How to make AI-written tweets not sound like AI
Give the model your own material first, then edit out its habits, then add one detail only you could know. Most AI tweets fail at the first step. The model got a topic and no facts, so it wrote the average of everything it has read.
Why swapping words won't fix it
Many guides give you a list of banned words. That helps a little. Wikipedia's editors keep a long guide to signs of AI writing, and it warns that the overused words change over time and that fixing surface signs alone won't solve the deeper problem.
These signs aren't proof, since people write this way too. The same guide says detection tools have real error rates, and it cites mixed research on people, including a 2025 study that found them no better than chance. So don't chase a detector score. Aim for a post your friends would believe you wrote.
The patterns that give it away
These come from Wikipedia's guide and other public lists of AI tells. Every example below is invented.
The fake correction. The model knocks down a claim nobody made.
- Before: "Consistency isn't about posting more. It's about showing up with intention."
- After: "I posted every weekday in March. Only the posts with a screenshot got replies."
The neat three. Three parallel items, with a rhythm like a slogan.
- Before: "Great founders listen, learn and adapt."
- After: "Our best change came from one annoyed customer email."
The closing moral. A final line that explains what you just read.
- Before: "We dropped our free plan and revenue rose. Sometimes less really is more."
- After: "We dropped our free plan and revenue rose. Support tickets fell too, which I didn't expect."
The trailing -ing phrase. A vague add-on at the end of a sentence.
- Before: "We redesigned the signup, highlighting our commitment to simplicity."
- After: "We redesigned the signup. It has two fields now."
Stock openers and closers. "In today's fast-paced world" at the top. "What do you think?" at the bottom. Phrases such as "game-changer" and "delve" belong on this list too.
Tidy punctuation. Wikipedia's guide flags heavy use of the long dash as a sign from recent years. Hashtags and bold text look odd in a casual post too. Write the way you'd type on your phone.
Feed it before you ask it
A model can't invent your experience, and you don't want it to. Give it three things:
- Five to ten of your own posts, as samples of how you write.
- The raw facts, in any order. What happened, when, and what surprised you.
- A short list of what to avoid.
Here's a prompt you can copy:
Here are posts I wrote: [paste 5 to 10]. Here are my notes for a new post: [paste notes]. Write three versions under 200 characters, using my usual sentence length and word choice. Use only facts from my notes. If a detail is missing, leave it out and tell me. Don't use hashtags, emoji, long dashes, lists of three, or a closing line that sums up the post.
The line about facts matters most. Models will produce a tidy statistic if you let them, and you're the one who has to stand behind it.
A 60-second edit
One guide suggests spending 30 to 60 seconds on each AI tweet, and that seems about right for a short post.
- Read it out loud. Rewrite anything you wouldn't say to a friend.
- Cut the first line if the post still makes sense without it.
- Cut the last line if it only repeats the point.
- Swap one general word for a real name, number or place. Keep it true.
Here's a full invented pass from a made-up coffee shop owner.
Draft: "Running a small business teaches you resilience, patience and the power of community. Every customer is a story waiting to be told. What has your business taught you?"
Edit: "A regular brought his own mug for six years. When we closed for repairs he left a note on the door asking if we were okay. I still have it."
The edit has no lesson attached. That's why it reads like a person. It also needed facts the model never had, which is the whole argument for step one.
Does X penalize AI-written posts?
Nobody outside X can say for sure. The README for X's published feed code describes ranking by predicted reactions and doesn't list an AI-text penalty, but some rules and the language-model prompts stay private, so that absence proves little.
Ranking does use negative signals. The published weights include negative values for "not interested", mute and block, so bland posts that annoy readers carry their own cost. X also has a rule for duplicate text in replies, which matters if a tool gives you and everyone else the same phrasing.
Automation is a separate question from authorship. X's developer guidelines say AI-generated replies from automated accounts need prior approval, and that even approved bots can't pretend to be human. Writing with help and posting by hand is a different activity. Keep a person in the loop, above all on replies.
A tool that checks for this
Repli learns from your past posts and checks drafts for these patterns before you see them. It can still miss, so read each draft like the steps above suggest.
Paste a post into the AI tweet checker on our free tools page for a straight answer and a more natural version.