Does the X algorithm penalize pasted replies?

I built a tool for this, and it’s free: TypeOut for iPhone and iPad, and TypeOut for Chrome.

The post that started it

On 7 November 2026, Konsti Wohlwend posted on X that the X algorithm ranks a pasted reply lower than a typed one. More than 600 people liked it within a day, and it was bad news for anyone who drafts their posts somewhere else first.

That’s me. I write my replies in a notes app, read them over, sometimes run them past a grammar checker, and then paste them in. So before changing anything, I went through X’s code to see whether the claim holds up.

What X’s code says

X, formerly Twitter, publishes its recommendation code on GitHub as xai-org/x-algorithm. Everything below is from the 7 November 2026 version (commit 3b6919f), and each step links to the lines it describes.

1. Replies go to an AI judge

X runs replies through an AI model that scores them. Which model depends on how big the conversation is:

• If the post you’re answering, or the first post in the thread, is from an account with more than 250,000 followers, the judge is Grok 4 mini, with Grok 4.1 Fast as a backup (task_filter.py, classifier_reply_ranking.py).
• Otherwise it’s Gemma 4, Google’s open model (classifier_simple_reply_scorer.py).

Replies to your own post, and replies anywhere in a thread you started, are skipped.

2. The judge is told whether you pasted

Before the judge reads your reply, X adds a short list of “Additional Signals” about you. One of them is whether the reply was pasted:

if post.is_pasted is not None:
    lines.append(f"  - Reply Was Pasted: {post.is_pasted}")

That’s thread.py, lines 62 and 63. Both judges get the list. The rest of it covers:

• how many replies you’ve posted in the last 24 hours
• how many blocks you’ve received in the last 24 hours
• your follower count
• your account’s top country and language
• whether the activity X’s apps normally log is missing (“Has Missing Client Events”)
• whether your account already has a risky safety label

The pasted flag arrives with the post’s metadata (data_types.py), and the reply judges are the only part of the published code that uses it. It isn’t new, either: the 13 August 2026 release has the same line.

3. It scores your reply from 0 to 3

The judge returns a score and a reason. Anything outside 0 to 3 is thrown out as “outside the 0-3 rubric” (classifier_reply_ranking.py, lines 164 to 170). The score is saved as the reply’s “reply ranking score”. The code that reads it back to order replies isn’t part of the release.

4. A 0 hides your reply from people who don’t follow you

A score of 0 puts a label called RiskyHighVizReply on your reply (task_write.py, lines 140 and 141). X’s visibility rules then drop labeled replies for everyone who doesn’t follow you (tweet_rules.rs, lines 1022 to 1027), and that rule is applied to mention and reply notifications (levels.rs, line 217 and line 261).

In practice, if the account you replied to doesn’t follow you, your reply never shows up in their notifications. When you reply to a big account, being seen by that account is usually the point, so this is the cost that matters.

Two kinds of accounts never get the label: accounts with an internal credibility score of 66 or more (X doesn’t publish the scale), and grey-checkmark accounts, which X gives to governments and multilateral organizations (task_write.py, lines 39 to 66).

5. How much pasting counts is private

The judge’s instructions aren’t in the release, and the code says why: the prompts are left out “to reduce gameability of the system” (prompts.py, line 11). So the code proves the judge knows you pasted. It can’t tell anyone how much that matters next to everything else the judge reads.

So is the post right?

Mostly. Here’s each claim next to what the code shows.

The screenshot in Konsti’s post is an AI-written summary of the code. Its numbers match what I found: the 250,000-follower cutoff, the credibility score of 66 and the grey-checkmark exemption. Its remark that pasting is “the strongest tell” of an AI-written reply is the summary’s own opinion, though. Nothing in the code ranks the signals.

Why I built TypeOut

I wasn’t going to stop drafting my replies, and retyping them by hand defeats the point of writing them carefully. So that weekend I built TypeOut: it types what you copied into the box you’re in, one key at a time, at a human pace. The words go in as typing, so there’s no paste for an app to record.

• On iPhone and iPad: get it free on the App Store, switch to the TypeOut keyboard and tap Type clipboard. The keyboard enters your text one character at a time, the way you would.
• In Chrome: add it from the Chrome Web Store, copy your reply, click into the reply box and press ⌥⇧V on a Mac, or AltShiftV on Windows. TypeOut sends real key presses through Chrome’s DevTools protocol. When I tested it in plain text boxes and rich-text editors, the page saw only key presses and no paste.

There are four speeds, from about 45 to 95 words a minute in Chrome and 30 to 75 on iPhone. It pauses after punctuation, and it can make the odd typo and fix it with Backspace. The finished text always matches what you copied, and you still read it over and press Reply yourself.

There’s no account and no analytics, and nothing you copy or type leaves your device.

What TypeOut won’t do

• It won’t make a weak reply strong. The judge reads the words and the thread. A typed reply that’s off-topic or spammy can still score 0.
• It doesn’t touch your other signals. Replies in the last 24 hours, blocks and your account’s history count the same however the text got there.
• It isn’t a bot. It types one reply into the box you chose, at human speed, when you ask it to. It doesn’t post, click or reply for you.
• The code can change. This page reflects X’s 7 November 2026 release. If a later release changes how pasted replies are handled, I’ll update it.

Sources

Checked on 8 November 2026 against commit 3b6919f of xai-org/x-algorithm, published on 7 November 2026. The links point to that commit, so the line numbers won’t drift.

• Konsti Wohlwend’s post on X, 7 November 2026
• thread.py, lines 36 to 66: the “Additional Signals” list, with “Reply Was Pasted” on lines 62 and 63
• data_types.py, lines 894 and 895: the flag comes from the post’s metadata
• task_filter.py, lines 185 to 249: which replies go to the Grok judge
• constants.py, lines 5 to 9: the 250,000-follower cutoff and the credibility score of 66
• classifier_reply_ranking.py, lines 55 to 69: the Grok judge gets the signals
• classifier_simple_reply_scorer.py, lines 51 to 67: the Gemma judge gets them too
• task_write.py, lines 39 to 66 and lines 140 and 141: what a 0 does, and who is exempt
• tweet_rules.rs, lines 1022 to 1027 and levels.rs: the label hides the reply from non-followers’ notifications
• prompts.py, line 11: why the judge’s instructions are private

Questions