AI phrases finder
The words and phrases AI overuses, highlighted in your text
Paste a passage and see every one of the 300+ phrases that show up far more often in AI writing than in human prose, with how much more often.
Find AI phrasesHow it works
Paste, read the highlights, fix.
Paste the text
Any English text: an essay, a blog draft, an email, a LinkedIn post. Matching is instant and runs in your browser.
Read the highlights
Every match is highlighted in place. Hover to see its multiplier and category, then decide whether it is your voice or the model's.
Fix, then re-check
Rewrite the worst offenders first (highest multiplier), then run the passage through the AI detector to see whether the score moved.
What the AI phrases list is
Language models have favourite words. Ask one to write about almost anything and you will meet delve, tapestry, vibrant, pivotal and underscore; a paragraph that stands as a testament to something; a closing line that begins in conclusion. None of these words is wrong. They are simply far more frequent in model output than in the writing of people who have not been trained on the same data, and that gap is what makes them useful as a signal.
This page holds a list of 300+ such words and phrases. Each entry carries a category, so you can see what kind of tell it is, and a multiplier: our estimate of how many times more often it appears in AI-generated text than in human writing. The tool above matches the list against your text, whole words only and longest phrase first, and highlights every hit in place.
How the list is built
Three sources feed it. The first is corpus research: studies that compare the frequency of words in academic abstracts, reviews and web text from before 2022 with the same genres after models became common, which is where the famous delve spike came from. The second is the working checklist editors use to spot model-written text, which contributes the structural categories: symbolism and importance, superficial analysis, promotional language, attribution to unnamed experts, chatbot mannerisms and knowledge-cutoff disclaimers. The third is our own detector benchmark, which keeps turning up new tells as models and the people prompting them change.
That last source is why the list keeps growing. The 2026 additions include trendy LinkedIn phrasing, fake engagement sign-offs (let me know in the comments), contrast framing (it's not about X, it's about Y) and heavy signposting (here's the thing, let's break it down). The multipliers are strength ratings, not laboratory measurements: they tell you which hits to fix first, not the odds that a passage came from a model.
A sample of the strongest tells
- as an AI language modelChatbot Mannerisms25x
- as of my knowledge cutoffKnowledge Cutoffs25x
- up to my last trainingKnowledge Cutoffs24x
- I cannot offer medical adviceChatbot Mannerisms22x
- I hope this helpsChatbot Mannerisms20x
- I don't have access to real-timeKnowledge Cutoffs20x
- delveOverused Vocabulary18x
- you're absolutely rightChatbot Mannerisms18x
- curious what others thinkFake Engagement18x
- delving intoOverused Vocabulary17x
- continues to captivatePromotional Language16x
- I'm happy to helpChatbot Mannerisms16x
19 categories in total: Overused Vocabulary, Symbolism & Importance, Superficial Analysis, Promotional Language, Structure & Conclusions, Attribution & Media, Chatbot Mannerisms, Punctuation Pattern, Knowledge Cutoffs, Transition Phrase, Hedging Language, Filler Phrase, Formal/AI Word, Emphatic Phrase, Academic Phrase, Trendy AI Phrase, Fake Engagement, Contrast Framing, Over-Signposting.
How to use it with the AI detector
The two tools answer different questions. The AI detector gives a score and highlights the sentences that pushed it up; it reads rhythm, predictability and structure, not just vocabulary. This page tells you which words are doing the damage, so it is the one to open when the detector has flagged a passage and you want to know what to change.
A workflow that works: scan the draft with the detector, paste the flagged section here, rewrite the highest-multiplier hits in your own words, then scan again. If the score barely moves, the problem is not the vocabulary but the shape of the sentences, and the humanizer is the better tool because it rewrites for rhythm and re-checks its own output with the detector.
Two cautions. Humans use every phrase on this list, so an isolated hit is not evidence of anything; look at density, not presence. And stripping tells out of text that a model wrote does not make it yours. The useful case is a draft you wrote with a model's help, or a draft that has absorbed the model's habits, that you want to sound like you again.
What guests and accounts see
Without an account you see the first four matches with their multipliers and categories. A free account shows every match plus the category summary; each scan uses one daily credit, and Pro removes the daily cap. The matching itself happens in your browser, so the text never leaves your device for this tool. Copied the text from a chat window? Run it through the invisible character remover first so hidden characters do not split words the matcher is looking for.
FAQ
AI phrases questions, answered.
- What counts as an AI phrase?
- A word or phrase that appears far more often in text from ChatGPT, Claude, Gemini and similar models than in human writing on the same topics. The list holds 300+ of them across 19 categories: overused vocabulary such as delve and tapestry, symbolism such as stands as a testament, stock transitions, hedging, filler, over-signposting, fake engagement sign-offs and the it's-not-X-it's-Y contrast framing.
- Does using these words mean my text was written by AI?
- No. Every phrase on the list is legitimate English that people used before language models existed. A single hit means nothing. A cluster of high-multiplier hits in a short passage means the text reads the way models write, whether a model wrote it or you absorbed the style. Treat each highlight as a question, not a verdict.
- What does the multiplier mean?
- How many times more often the phrase turns up in AI-generated text than in human writing, by our estimate. 18x next to delve means it is roughly eighteen times more common in model output. The numbers are strength ratings drawn from published corpus comparisons and our own benchmark samples, not precise statistics, so use them to rank what to fix first.
- How is the list built and kept current?
- Entries come from corpus studies that compare word frequencies before and after 2022, from the tells editors use to spot model-written text, and from the phrases our detector benchmark keeps seeing. New patterns are added as they spread, which is why the list now includes trendy LinkedIn phrasing, fake engagement sign-offs and contrast framing alongside the older academic tells.
- Is my text uploaded?
- No. The phrase matching runs entirely in your browser against a list shipped with the page. Signing in only records that you used a credit; the text itself stays on your device.
- Can I see the results without an account?
- Yes, the first four matches with their multipliers and categories. A free account shows every match and the full category summary; each scan on a free account uses one daily credit, and Pro scans without the daily cap.
- Will removing these phrases make my text pass an AI detector?
- Not by itself. Detectors also weigh rhythm, sentence-length variation and how predictable each word is. Swapping delve for look into fixes one tell. Use this page to find the wording, the AI detector to see the score with sentence highlights, and the humanizer if you want the whole passage reworked and re-checked.