A Quarter of Long-Form Social Posts Are Fully Machine-Written

If you have been reading long-form posts on LinkedIn lately, the odds of the one you are reading being written entirely by a machine are not small. Pangram, an AI content detection company, put that number on its Chrome extension. It scanned more than one million posts between April and June 2026 and found that one in four long-form articles across every platform it looked at was fully machine-generated.

Pangram defines long-form as anything over 250 words. The average is not remotely even. On LinkedIn, 41 percent of long-form posts were flagged as entirely AI-written. Only 55.2 percent were actually human-written, and just 4.3 percent were AI-assisted.

The number worth remembering is not the figure itself. It is that this number cannot be used to accuse anyone.

Where the slop is piling up

Here is Pangram's per-platform breakdown, for articles over 250 words:

• LinkedIn — 40.5 percent AI, 4.3 percent mixed, 55.2 percent human
• Medium — 30.6 percent AI, 9.7 percent mixed, 59.7 percent human
• X — 24.1 percent AI, 23.2 percent mixed, 52.7 percent human
• Substack — 10.0 percent AI, 11.9 percent mixed, 78.1 percent human

Two details are worth holding onto. First, on four of the five platforms, longer content was actually more likely to be machine-written. Second, Substack is the exception: there, longer articles were slightly less likely to be AI.

On X, counting mixed content, nearly half of all articles touch AI in some way. That is why X feels the worst to read. Not because it has the most AI, but because it has the most half-finished AI. And two thirds of every item flagged as AI came from LinkedIn.

Why these numbers cannot be used to judge anyone

This is the part that matters most, and it almost never gets reported alongside the numbers above.

Detection has improved fast. Pangram now claims a false positive rate around 0.0041 percent, roughly one false positive per 24,000 documents. Humanizer tools have been beaten too: in published testing, humanizer output was almost always reported as human-written. A sentence like "The numbers are no longer small enough to ignore" became "The sheer size of these usage figures can no longer be ignored", and the detector called that human.

But getting faster has not removed the deeper problem. Several cases have become public:

• A self-published novel had its publishing deal cancelled after Pangram's chief executive said it was 78 percent AI, while the author denied writing it with AI
• An instalment of The New York Times' Modern Love column was scored at 100 percent AI
• Portions of Pope Leo XIV's encyclical on the dangers of AI were accused of being AI-written

The most important part of that list: most of these accusations were never substantiated, and some of them are demonstrably wrong. A University of Maryland study found that people who frequently use LLMs for writing are actually more accurate at telling AI text from human text, without any special training. So this is not merely a vendor problem.

The unfairest part of the whole record: detectors have been reported to disproportionately flag non-native English writers and neurodivergent writers as machine-generated. The reason is understandable, and that is exactly the problem. Writing that is "too tidy" is more likely to look like, and that structural bias is baked into the tooling.

There is also a working paper from Notre Dame finding that light AI editing on academic abstracts was flagged as AI writing 64 to 80 percent of the time. And a teacher in a New York City high school put it more bluntly: he has run some of his students' papers through it and it shows up as 100 percent human, which he does not believe.

What humanizers actually do

There is a piece of context a regular reader needs here. A humanizer is a tool that makes AI output look more human. A journalist who tested one found it swapped one clunky transition clause for another clunky transition clause, while introducing deliberate grammatical oddities.

So running a humanizer on your own writing carries two risks at once. The text gets worse, and you train your readers to assume slightly sloppy writing is human.

What you can actually check

Far more of this is mechanically checkable than the marketing around it suggests. And none of it requires guessing:

• Opening cliches. "In today's fast-paced world", "when it comes to", "at the end of the day"
• The "not just X, but Y" construction. Almost always a sign of a template
• Stacked connectors. "Furthermore", "Moreover", "Additionally" inside one paragraph
• Stacked hedges. "may potentially", "can often be", "somewhat" inside one sentence
• Em-dash overuse. Roughly more than one and a half em dashes per thousand characters
• Flat sentence rhythm. If every sentence is nearly the same length, the rhythm reads dead
• Repeated sentence openers. The same word opening five sentences in a row

All of it can be checked without guessing, and all of it is fixed by editing. That is the genuinely useful step.

What is worth being blunt about is this: scanning someone else's writing for "proof" they cheated is not detection, it is guesswork. And you already have better evidence than that, namely how badly the detectors themselves fail.

Check your own writing. Open AI Writing Tells on Loonix, paste a paragraph, and see which patterns are worth fixing. It deliberately gives you no "percent AI" score, because that number is not trustworthy and has already been used to punish the wrong people.

And if you genuinely are writing with AI assistance, the safest rule is simple: does the piece contain a name, a number, or a lived experience that you could not have invented yourself?