AI watermarks: how to take the ChatGPT feel out of your text

You ask ChatGPT for a post, read it, find it reasonable and publish. Three people comment "AI text" and everyone else scrolls past. The problem is not using AI. The problem is publishing the text the way it came out.

Language models write with tics. They are statistical patterns that show up in almost every generated text, whatever the topic, and LinkedIn readers have learned to spot them. This guide lists those patterns, shows how to remove them and explains why it matters for reach.

What an AI "watermark" is

There is research on technical watermarking, where the model picks words following a hidden pattern a detector can read later. In practice, no model you use day to day ships this in a way a reader would notice.

The watermark that matters is a different one. It is the set of style habits the model repeats because they were rewarded in training: symmetrical structures, easy contrasts, generic adjectives, openers that say nothing. An attentive reader spots it in two lines. A recruiter or client who reads 50 posts a day spots it in one.

And it has a cost. In March 2026 LinkedIn announced feed changes to reduce the reach of generic content and artificial engagement tactics. In Postbridge's data, what separates the top 5% of posts from the bottom half is format and specificity, and unedited AI text tends to fail at both.

The 14 most common tells

This list comes from the posts that go through Postbridge and from the drafts I generate myself before editing. If your text has three or more, it reads as AI.

1. An em dash in every sentence

The model loves the long dash to splice ideas together — like this — several times per paragraph. Almost nobody writes LinkedIn posts that way. Replace it with a period or a comma.

2. "It's not about X. It's about Y."

The ready-made contrast: "It's not about selling. It's about connecting." It works once every twenty posts. The model uses it in all of them.

3. Everything in threes

"Clarity, consistency and courage." "Faster, cheaper, more human." Three parallel items, ideally alliterated. One list of three is natural; five lists of three in the same text is not.

4. An opener that says nothing

"In today's fast-paced world", "In an increasingly connected landscape", "In this article we will explore", "Let's dive in". The first line is the most expensive line in the post. In our data, hooks over 140 characters get half the median interactions. Start from the second sentence of the draft; it is almost always the real opener.

5. Closing with a rhetorical question

"What do you think?" "Agree?" "What's your experience?" The model ends this way because it learned that posts end with a question. The reader hears a request for comments, not a conversation.

6. School-essay connectors

"It's worth noting", "it's important to highlight", "in summary", "that said", "in this sense". Each one can be deleted without losing anything.

7. Empty adjectives

"Robust", "powerful", "transformative", "essential", "incredible", "game-changing". None of them describes. If the tool is powerful, say what it does in 20 seconds that used to take an hour.

8. Marketing verbs

"Leverage", "unlock", "elevate", "empower", "supercharge", "optimize". Verbs that fit anything fit nothing.

9. Formula titles

"The 7 pillars that matter", "The ultimate guide", "Everything you need to know", "X things nobody tells you". Replace with a concrete claim or the result.

10. An emoji and bold text in every item

A list with a different emoji in front of each line and one bold word in every sentence. In Postbridge's data, posts with emoji have a higher median, but those are posts with one or two emoji doing a job, not one per line.

11. Sentences of the same length

Paragraphs of three sentences, each 15 words long. Human text has short sentences. And it has long ones, which take a little longer to close the idea and change the rhythm precisely because of that.

12. No detail only you would know

No name, no date, no specific number, no mistake of your own. AI text talks about "a client" and "one time". You talk about Tuesday's meeting with Renata, where the deck had 40 slides and nobody got past the eighth.

13. Double hedging

"Could potentially help", "tends to generally improve". One uncertainty per sentence is already a lot.

14. Keynote vocabulary

"Journey", "delve", "tapestry", "ecosystem", "landscape", "synergy", "mindset", "resilience", "navigate". If you would not say the word in a conversation, do not use it in the post.

A 7-step editing method

The goal is not to hide that you used AI. It is to publish something you would sign. It takes 5 to 10 minutes per post.

Step 1: delete the first paragraph

Read the draft from the second sentence. If it still makes sense, the first sentence was decoration. In 8 out of 10 drafts, that is what happens.

Step 2: replace every generality with a fact of yours

For every sentence that could sit in anyone's post, ask: which example of mine proves this? A number, a name, a date, what went wrong. No example, no sentence.

Step 3: break the symmetry

Look for triplets, "it's not X, it's Y" contrasts and twin paragraphs. Keep at most one of each per post. Vary sentence length on purpose.

Step 4: hunt dashes and connectors

Use the editor's search: em dash, "it's worth noting", "in summary", "it's important". Delete or replace with a period.

Step 5: swap adjectives for description

"Incredible result" becomes "from 200 to 1,400 views in a week". "Powerful tool" becomes "drafts in 20 seconds". If it cannot be described, cut the adjective.

Step 6: read it out loud

Anything you would not say to a colleague over coffee, rewrite the way you would say it. It is the fastest test there is and it catches what the other steps missed.

Step 7: one ask at the end

Instead of a rhetorical question, close with the conclusion or a specific request: "if you've tried this, tell me what changed". In our data, posts that open with a question get more comments; posts that close with a generic question do not.

Before and after

AI-generated draft, unedited:

In today's fast-paced world, productivity isn't about doing more. It's about doing what matters. After years leading teams, I learned three fundamental lessons: clarity, focus and consistency. It's not easy, but it's transformative. What about you, how do you organize your day?

After editing with the seven steps:

In 2024 I opened the day with 40 tabs and closed it with 38. I changed one thing: the first hour of the morning has no email and no Slack, only the task I would otherwise postpone. In three months, the "I'll look later" backlog went from 60 items to 9. If you've tried blocking the first hour, tell me what changed.

Same topic, same person, half the adjectives and three details only the author would know.

About AI detectors

Tools like GPTZero and similar promise to tell whether a text was generated by a model. Do not optimize for them. OpenAI itself shut down its detector in July 2023 for low accuracy, and a Stanford study published in Patterns showed that detectors flag text written by non-native English speakers as AI at high rates. The detector that matters is the reader, and the reader does not use statistics: they notice the text could be anyone's.

How to use AI without leaving a mark

AI is good at three stages: getting the idea out of nothing, organizing a mess of notes and suggesting hook variations. It is bad at writing like you. So flip the order: give it your facts (the case, the number, what went wrong), ask for the draft and run the seven steps above. The result carries your voice because the raw material is yours.

If you want to test hooks before writing the body, Postbridge's hook generator builds variations from your topic, and the hooks guide explains what we measured about first lines. For the structure of the text, see how to write LinkedIn posts. Postbridge learns from your past posts before generating new ones, which handles most of the list above before you start editing.

Frequently asked questions

Does LinkedIn penalize AI-generated text?

LinkedIn has not announced a penalty for using AI. What it announced, in March 2026, was reduced reach for generic content and artificial engagement tactics. Unedited AI text usually lands in the first category because of its content, not its origin.

Do AI detectors work?

With many false positives. OpenAI shut down its own detector in 2023 for low accuracy, and research shows bias against text by non-native writers. Use the reader as the reference, not the detector.

Can I use em dashes?

Yes. The problem is frequency. One dash every three or four paragraphs is fine; three per paragraph is the best-known signature of generated text.

How long does it take to edit an AI post?

Between 5 and 10 minutes following the seven steps. Less than that usually leaves marks; more than that means it would have been faster to write from scratch.

Is it better to write from scratch?

It depends on the bottleneck. If your problem is starting, AI solves the blank page and editing solves the voice. If your problem is editing time, write from scratch starting from one fact and ask AI only for hook variations.

How do I know if my text still reads as AI?

Count the tells on the list. Three or more, edit again. Then read it out loud: whatever you would not say to a colleague, rewrite.