Getting caught is not the risk. Sounding the same is.

On 9 July, Pangram published a scan of just over a million social posts. The data came from people who opted in through their Chrome extension, and the interesting slice was longform — anything past 250 words.

LinkedIn: 41% of it fully AI-generated, the highest of the five platforms. X: 25% fully generated, plus roughly another 23% written with AI assistance. Reddit and Substack both sat around one in ten — about a quarter of LinkedIn's rate. One other pattern held across platforms: top-level posts are far more likely to be machine-written than the comments underneath them.

Three weeks later LinkedIn acted on it. On 30 July they added a "Seems like AI slop" option to the menu on every post, and anyone reported gets told privately in their own analytics. At the same time they retired the "rewrite with AI" button they had spent years promoting, replacing it with a tool that only proofreads and leaves your voice alone.

Most of the coverage read this as a flood. I read it as a price list.

Getting caught stopped being the risk

When two in every five long posts on a platform are machine-written, readers have stopped trying to tell. They are scrolling, and they decide inside a second whether to stop.

So being found out is not what worries me. This is: the thing you spent time producing reads like the other 40%, gets sorted into the same pile, and gets scrolled past with it.

That is a harder problem than being reported. Nobody steps forward to accuse you. You simply are not read.

And the people most exposed to it are exactly the small teams reaching for AI to lift their output — us, and most of our clients.

What that button is telling you

A platform willing to put a report option next to every post, while killing the AI rewriting feature it promoted for years, has done the arithmetic: reader patience is worth more than posting volume.

LinkedIn will not be the last. "How high can we push output" is being replaced by a different question — what is still worth finishing.

What machine writing actually sounds like

"AI-sounding" feels subjective. It is fairly mechanical. The reliable ones:

• The antithesis reflex. "It's not X — it's Y." The single loudest tell, in English and Chinese alike.
• Everything in threes. Three items whether or not the third one earned its place.
• Signposting. "First… Second… More importantly…" all the way down.
• The TED lift. Four paragraphs of practical detail, then a closing move up to what this says about being human.
• Rhetorical question, immediate answer. "So what should you do? The answer is—"
• Em-dashes everywhere. One is fine. Thirteen in a piece is a confession.
• Words that sound full and are empty. Leverage, robust, holistic, ecosystem. The ones you can delete without changing the sentence.

The list is worth more before you write than after. Knowing where the holes are costs far less than being told you fell in one.

One line that matters more than style

Opinions, yes. Invented experience, no.

The most dangerous habit an AI writer has is reaching for "we tested this" or "a client told us" to make a paragraph land harder. Those sentences read as the most human on the page, and they do the most damage — because they can be checked, and eventually they are.

Have as strong a view as you like. Do not manufacture something that did not happen.

The part no tool fixes

A draft can read as completely human and still tell you nothing about whether today is the day to post it.

Timing, what you already said last week, which client happens to be reading — none of that is in the draft, because the model was not in the room. Someone still presses the button, and here is why.

Tools keep the draft from being bad. You decide whether it should go out at all.

If you want to see what your own double would write, building and previewing are free — no card required.