is it slop? Check a passage

AI words and phrases

AI words and phrases are the vocabulary large language models reach for far more often than people do. Much of it has been counted: delve, underscore, showcase, meticulous, intricate, pivotal, landscape, realm, testament, tapestry, joins like additionally, and frames like it is not just X, it is Y. The counting does not give you a test. These are ordinary English words, most of them common in academic and professional writing long before 2022, and the set turns over with every model generation. Density is the signal, not the word. One delve means nothing. Fifteen in four paragraphs means something.

An illustrated robot at a laptop, pencil in hand, with notebooks and loose paper spread across the desk.
A word is a hint, never a verdict. People have been accused, marked down and disciplined over a single "delve". Read the list as a habit, not a rule of evidence.

Which words does ChatGPT overuse most?

Delve is the famous one and the best measured. A study of more than 15 million PubMed abstracts found the words that jumped after ChatGPT were style, not subject: delves ran at about 28 times its expected 2024 frequency, underscores about 14 times, showcasing about 11 times. A count of AI conference peer reviews found commendable, meticulous and intricate rising 9.8, 34.7 and 11.2 fold in their chance of appearing in a sentence. Most lists online are memory, not measurement, so the groups below stick to words somebody counted or documented, and flag the two that nobody has.

Verbs that arrive too often

Each stands in for a plainer verb, the one a hurried writer uses.

Where this goes wrong. These are not unusual words. Underscore and leverage have been business and policy standards for decades; delve, harness and foster are ordinary academic English. Wikipedia's field guide to AI writing adds that overuse says nothing about synonyms: a suspect delve does not make explore suspect.

The connective tissue

Models glue paragraphs with visible hardware. This group beats the verbs: humans vary the joins or drop them.

Where this goes wrong. This group gets innocent people accused, because formal English instruction teaches it: signpost with moreover, close with in conclusion. Detectors punish that. Stanford researchers ran seven detectors over 91 TOEFL essays by non-native English speakers: more than half came back labelled AI, one flagged nearly 98%, and the same tools correctly called over 90% of essays by US eighth-graders human (GPT detectors are biased against non-native English writers). It is the most important thing to know before you use a word list on somebody.

Additionally has counts behind it; moreover and furthermore are folk entries nobody measured, and the disclaimer habit belongs mostly to 2023 and 2024 models. Where a word list is used on a student, the tool behind the accusation matters as much as the words do: see the AI detectors schools and colleges use and what their scores are worth.

Nouns and abstractions

The vocabulary of writing about a subject without knowing much about it: elevated, abstract, attached to nothing checkable. That gap between sounding fluent and knowing anything is what the word AI slop points at.

Where this goes wrong. Landscape and realm are normal in strategy, policy, ecology and games writing, and the PubMed researchers annotated landscape as both content and style, since it carries a technical meaning as well as a decorative one. Their full list of 900 excess words is public, 407 marked style rather than content, far larger and duller than the twenty everyone quotes.

Sentence shapes

These matter more than the vocabulary: they survive the find-and-replace anyone does to make output pass.

Where this goes wrong. Humans write like this constantly. Negative parallelism is listicle and LinkedIn house style; serves as for is is a tic of committee prose and press releases, both older than the models. Editors strip these frames out too, so their absence proves nothing.

The list has a shelf life

The vocabulary turns over with the models. Wikipedia's guide, kept by editors cleaning up AI text since 2023, splits it into eras: delve, tapestry, testament, meticulous and intricate for 2023 and the first half of 2024, align with, fostering and showcasing for the middle, and by mid-2025 only emphasizing, enhance, highlighting and showcasing survive. Delve dropped off sharply during 2025. Grok has its own taste, for causal, empirical and correlate.

So a list built in 2024 misses recent output and flags old human writing that merely sounds formal. And the words are spreading: a 2026 preprint analysing open-access biomedical papers reports that by the end of 2025 89% of them carried an excess of LLM-associated vocabulary. Whatever the true figure, a marker that common has stopped telling one paper apart from another.

The list is contaminating its own evidence. The Max Planck Institute for Human Development tracked these words through hundreds of thousands of podcast episodes and found them rising in unscripted human speech after ChatGPT. We are learning the machine's vocabulary, so the list works worse every year.

Where the words came from

Nobody trained a model to love delve. The leading theory is reinforcement learning from human feedback, where paid annotators rank responses. Alex Hern noticed delve is more common in Nigerian business English than in British or American English, and annotation work is heavily outsourced to Nigeria, a theory summarised by Simon Willison. Researchers later tested seven explanations, found training data and architecture unconvincing, and came back to the feedback stage on mixed evidence. The words are an artefact of tuning, not a property of machine writing. Change the tuning and the list changes.

Does using these words mean something is AI?

No. Every word on this page is ordinary English, most of it common in academic and professional writing well before ChatGPT. Models did not invent any of it; they overuse it. A single word settles nothing about who wrote a sentence, and the people most likely to be wrongly caught by a word list are those who learned formal English as a second language, whose TOEFL essays a 2023 Stanford study found detectors labelling AI more than half the time. Density and combination carry the signal: many of these words, many times, in a short passage, alongside the sentence shapes described above. Wikipedia's guide calls its own list "descriptive, not prescriptive", and that is the right way to hold this one.

Checking a passage instead of a word

A word list turns you into a word hunter: you scan for delve, find it, stop. That is how most false accusations happen, and it fails both ways, catching careful non-native writers who signpost properly and missing machine text with its vocabulary swapped out. Vocabulary is the shallowest tell, and the structural ones hold up better, which is why the signs of AI writing works through each of them and where each goes wrong. One study found people who often use LLMs for their own writing are strong detectors of machine text, weighing vocabulary alongside formality, originality and clarity rather than hunting single words.

Paste a passage into the detector on the home page and you get a likelihood plus the parts carrying the signals, so you can argue with the reasons. It judges where the words came from, not whether they are good. It is not proof, it will not tell you who typed it, and it should never be the only thing you act on. Use it to decide whether to look harder.

Check a passage →

Last updated 2026-08-17.