XThe X Algorithm
X change: November 6, 2026

Does X penalize pasted replies? What the code says

X's code does tell its reply-scoring AI whether a reply was pasted, on every reply, not only under big accounts. A score of 0 gets a label. The code doesn't say pasting causes a 0.

← All updatesWritten By Bill G. Wolcott

A claim is going around X: stop pasting your replies, because X's open-source code has an "is_pasted" flag, Grok reads it, a pasted reply gets a 0, and a 0 buries the reply for good. I read every file in X's public code that touches it. Some of the claim is right, one part is too narrow, and the scariest part isn't in the code at all.

Each claim, checked against X's public code as of November 7, 2026
Each claim, checked against X's public code as of November 7, 2026

Is there really an is_pasted flag?

Yes. When X scores a reply, its code builds a short report about the person who wrote it and hands that report to an AI model. The report has a section called "Additional Signals" with eight lines. The last one reads "Reply Was Pasted: True" or "Reply Was Pasted: False." The line has been in the public code since X first published this part of it on September 18, 2026.

Which replies get scored?

Every reply to someone else's post. The only ones skipped are replies on your own post or your own thread. The claim says this happens to replies under big accounts. That's true, but it's only half of it. Which model does the scoring depends on the size of the thread: - Threads over 250,000 followers. If the account you replied to, or the account that started the thread, has more than 250,000 followers, Grok scores the reply. The code names the model as Grok 4 Mini, with Grok 4.1 Fast as a backup. - Every other thread. Gemma 4, an open AI model, scores the reply. In threads over 125,000 followers it's a version of Gemma tuned for reply spam. Quote posts always go to Grok.

Both models get the same report, including the pasted line. So the pasted flag is read on replies to small accounts too, not only big ones.

What happens to every reply you post, as X's code lays it out
What happens to every reply you post, as X's code lays it out

What else does the model see?

The pasted line is one of eight. The model is told: - whether X is missing the usual app activity for the account - the country the account is mostly used from - the account's language - how many replies the account posted in the last 24 hours - the account's follower count - whether the account already has a risky-user label - how many people blocked the account in the last 24 hours - whether the reply was pasted

It also gets the reply author's profile and up to 10 posts of the thread, with up to two pictures or videos from each. The code also records which app the reply came from and whether that app passed X's app check, but those go to X's logs. They aren't in the report the model reads.

The block X's code adds to every reply it sends for scoring. The values are examples.
The block X's code adds to every reply it sends for scoring. The values are examples.

How is a reply scored?

From 0 to 3. If a model returns anything outside 0 to 3, the code throws it out as an error. The score is stored with up to 500 characters of the model's reasoning.

Can a score go back up?

No. If a reply is scored again, the new score is saved only if it's lower than the one already stored. A higher score is skipped. So a reply's score can drop over time, but it can't recover. That part of the claim is correct.

What happens at a score of 0?

The reply gets a label called RiskyHighVizReply. Three kinds of accounts are exempt: grey-badge accounts, X's test accounts, and accounts with a credibility score of 66 or more. X has raised that cutoff three times: 60 on October 1, 62 on October 3 and 66 on October 8.

Does the label bury the reply?

The public code shows one effect. Since November 6, a reply with this label isn't sent as a reply or mention notification to people who don't follow you. If you reply to a big account that doesn't follow you, and your reply is labeled, that account won't be notified of it. People who already follow you aren't affected by that rule.

The code doesn't show whether the label or the score also moves the reply down in the reply list. The score is stored under the name "reply ranking," but the part of X that reads it to order replies isn't in the public code. "Buried" may be right. The code doesn't prove it.

Does pasting get you a 0?

The code doesn't say. It shows that the model is told whether a reply was pasted. It doesn't show what the model is told to do with that. The instructions that tell Grok and Gemma how to judge a reply are kept out of the public code. X's own note in the code says prompts are excluded "to reduce gameability."

So nobody outside X can say how much pasting counts, whether it counts against you at all, or whether it only matters together with other signals, such as dozens of replies in a day or a pile of blocks. Anyone who says "pasting gets you a 0" is guessing at the part X keeps private.

Did X change any of this recently?

Yes, several times since September. The follower line between Grok and Gemma went from 200,000 to 225,000 on September 23 and to 250,000 on September 24. The cutoff for the tuned Gemma dropped from 150,000 to 125,000 on September 26. The credibility exemption arrived on October 1 and was raised twice by October 8. The notification rule came on November 6. Separately, X's enforcement service has been able to add the same label since September 18, for 30 days, to posts it rates as low-quality AI-written text.

Should you stop pasting replies?

The code gives a plain reason to think pasting is noticed, and no proof that it's punished. Typing a reply in X costs you nothing and takes the question off the table. Whatever you do, the other signals in that list are ones the code shows just as clearly: a lot of replies in a day and a lot of blocks are both reported to the same model.

The code is in X's repository, and X announces changes at x.com/XOpenSource. Your edition of the book is updated in the November 25 drop.

Source: X's published code, commit 2e2295b.

What to do about it is in the book. Updates say what changed. The book explains how the feed works and what to do, and it's revised every month from X's code.

See what it costs