A few weeks ago, X cut the reach of my posts and replies to almost zero. They put a label on my account.

I left a comment under one of Naval’s posts, and the interface immediately blocked me from doing anything on X. The number of accounts I was following also dropped to zero. At first, I thought I had been hacked.

But there were no security alerts. After I logged out and back in, all the functions and the accounts I followed returned. Later, in the notifications section, I saw a red flag next to a message: “We’ve added a temporary label to your account which may impact its reach. Learn more here.”

I couldn’t understand what I had done wrong. I hadn’t even sworn in a single reply. Grok suggested it might be a false positive and that a Request Review should fix it.

In my case, nothing was resolved, no matter how many times I pressed that button. From posts and discussions online, though, it was clear that some people had their labels removed within a few hours of requesting a review. For others, it took several days.

I pinged Premium support and eventually managed to get through to a real person. The support specialist explained that their team only deals with Premium features and services, not algorithms or account restrictions. Still, he hinted that the problem might be related to the quality of the posts.

That’s when the thought first crossed my mind: X had begun cutting the reach of accounts whose text looked AI-generated. In my case, AI was used to translate my writing into English and make small editorial changes. Stylistically, you could often feel it.

On the one hand, I was angry at the unfairness of it. X has no problem distributing slop videos and images, especially those generated by Grok. But on the other hand, I was almost glad: if the platform really has learned to distinguish mass-produced, overly polished text from more human writing, then an author’s voice still matters. Maybe more than ever.

Thirty Days as an LLM Slop User

For the first two weeks, I kept desperately pressing Request Review every day, reading comments and similar stories on Reddit, and wondering how long the label would stay on my account.

I was under no illusions that platforms owe authors free distribution. I understand that impressions and reach are ultimately at the platform’s discretion, especially if your account fits current trends and can hold people’s attention. Still, I have to admit that once the label appeared, my distribution dropped sharply. Fewer people were discovering me and reading articles on my site, Verba. The web analytics turned gloomy and slightly alarming.

The whole situation made me think that X has seriously started focusing on quality, while also trying to become more transparent. Clearly, the platform wants users to trust it.

Labels have become widespread, judging by the number of posts and reports from users online. On top of that, X made parts of its algorithm public and launched Under the Hood, which lets users see what restrictions are affecting their accounts and figure out how to fix the problem.

Users are divided. Some welcome the transparency. Others point out that X first rolled out widespread restrictions and only afterwards gave people a way to see why they had been restricted.

Under the Hood only works for accounts older than one year that post at least ten times a month. It also has not been rolled out to everyone yet. There is a random test group, and it seems the people with the strictest labels were more likely to get in.

The report showed that the label had already been attached to my account before I received the red flag in notifications and before my reach was cut.

What the Label Revealed

I was labeled as “SpamHighRecall.” Digging through the published code, especially the components related to enforcement, I found that the algorithm applies this label for 30 days to users it classifies as llm_slop_user.

I also found that a user can receive a 30-day restriction for a single post or several posts. In that case, the user gets the RiskyHighVizReply label and the llm_slop_post category. As I understand it, all other posts, including new ones that are not marked with that label, can still get normal reach.

As you can see, my case was one of the harsher ones. X restricted the entire account at once. The system appears to have decided that the problem was not a specific post, but a broader pattern, effectively classifying me as an author inclined to post LLM slop.

That also made me think that several posts in a row with a similarly “smooth” style, translation plus light editing, may have produced a combined score high enough to trigger llm_slop_user or a similar inauthenticity signal.

What reassures me slightly is what Grok and some analysts on X have said about the high-recall filter: it is designed to cast a wide net. It deliberately errs on the side of catching more. False positives involving real people are part of the trade-off. Judging by the comments, though, when the system gets it wrong, the review process usually moves very quickly and the label is removed. That is clearly not what happened in my case.

Another conclusion: even though the ideas, facts and structure of my posts were mine, and AI was only used for translation and light stylistic editing, the detector does not care about authorship in that sense. What triggers it are the statistical traces in the text: rhythm, connective tissue, smoothness, predictability. And if enough of those signals build up at the account level, it puts the label on the whole profile.

There is a curious asymmetry here. X actively promotes its own generative tools and is flooded with AI images and videos. At the same time, text that has passed through an LLM, even just as a translator or editor, can potentially become a low-quality signal.

There is at least one good thing about all this: now you don’t have to guess as much about why the system has restricted your account. You can now see the name of the label, its effect and how long it lasts. Before that, you were left wondering whether you had been shadowbanned and relying on dubious third-party services to find out.

Transparency, With Limits

By dropping part of its code, X partially opened up how the recommendation algorithm works, while at the same time tightening the filters.

In August, alongside the Under the Hood tool, X published code on GitHub in the xai-org/x-algorithm repository covering parts of the system that determine which posts end up in the For You feed. It is not the entire algorithm, nor does it expose every internal model, but it does reveal some key components.

They let you see how the system decides whether to show a post or hide it, which labels affect visibility, and how the filtering works (including SpamHighRecall and other restrictions).

X has said the code will be updated regularly, with explanations of what has changed. The original idea was a cycle of updates every few weeks. How consistently that will actually happen remains to be seen. But the publication of the code, together with personalized reports on account labels, already changes the rules of the game.

The platform now has a stronger argument against accusations of opacity. This matters especially in Europe, where under the Digital Services Act big platforms are required to explain how their recommendation and moderation systems work. When parts of the code and the label mechanism are publicly accessible, accusing X of complete opacity becomes noticeably harder.

At the same time, the transparency is only partial. Under the Hood shows the result, which label is attached to the account, but not the exact internal rule that triggered it. In the code, you can see that llm_slop_user automatically leads to a 30-day SpamHighRecall label, but you cannot see why the system decided that a particular account should be classified as llm_slop_user in the first place. The curtain has been lifted, but not ripped away completely.

For an ordinary author, that is still a meaningful shift. Before, a restriction felt like an invisible wall. Now the wall at least has a name, a duration and some kind of explanation.

Why Human Voice Still Matters

Overall, I see these changes as positive. The platform is trying to take care of the quality and originality of content for users. At the same time, authors who work with AI will have to keep working on texts the way they used to: building sentences themselves, finding a sharper angle, choosing the right shade of meaning, and doubting their own wording.

I know that many writers relaxed quite a bit once AI started giving their own thoughts back to them in clean, rhythmic and convincing prose. It can look so good that some people, for example when working with translations, stop correcting nuances or even small differences in the meaning of verbs.

Neural networks write so smoothly and so well that it becomes tempting to stop thinking. It turns into a kind of cognitive outsourcing: first we hand grammar over to the model, then phrasing, then structure, then the argument itself. After a while, it becomes hard to tell where exactly the author remains in the text.

That is why I would support X in this direction. If the platform wants to preserve a space where human authorship still has value, it has reasons to fight not only bots pumping out 500 posts a day, but also the mass production of polished AI text. Otherwise, displacement is almost inevitable: human writing is slower, rougher and more expensive to produce, while model-generated text can be produced almost endlessly and at virtually no cost.