Editing AI Text: The Telltale Signs

What the measurements show

Across fifteen million scientific abstracts, 379 style words became abruptly more frequent in 2024 than the previous years' trend could explain. The largest jumps: delves rose twenty-eight-fold, underscores almost fourteen-fold, showcasing more than tenfold. For scale: this linguistic shock was roughly twice the size of the one the pandemic vocabulary caused in 2021.

The second finding matters more in practice. According to another study, 59.3 percent of the 2024 papers that used delve also used underscore — in 2022 that figure was 1.3 percent. It is not the individual word that gives text away but the cluster: three or four of them close together means something; one of them means nothing at all.

The third measurement goes deeper than any word list. Comparing human and machine text on the same set of linguistic features showed that models do not miss the average of style but its spread: the variation in sentence length, register and structure is narrower than a human's. And instruction-tuned models are narrower than their base versions — so the flattening does not come from the training data but from the process that made the model polite afterwards.

The twelve phrases

These give text away in English. Not because they are bad words, but because the model reaches for them out of all proportion — and because human writing rarely produces four of them in a row.

• The empty opener. "In today's fast-paced world", "In recent years, X has gained increasing attention." In most machine text the first sentence of the first paragraph can be deleted without loss.
• The importance markers. Crucial, vital, essential, pivotal. If something really is, say what it is crucial to and why. If it is not, cut it.
• The digging metaphor. "Let's delve into", "let's dive deeper", "let's take a closer look." This is the strongest single lexical marker measured anywhere, and it now reads as a signature.
• The highlighting family. Underscores, highlights, showcases, emphasises — four words for one act, and machine text likes all four inside a single paragraph.
• The not-just-but construction. "This is not just a technological question, it is a human one." The antithesis itself is fine; the problem is that the model builds every second paragraph this way.
• The rule of three everywhere. Three adjectives, three examples, three considerations — even where two would do and where five are needed. Human writing is uneven.
• The announced importance. "It is important to note that", "it is worth emphasising that." If it is important, write it; announcing that it is adds nothing.
• The summary close. "In conclusion", "all in all", "taken together, we can say." Machine text nearly always repeats itself at the end, because that is what it was trained to do.
• The landscape metaphor. Landscape, realm, world, ecosystem, maze, and the navigating that comes with them. "Navigating the ever-evolving landscape of AI" appears in a thousand places and says nothing in any of them.
• The even sentence length. If every sentence in a paragraph runs 15 to 20 words, that alone is suspicious. A human paragraph contains a four-word sentence somewhere.
• The adjective pile-up. Comprehensive, robust, seamless, innovative, holistic. Every one of them is usable on its own; finding three in one paragraph is a reliable sign.
• The markdown residue. Every other word in bold, every thought under its own subheading. That is not style, it is the model's default formatting, and in an email or an article it reads as foreign.

What a word list will not catch

Once the phrases are swapped out the text can still read as machine-written, because most of what makes it recognisable lives in the structure rather than the vocabulary.

• There is no uncertainty in it. Measurements show models use fewer hedges than humans — "I think", "I'm not sure", "I may be wrong about this". Machine text is more confident than the knowledge behind it warrants.
• There is no personal aside. A human mentions in half a sentence where they ran into the problem. The model has nowhere to have run into it, so that half-sentence is always missing.
• There is no question to the reader. Direct address and a real question are human markers; the model thinks in statements.
• Every paragraph is the same size. Four sentences, four sentences, four sentences. Human writing contains a one-sentence paragraph, because somewhere the point needed weight.

The em-dash myth

The most widespread belief is that a lot of em-dashes gives machine text away. This has been measured: the human essays in the study averaged 3.23 em-dashes per thousand words, but the range ran from 0.33 to 17.12. The models sit inside that range — one popular model at 10.62, another at 9.09, while some open-weight families used none at all.

Two things follow. First, the count proves nothing on its own, because the human range overlaps the machine range completely. Second, it is model-specific and instructable: in the same study one model dropped from 9.09 to 0.19 when asked for plain prose, while another did not move at all. If the dashes bother you, ask for fewer; but do not imagine their absence makes the text human.

About detectors, before you trust one

This section has practical consequences, so it is worth slowing down. The tools built to detect AI text do not work reliably, and that is not an opinion: one major provider released its own classifier at a 26 percent detection rate with a 9 percent false alarm rate, then withdrew it six months later for poor accuracy.

The number that should worry you is worse. In a study testing fourteen tools, none reached 80 percent accuracy, and for manually edited machine text the rate fell to 42 percent, for paraphrased text to 26 percent. Another study ran seven detectors over essays by non-native English writers: the average false-positive rate was 61.22 percent, against 5.19 percent for native-speaker student writing. The explanation is linguistic — simpler, more predictable sentence construction looks machine-like to a detector.

If you write in English as a second language, this is aimed at you: the system punishes clean, plain English and rewards the ornate kind. In the same study, detection of AI-written text fell from 100 percent to 13 percent once the model was asked to write in more literary language. So if someone "proves" with a detector that you used a machine, that number carries no evidential weight — worth knowing before you start defending yourself.

The twenty-minute editing pass

The order is what it is because the cheapest fixes come first.

1. Delete the first and last paragraph. Seven times out of eight the piece improves, and the article starts where its first real claim is.
2. Search for the twelve phrases. Literal search, one at a time. This takes five minutes and removes half of what is recognisable.
3. Break three paragraphs. Cut in a short sentence where emphasis is needed, and merge two where it is not.
4. Put in one thing only you know. A number from your own work, a case, a date. This is the single step the model cannot structurally do for you.
5. Read the first three paragraphs aloud. Wherever you stumble or get bored, the text is still machine-written. Fixing your own voice first shortens this step, though it does not replace it.

What to take away

Machine writing shows up in clusters, not in individual words, and in structure more than in vocabulary. Delete the first and last paragraph, comb out the twelve phrases, break the rhythm, and add one fact of your own. Do not trust detectors — not against yourself and not against anyone else. Whether any of this has to be disclosed at all is a separate post: when you have to disclose AI, and the answer is surprisingly narrow.