AI-isms

AI-isms are the words, phrases and sentence shapes that language models overuse, so often that readers learn to spot them: "delve", "it's worth noting", "not X but Y", a list of three in identical grammar. Each one appears in human writing too; what gives AI text away is the density, several in a short piece.

The numbers

What an AI-ism is

An AI-ism is a word, phrase or sentence shape that language models use far more often than people do. Think of the stock closers, a list of three items in identical grammar, or a paragraph where every sentence runs about thirty words. None of them is wrong. That's the whole problem.

Here's the thing: every AI-ism appears in human writing. People write "moreover" too. One hit proves nothing. A cluster in a short piece is the tell.

There's a number behind that. In the reference corpus of 299 human pieces, the median density is 0.4 AI-isms per 1,000 words, and the 90th percentile is 1.6. So a person who lands a stray "mechanism" or a "not X but Y" is well inside normal. The full list is on the AI-isms in marketing copy page.

The three kinds of AI-ism

Words come first. The best-known case is "delve". Juzek and Ward, in "Why Does ChatGPT Delve So Much?" (COLING 2025), analysed 26.7 million PubMed abstracts and found 21 focal words whose use rose sharply after ChatGPT's release. "Delves" rose 6,697% in those abstracts from 2020 to 2024. To be fair, that magnitude belongs to PubMed abstracts. It doesn't mean the word rose that much in blog posts. The overused words page has the sourced detail.

Phrases come next, in four groups. Generic transitions: "in conclusion", "in an age where". Nominalised abstract metaphors, the kind that stack nouns on nouns and call it insight. Hedged abstract closes, the ones that end on a vague soft note instead of a claim. And casual markers such as "you know" or "sort of", which count only when they recur, because a chatty tic repeated on schedule reads as a tic.

Then sentence shapes. These are the sneakiest, because no single word gives them away. The "not X but Y" balancing construction is the famous one. Others: "less X, more Y" contrasts, "perhaps X, perhaps Y" pairs, and triads in identical grammar. Three more work as rhythm checks: uniform sentence length, no short sentences, and no contractions across 300 words of contemporary prose.

Why models write them

Nobody has settled this. Juzek and Ward report failing to find evidence that the overuse comes from model architecture, or from training or fine-tuning data. The focal words are far rarer in corpora such as arXiv and Wikipedia than in ChatGPT output, so the data doesn't explain it either.

The hypothesis left standing is reinforcement learning from human feedback. The idea: preference tuning teaches models that this vocabulary is what good answers sound like. It's a hypothesis. It isn't proven.

The tells also date. "Delve" was heavily overused in 2023 and early 2024, then dropped off sharply in 2025. So a list ages as models change, and a list nobody maintains ends up catching last year's habits.

One measurement shows how far they move. In ScriptGrain's SGR-002 study (two briefs, one author), bare GPT-5 produced 0.2 listed tells per 1,000 words but averaged 17.5 em dashes per draft, and wrote 65% more words than asked. Different habits, same lesson: the tells move.

How to find and fix them

Count them. Don't hunt by eye. Your eye skips what it has seen a hundred times, and it flags the wrong things. The free AI cliché checker runs the list in the browser and reports density per 1,000 words. Under 1 is clean, 1 to 3 is some, over 3 is heavy. Density counts phrase hits plus the first six sentence shapes. Em dashes and the three rhythm meters are reported separately.

Then fix the sentence, not the word. A synonym for "delve" keeps the shape. You've swapped the costume and left the body.

And the honest limit: removing AI-isms makes text less generic. It doesn't make it yours. Your own voice is the real fix.

Worked example

Take this invented paragraph of about 200 words, offered as an illustration. It opens, "In today's world, remote work reshapes how teams operate. The change is not a threat but an opportunity. Managers gain flexibility, staff gain autonomy, and companies gain access to wider talent." The rest of the paragraph carries on in the same register and contains no other hits.

The checker finds three: the transition phrase "in today's world", one "not X but Y", and one triad in identical grammar. Three hits in 200 words is 15 per 1,000. Against the bands, that's heavy. The numbers are illustrative, not measured from a real piece.

Rewrite the sentences instead of swapping words. "Remote work changes how teams run. Some managers hate it. Others finally get to hire the best person, wherever they live." Nothing was replaced with a synonym; the shapes are gone. It's still generic, mind you. Add what you actually saw on your own team and it starts to sound like you.

Sources

More terms