X Under the Hood Decoder

Adult content, high confidence

An automated classifier is confident the post contains adult nudity or sexual content. The post stays up, but it is held back from For You, search and trends for anyone who has not opted into sensitive media.

Fix: If the post really is adult content, mark your media as sensitive by default in Settings, Privacy and safety, Your posts. That keeps the label from being read as an attempt to evade filtering. If the call is wrong, use the appeal link on the post, since a manual review is the only route that removes an automated adult-content label.

Adult content, wide net

The same adult-content question asked with a much lower bar for saying yes. High recall means X would rather catch everything and be wrong sometimes, so this label carries far more false positives than the high precision one.

Fix: This is the label most worth appealing, because the false-positive rate is the whole point of the high recall setting. If several of your posts carry it and none are adult, appeal them individually rather than assuming the account is done for.

Adult content, from reports

Applied because people reported the post as adult content and a heuristic agreed, rather than because a model flagged it on its own.

Fix: Appeal it, and look at whether the post reached an audience it was not written for. Posts that land in front of people with no context attract reports that the same post inside your normal audience would never get.

Sexually explicit text

The text of the post was read as sexually explicit. No image is required. A text-only post can carry this.

Fix: Rewrite the wording if the flag is fair. If it is not, appeal. Deleting and reposting the same text usually gets the same label back and adds a duplicate-content signal on top.

Adult content in the link preview

The flagged image is the link preview card, not anything you uploaded. X pulled the og:image from the page you linked to and that image tripped the adult-content classifier.

Fix: Change the og:image on the destination page if it is yours, then repost so X fetches a fresh card. If the page is not yours, post the link without the card, or link to a different page on the same site.

Media marked sensitive

Reads as the ordinary sensitive-media marker: the media is behind a click-through warning. That is a filtering state, not a strike.

Fix: If you set it yourself, nothing to fix, this is the system working as intended. If you did not, check Settings, Privacy and safety, Your posts, and turn off the default sensitive marker if it is on.

Graphic violence, high confidence

The media was classified as graphic violence or gore. Recommendation surfaces drop the post.

Fix: For newsworthy footage, mark it sensitive yourself and say in the post what it shows. That is the treatment X asks for, and it is the strongest position to appeal from.

Graphic violence, from reports

Reads as the reports-driven twin of the graphic violence label, applied after users reported the post rather than from the model alone.

Fix: Appeal, and add context to the post. Graphic media with no explanation attracts reports even when it is legitimate.

Spam, wide net

A broad spam classifier matched the post. High recall again means a low bar, so ordinary posting patterns can produce it. This is the label most often behind a sudden unexplained drop in reach.

Fix: Stop the pattern before you appeal, because the classifier scores behaviour over a window rather than one post. Space posts out, vary the wording rather than reposting the same text, use one link per post, and drop link shorteners for a direct domain.

Duplicate content

The post is the same as, or very close to, something already on the platform. That can be your own earlier post or someone else, and the label lands on the copy either way.

Fix: Rewrite rather than repost. Change the opening line, the framing, or the example, so the recycled post is a variant instead of a copy. If you recycle on a schedule, put a real gap between runs and vary the text each time.

Automated behaviour suspected

Reads as X flagging the posting behaviour, not the content. Scheduling is allowed, so this is about the shape of the activity looking machine-generated.

Fix: Add jitter to the schedule so posts do not land on the exact minute every time, break up long automated runs, and never send the same reply text to more than a handful of accounts. If you post through an API or an agent, throttle it.

High toxicity score

Reads as a toxicity classifier scoring the text above a threshold. It is a score, not a policy finding, so it can land on heated but rule-abiding posts.

Fix: Look at whether the post attacks a person or an argument. Rewriting toward the argument usually clears the score. Quoting someone else abusive language can also trigger this, so paraphrase instead of quoting verbatim.

Untrusted link

Reads as a link in the post pointing at a domain X does not trust. The post is held back, and in some cases the link is interstitialed.

Fix: Link the destination domain directly instead of a shortener. If your own domain is flagged, that is a domain-level problem and needs an appeal through X support, not a repost.

Do not amplify

A blanket instruction to the recommendation systems to leave the post out. It is not a takedown, the post is visible to anyone who goes looking, and it is the heaviest of the reach-limiting labels.

Fix: Appeal it, and treat it as the top priority in this file. There is no posting-pattern change that clears it, because a decision produced it rather than a threshold.

Low quality content score

Reads as a quality score rather than a rule violation. Low quality posts are ranked down instead of removed.

Fix: Put the substance in the post rather than behind a link, and cut hashtags to one or two. This is the label that clears on its own fastest once the pattern changes.

Engagement farming

Reads as the label behind X public enforcement against posts that ask for the engagement directly rather than earning it.

Fix: Remove the ask. Ask a question people want to answer instead. This one compounds, since the same behaviour also drives the not-interested signal that costs 43.2 points in the ranking weights X published.

Withheld for copyright

A copyright holder filed a DMCA notice and the post or its media was withheld. This is a legal process, not an algorithmic one.

Fix: X emails the account when a notice lands. A counter-notice is the only route, and it is a legal filing with real consequences. Repeated notices lead to account-level action, so do not repost the media in the meantime.

Freedom of Speech, Not Reach

The FOSNR prefix is X public enforcement approach: instead of removing a post that breaks a rule, X restricts how far it travels and leaves it up behind a notice. The post is excluded from For You, search, trends and recommended notifications, and replies to it are usually collapsed.

Fix: Every FOSNR label has an appeal on the post itself, and the appeal is worth filing because a human review is the only thing that lifts it. While it is applied, do not delete and repost the same content, since that reads as evasion and escalates to account-level action.

X has documented the Freedom of Speech Not Reach approach. It has not published the list of per-policy label names in this export, so the suffix below is read from the key itself rather than from documentation.