switch today. A namespace of `NoCopy` in `limited_actions_copy.rs` returns no prompt, and `Class::BareDrop` is a first-class verdict meaning a drop that carries no reason.
Quick answer
X AI moderation works in two steps: a classifier attaches a named safety label to your post, then a separate rules engine decides what that label costs you. On 10 November 2026 X's head of safety previewed labels that would finally show you both halves plus an appeal. They are not live. Until they are, the reason you cannot see the damage is structural: of the 117 audience: declarations in X's published post rules, 66 read NonAuthors, so the author is deliberately outside the audience of the rule that is limiting the post.
Last updated: November 2026
Everything below the news is checked against the live xai-org/x-algorithm tree and two X help pages, read on 14 November 2026.
What was actually previewed, and when
Michael O'Herlihy is a Twitter veteran who recently rejoined X (formerly Twitter) as head of safety. On Tuesday 10 November 2026, at the Brand & AI Safety Summit in New York, he showed designs for a label that would sit underneath a flagged post and carry three things: what the AI detected, how exactly the post's reach was limited after being flagged, and a way to appeal. Adweek published the designs the same afternoon under the headline "X Previews Labels That Tell Users Why Its AI System Flagged Their Posts".
Kendra Barnett, who wrote that piece and posted the screenshots, framed the appeal narrowly: you appeal by telling X what the model got wrong. Her example is the one worth remembering. The same rocket launch image gets tagged graphic violence if the AI reads the exhaust plume as an explosion. Adweek's lede is a sibling of it: X's system once labelled a SpaceX launch as adult content, and nobody, including the person who posted it, could see why.
O'Herlihy's pitch to the room was a "glass box": transparency paired with independent verification, and an argument that AI-led moderation leaves an auditable trail that human review never did. That is a brand safety sales pitch as much as a product plan, which is the honest frame for a company whose ad revenue has roughly halved since 2022.
One widely-read report described a different product
Social Media Today covered the same preview on 11 November 2026 and described labels that "indicate whether AI was used to generate an update". That is a different feature in a different category: provenance labelling for synthetic media, the thing most platforms call "Made with AI".
That preview points the opposite way. These labels are about X's AI reading your post, not about an AI having written it. If you searched this topic in the last few days and came away thinking X is about to start tagging AI-generated images, that is where it came from. Both products are plausible. Only one was on the slides.
Why the author is the last person to find out
This is the part no coverage has connected, and it is checkable in public code.
X's visibility system is open source at xai-org/x-algorithm. The post rules live in one file, visibility-filtering/rules/tweet_rules.rs, and every rule declares who it applies to. The split across its 117 audience: declarations:
| Audience | Declarations |
|---|---|
NonAuthors | 66 |
Everyone | 40 |
NonFollowers | 10 |
OnlyAuthor | 1 |
A rule with audience: NonAuthors does not fire for you. Take the live NSFW family. NSFW_HIGH_PRECISION_BLUR_NUDITY blurs the post behind an interstitial, NSFW_HIGH_PRECISION_DROP_NSFW_MEDIA removes it from the surface entirely, and several ..._TOMBSTONE_... variants replace it with a notice in specific countries or on clients that cannot render a blur. Every one of them is NonAuthors.
So the post looks normal in your own timeline, your own profile and your own notifications, while a stranger gets a grey box or nothing. That is not a bug or a hidden penalty. It is the stated audience of the rule. The tooling to notice it from outside is thin, which is why the practical check is still social: have someone who does not follow you open the post logged out. Our write-up of why X views stay low covers the ranking side of the same blind spot, where the ceiling is a feed slot rather than a label.
There is one counter-current worth noting. GORE_AND_VIOLENCE_HIGH_PRECISION_BLUR_ALL_USERS carries audience: Everyone, gated on a client switch called ModernBlur. Where modern blur is on, the author does see the interstitial. X has started closing this gap in the newer rules, one family at a time.
The labels already have names
tweet_rules.rs names 32 distinct SafetyLabelType values. Counted by how often they appear in rule conditions, the top of the list is NSFW_HIGH_PRECISION (14), GORE_AND_VIOLENCE_HIGH_PRECISION (9) and NSFW_CARD_IMAGE (8). Below those sit SPAM_HIGH_RECALL, LOW_QUALITY, DO_NOT_AMPLIFY, DOWNRANK_SPAM_REPLY, RISKY_HIGH_VIZ_REPLY, DUPLICATE_CONTENT, DUPLICATE_MENTION, MALICIOUS_URL, UNTRUSTED_URL, BOUNCE, PDNA, five FOSNR_ conduct labels, and four misinformation variants: civic, medical, crisis and generic.
Read the two names at the top again next to the rocket launch. HIGH_PRECISION is the model family's own claim about itself, and a plume misread as an explosion is precisely a precision failure. The preview is X proposing to print the classifier's guess on the post that the guess just cost.
Provenance is recorded too. SafetyLabelMap keeps a by_agent set for labels whose source is a tool action, and exposes is_by_agent() so a caller can ask whether an agent applied the label rather than a human or a batch job. No rule in the published tweet_rules.rs reads it yet. The hook for "an AI did this to you" exists before the label that would say so.
X already built the explainer. You cannot run it.
X's repository ships visibility-filtering/why.rs and a binary called vf-why. You give it a post id, optionally a viewer id, a country and a client, and it evaluates every safety level and prints, per level:
<level> (<rpc>): <class>/<reason>
rule: <the rule that decided>
viewer: <logged_out | not_read | found | no_user | lookup_failed>
failed lookups: <...>
fail-open defaults: <...>
That is the previewed label, in text, already written. The catch is the first line of main: it reads a DATACENTER environment variable and connects to live internal data sources. It is a staff tool.
One level up, the same pattern repeats. visibility-filtering/rules/catalog.rs generates a document titled "Policy catalog" that lists every rule with its audience, its conditions spelled out in words and its action, sorted by severity, and a test asserts the generated file matches the rules. The generator is public. The generated catalog, and the GLOSSARY.md it cross-references, are not in the tree. X publishes the machine that prints its rulebook and withholds the printout.
If you want the rest of what that repo does and does not show, we went through it in the open-source X algorithm.
Today, an explanation is a config switch
When X does limit an action rather than hide a post, the user-facing text comes from params/limited_actions_copy.rs. It covers 26 LimitedActionType values, which is the real list of what X can switch off on a single post: reply, repost, quote post, like, react, send via DM, bookmark, add to Moment, pin to profile, view post activity, share via, follow, list add or remove, mute conversation, embed, view hidden replies, hide community post, copy link, vote on poll, remove from community, show repost menu, reply downvote, autoplay, edit post, highlight and view post engagements.
Each one looks up a headline and a subtext from a translated string bundle, keyed as namespace plus action plus _Headline or _Subtext. Three details matter more than the list:
- A namespace of
NoCopyreturns nothing. No headline, no subtext, no explanation rendered. - Whether there is a "Learn more" link at all depends on a configured URL being non-empty.
result_label.rsdefines seven verdict classes, and one of them isBareDrop: a drop that carries no reason. It is treated as rendering the same as a reasoned drop for every reason but nine named ones.
So the current state is not that X cannot explain itself. It is that the explanation is an optional string, and "no reason" is a supported value.
What X's own policy pages say about labels
Two X help pages disagree, and both are easy to check.
The old synthetic and manipulated media URL now redirects permanently to the Authenticity policy, dated April 2025. Its synthetic media section is strict about fabricated media of real people, but notice what its enforcement list contains: anti-spam challenges, denylisting URLs, restricting reach, temporary loss of features, profile modifications and suspension. Labelling is not on the list. Appeals get one line and one link.
The separate Notices on X page, archived 11 October 2026, still documents a "Manipulated media" label, and adds a limitation worth copying down: violation notices appear on the post detail page and the profile, and are "not available" anywhere else, including the Home timeline, notifications and search. A label nobody scrolls past is not transparency either.
Two more things follow from the policy text. Plain AI involvement is not a violation on its own under the Authenticity policy, so an AI label would be information rather than a strike. And nothing in either page describes a hidden account-wide penalty state. X restricts reach per rule, which is why we use visibility limiting rather than the folk term, and why X's own report is the first place to look. We cover what it shows and the gaps it leaves in the X shadowban check.
What to do before the labels ship
Four things, in order of payoff.
- Check from the outside. A logged-out window and a non-follower account will catch a
NonAuthorsrule that your own timeline hides. Do it on any post that underperforms badly rather than mildly. - Treat media as the risk surface. Three of the four most-used labels in the rules are media labels. Thumbnails, card images and video stills get classified, and a plume, a wound, skin tone or a dense red frame is what the model sees.
- Appeal through the documented route, not the preview. The in-app appeal linked from the Authenticity policy is the only live path, and contacting X support is worth pairing with it when a post is business-critical.
- Keep your posting cadence human. Labels like
DUPLICATE_CONTENT,DUPLICATE_MENTIONandDOWNRANK_SPAM_REPLYare triggered by patterns, not by content, which is the one part of this entirely in your control. It is also why we run X-Autopilot on your own machine, in your own browser session, at human pacing: a lower detection surface than a shared cloud IP pool, not a guarantee of anything. Browser automation on X is a gray area and we will not tell you otherwise.
The short version
X moderates with AI and has stopped pretending it does not. The honest summary of the 10 November preview is that X proposes to show you the classifier's guess, the reach it cost you, and a box to say the model was wrong. The machinery to do it has been in public for months: a rules file that names the labels, a why binary that names the deciding rule, and a catalog generator that spells out the whole rulebook in words. What has been missing is the decision to point any of it at the author, and 66 NonAuthors declarations are a reasonable measure of how far that decision has to travel.
Frequently asked
Answers indexed by Google + AI assistants.
Does X use AI to moderate posts?+
Yes, and X now says so openly. At the Brand & AI Safety Summit in New York on 10 November 2026, head of safety Michael O'Herlihy argued that AI-led moderation produces auditable records that human review does not, and previewed labels that would expose the model's reasoning to the person who posted. The mechanism underneath is a label: a classifier attaches a named safety label to a post, and a separate rules engine decides what that label does to the post's reach.
Why did X flag my post as sensitive when it is not?+
Because the classifier works on the media, not on your intent, and the two labels most often involved are named `NSFW_HIGH_PRECISION` and `GORE_AND_VIOLENCE_HIGH_PRECISION` in X's published rules. Adweek's example is a SpaceX rocket launch that X's system labelled adult content. Kendra Barnett, who saw the preview, described a second version where the same launch reads as graphic violence because the model takes the exhaust plume for an explosion. High precision is the design goal, not a guarantee.
Can I see why my post's reach was limited?+
Not in full, not yet. X ships a visibility report called Under the hood that tells you whether your account carries a restriction, and the previewed labels would add the per-post reason. Today the strongest signal available to you is indirect: ask someone who does not follow you to open the post in a logged-out browser, because 66 of the 117 audience declarations in the public post rules exclude the author from the rule they trigger.
How do I appeal an X moderation decision?+
The appeal route X documents today is the in-app appeal form linked from the Authenticity policy, which covers account-level and content-level enforcement. The previewed labels would add a per-post appeal that works by telling X what the model got wrong, which is a different thing: feedback on a classifier rather than a review of a rule. Until that ships, an appeal plus a support thread is the only path.
What are X's safety labels called?+
The public rules name 32 of them. The most used are `NSFW_HIGH_PRECISION`, `GORE_AND_VIOLENCE_HIGH_PRECISION` and `NSFW_CARD_IMAGE`, and the set also includes `SPAM_HIGH_RECALL`, `LOW_QUALITY`, `DO_NOT_AMPLIFY`, `DUPLICATE_CONTENT`, `DOWNRANK_SPAM_REPLY`, `RISKY_HIGH_VIZ_REPLY`, `MALICIOUS_URL`, `UNTRUSTED_URL`, `BOUNCE`, and four misinformation variants for civic, medical, crisis and generic claims.
Are the new AI transparency labels live on X?+
No. O'Herlihy previewed designs at a summit on 10 November 2026. X has not announced a ship date, has not published a help page for them, and has not added them to its Notices on X page, which still documents only the Manipulated media label. Treat them as a direction of travel, not a feature.
- Adweek, 'X Previews Labels That Tell Users Why Its AI System Flagged Their Posts' by Kendra Barnett (published 10 November 2026, 17:18 ET): the SpaceX rocket launch labelled adult content with nobody including the author able to see why, Michael O'Herlihy described as a Twitter veteran who recently rejoined X as head of safety, the Brand & AI Safety Summit in New York on Tuesday 10 November, and the three things the previewed labels would show - what the AI detected, how the reach was limited, and an appeal
- Kendra Barnett on X, posts 1 and 2 of her thread from the summit (posted 10 November 2026, 22:36 UTC): the three things each label shows, and the worked example of a rocket launch tagged graphic violence when the AI reads exhaust as an explosion
- xai-org/x-algorithm, `visibility-filtering/rules/tweet_rules.rs`: 117 `audience:` declarations split 66 NonAuthors, 40 Everyone, 10 NonFollowers and 1 OnlyAuthor, 32 distinct `SafetyLabelType` values, and the NSFW_HIGH_PRECISION and GORE_AND_VIOLENCE_HIGH_PRECISION rule families with their blur, drop and tombstone actions (read 14 November 2026)
- xai-org/x-algorithm, `visibility-filtering/why.rs` and `visibility-filtering/bin/vf-why.rs`: the internal explainer, its `Ask` arguments of post, viewer, country, client and level, and the exact output of class, reason, rule, viewer resolution, failed lookups and fail-open defaults per safety level (read 14 November 2026)
- xai-org/x-algorithm, `visibility-filtering/params/limited_actions_copy.rs`: the 26 `LimitedActionType` values, the StringCenter key scheme of namespace plus action plus Headline or Subtext, the `NoCopy` namespace that returns no prompt, and the Basic, SeeConversation and LearnMore prompt kinds (read 14 November 2026)
- xai-org/x-algorithm, `visibility-filtering/result_label.rs` and `visibility-filtering/rules/catalog.rs`: the seven verdict classes including BareDrop, and the Policy catalog generator whose output file and GLOSSARY.md are not in the public tree (read 14 November 2026)
- X Help Center, 'Authenticity' (policy dated April 2025, latest archived copy 11 November 2026): the destination of the old manipulated-media URL, the synthetic and manipulated media rules, and an enforcement list of anti-spam challenges, denylisting URLs, restricting reach, temporary loss of features, profile modifications and suspension that never mentions a label
- X Help Center, 'Notices on X and what they mean' (archived 11 October 2026): the Manipulated media label, and the statement that violation notices are not available outside profiles and post detail pages, including the Home timeline, notifications and search
Product designer and indie hacker. Runs the agent on his own X account every day and writes up what the data shows, including when it's inconvenient.
Follow on X →