AI Writing Tells: What We Found in 218 Blog Posts

Key Takeaways

• We scanned our own blog archive for AI writing tells: 218 published posts, 311,845 words, going back to 2007.
• 67% of those posts tripped at least one check, and most of that writing predates generative AI entirely.
• The flagged language splits into two groups: filler that humans already overused, and vocabulary that humans almost never choose.
• 12 of the 25 phrases we checked appeared zero times in 19 years of human writing. Those are the words that actually signal a machine.
• Dash punctuation was the single loudest signal, with 209 em-dashes and 158 en-dashes across the archive.

AI writing tells are the words and punctuation patterns that make a piece of content read as machine-generated, whether or not a machine wrote it. Every marketer publishing with AI assistance wants to know if their content carries them.

So we ran the test on ourselves first. Instead of scanning a competitor or a sample of random blogs, we pointed a checker at our entire published archive and counted.

The result was uncomfortable enough to be worth writing up. Two thirds of our own content marketing archive failed, and almost none of it was written by a machine.

What We Scanned

The dataset was every published post on this blog: 218 posts, 311,845 words, spanning December 2007 through September 2026.

We checked each post against 25 phrases associated with machine-written copy, plus em-dash and en-dash punctuation used as a pause device. Hyphens inside compound words like “same-day” and “family-owned” were excluded, since those are ordinary English.

The scan was deterministic. It counted exact matches rather than scoring anything, which matters for a reason we will get to.

67% of Our Archive Failed

146 of 218 posts tripped at least one check. Only 72 came back completely clean.

Breaking that down, 99 posts contained at least one flagged phrase and 79 contained dash punctuation. The archive held 226 total phrase hits, 209 em-dashes, and 158 en-dashes.

Here is the part that reframes the whole exercise. Most of this content was published years before generative AI was available to anyone. Human writers produced it.

That single fact should make you skeptical of any tool claiming to tell you whether a machine wrote something.

The Two Kinds of AI Writing Tells

Sorting the results by frequency revealed something more useful than a pass or fail score. The flagged language falls into two distinct groups, and they mean completely different things.

Filler Humans Already Overused

The top offenders were phrases people have been writing badly for decades.

These are not machine vocabulary. They are weak writing, and they were weak writing long before anyone had access to a language model.

Large language models trained on that same corpus learned the habit and amplified it. The phrase became a signal only because the machines lean on it harder than we did.

Cut these because they are filler, not because a robot might be blamed. The sentence is shorter and clearer without them either way.

Words Humans Almost Never Write

The more interesting finding was the silence. Twelve of the 25 checked phrases appeared exactly zero times across 311,845 words of human writing.

Not once did our archive contain “bespoke,” “underpins,” “our suite,” “in the heart of,” “unlock the secrets,” “it is advisable,” “amongst,” or “designed to enhance.”

That is the real detector. When a draft contains vocabulary that nineteen years of human writing never reached for, something else composed it.

A single instance of “delve” or “bespoke” in a business blog post carries more signal than fifty instances of “not only.”

Dash Punctuation Was the Loudest Signal

The em-dash deserves its own section because it produced the most extreme individual result in the scan.

One post from our archive contained 42 en-dashes across 2,125 words and zero flagged phrases. Every word choice passed. The punctuation alone made it read as generated.

Language models use the em-dash as a default pause device at a rate that human business writing never matched. Readers have started to notice, even when they cannot articulate why a paragraph feels off.

Rewrite with a comma, a period, a colon, or parentheses. The meaning survives and the tell disappears. Curious what your own count is? We will run a free scan and send the number back to you.

Why AI Detectors Are the Wrong Tool

Our 67% figure is the argument against the entire AI detection product category.

If a checker flags two thirds of an archive written mostly by people, then any tool converting those same signals into a confidence score is going to accuse innocent writing constantly. That is exactly what has happened to students and freelancers over the past few years.

Google’s own position supports treating this as a quality question rather than an origin question. Its guidance on AI-generated content focuses on whether content is helpful and original, not on how it was produced.

A linter is honest in a way a detector cannot be. It reports that you used a specific word eleven times. It makes no claim about who or what wrote the sentence, and it cannot be wrong about the thing it measures.

That distinction matters if you publish at volume. One tool gives you an editable list. The other gives you an accusation you cannot verify.

How to Check Your Own Content

Start with your archive rather than your next draft. Historical content shows you your actual habits, and it is already indexed and earning traffic.

Count exact matches instead of asking a model to evaluate its own writing. A language model asked to avoid sounding like a language model will agree enthusiastically and then produce the same vocabulary.

Separate the two categories when you review. Treat the rare vocabulary as a hard stop and the common filler as an editing pass, because they carry very different weight.

Fix the punctuation first. It takes minutes, it requires no rewriting, and it removes the signal readers pick up on fastest.

Then leave the rest alone. Passing a phrase check does not make writing good, and rewriting every flagged sentence produces copy that is clean and lifeless. Content still has to answer a question better than the pages competing with it, which is the same standard that has always applied to SEO content.

Where This Fits in Search

Search engines are not running phrase checkers. Nothing in this scan maps to a ranking factor.

What it maps to is trust. A reader deciding whether to hire you reads three paragraphs, and vocabulary that pattern-matches to generated filler costs you that reader before any algorithm is involved.

That cost is rising as more content gets published this way. It also runs alongside the shift in how results are assembled, which we covered in our breakdown of Google AI Overviews and in our guide to generative engine optimization.

For the record, this post fails its own checker fifteen times. Every one of those hits is a phrase quoted as an example, which is the limitation of counting exact matches: a checker reads words, not intent. That is also why the output is a list to review rather than a verdict to obey.

The practical takeaway is small. Run a real count against your own published work, fix the punctuation, cut the twelve words humans never use, and then go back to worrying about whether the content is actually worth reading.