What the AI phrase finder looks for
Chatbots like ChatGPT, Claude and Gemini have favorite words. Ask for an email, a product description or an essay, and the same handful keep coming back: delve, intricate, tapestry, pivotal, seamless, a testament to, plays a crucial role. Add a few habits, like starting sentences with Moreover, building every point as "not just X, but Y" and closing with "I hope this helps!", and the text starts to sound like every other chatbot answer.
This style checker scans your text for more than 120 of these words, phrases and patterns (about 300 spellings and word forms in all), highlights each one in place, and sorts them into six groups:
- Overused AI-era words such as delve, underscore, showcase, meticulous and realm.
- Stock phrases such as "it's important to note", "in today's fast-paced world" and "navigate the complexities of".
- Overused transitions such as Moreover, Additionally, Furthermore and In conclusion.
- Sentence patterns: "not just X, but Y", "it's not X, it's Y", "whether you're a X or a Y" and long dashes.
- Chatbot leftovers: "Certainly!", "I hope this helps", "let me know if" and "as of my last update".
- Filler intensifiers such as very, really, truly and incredibly.
Every finding comes with plain alternatives, and the list under the text counts how often each one appears. Switch groups on or off with the buttons above the text. The score at the top is simply the number of findings per 100 words. The lists are for English text.
Where the word list comes from
The single words are not guesses. Most of them come from studies that compared large collections of texts written before and after ChatGPT arrived at the end of 2022:
- Kobak and colleagues, Science Advances, 2025. The team looked at more than 15 million biomedical abstracts in PubMed from 2010 to 2024 and searched for words that suddenly became more common than earlier years predicted. In 2024, delves showed up about 28 times as often as the earlier trend predicted, underscores about 14 times and showcasing about 11 times. Earlier jumps, like during the Covid pandemic, were nouns about the topic, such as respiratory. The 2024 jump was style words, mostly verbs and adjectives. Every word in our list that cites this study appears on the authors' published list of 2024 excess style words.
- Juzek and Ward, COLING 2025. They found 21 words that both spiked in scientific abstracts and that ChatGPT used significantly more often than the human authors when it generated abstracts from the same originals, including delve, intricate, realm, boasts, garnered and groundbreaking.
- Liang and colleagues, ICML 2024. In peer reviews for AI conferences, adjectives like commendable, meticulous and intricate became many times more common after ChatGPT came out.
Two words with some of the biggest jumps, potential and findings, are left out on purpose: they are so ordinary that a single text says nothing with them. The stock phrases, sentence patterns and chatbot leftovers mostly come from Signs of AI writing, a field guide kept by Wikipedia editors who clean up chatbot text. It is a list of observations, not a study. The rest, like "game-changer", "at the end of the day" and the filler intensifiers, is general style advice. We mark those as style advice, not as AI markers. Click any highlight to see which source it comes from.
Word lists also age. A 2025 paper by Geng and Trotta found that delve dropped in arXiv abstracts soon after it was publicized as a ChatGPT word in early 2024, while other favorites such as significant kept rising.
Why this is not an AI detector
The research above works on millions of texts at once. Kobak and colleagues say plainly that their method works across the whole collection and cannot identify individual abstracts that went through a chatbot. That limit applies to any word list, including this one.
- People use these words. Chatbots learned them from human writing, and the influence now runs the other way too: a paper on more than 700,000 hours of podcast conversation by Yakura and colleagues found that people now say words like delve, meticulous and boast more often in unscripted talk since ChatGPT came out.
- Edited AI text can score zero. Anyone who deletes the obvious phrases removes the signal.
- False accusations do real harm. Even dedicated AI detectors get it wrong. In a study in the journal Patterns, Liang and colleagues ran seven popular detectors on 91 TOEFL essays written by non-native English speakers, and on average the detectors wrongly flagged 61.3% of them as AI. A word count is far cruder than those tools.
So read the score as a style signal about the text in front of you, never as a verdict about the person who wrote it.
How to rewrite text that sounds like a chatbot
Swapping delve for explore helps a little. Real improvement comes from saying something specific. Most stock phrases are there because the sentence has no concrete content yet. A few before and after examples:
| Before | After |
|---|---|
| I wanted to delve into the intricate details of our Q3 plan, which underscores our commitment to growth. | Here is our Q3 plan. In short: we want to grow sales in two new regions. |
| This app is not just a to-do list, but a powerful tool that can elevate your workflow. | This app keeps your to-do list and your calendar in one place. |
| Nestled in the heart of downtown, the cafe boasts a vibrant atmosphere. | The cafe is on Main Street, two blocks from the station, and it gets loud at lunch. |
| Moreover, the new policy seamlessly aligns with our values. | The new policy also fits how we already work: no meetings on Fridays. |
| Certainly! Here is a comprehensive overview of the options. I hope this helps! | Here are the three options. (Delete the greeting and the sign-off.) |
- Lead with the point. Delete "It's important to note that" and start with what is important.
- Replace praise with facts. Vibrant, groundbreaking and seamless tell the reader nothing they can check. A street name, a number or a date does.
- Say the positive claim once. "It's not X, it's Y" sets up a claim nobody made. Just say Y.
- Cut the frame. Opening lines like "In today's fast-paced world" and closing lines like "In conclusion" can almost always go.
- Vary your linking words. If three sentences in a row start with Moreover, Additionally and Furthermore, try "and", "also", or no linking word at all.
- Mind the dashes. The Wikipedia editors' guide notes that chatbots use long dashes much more than casual human writers, often where a comma or parentheses would do. One is fine. If every paragraph has several, try a comma, colon or period.
Pasted text from a chatbot often carries markdown symbols and hidden characters as well. The AI text cleaner removes those without changing a word, and the word counter checks length and readability once you are done.
How to use the AI phrase finder
- 1Paste your text into the box. The score, the counts and the highlighted copy update as you type.
- 2Switch off groups you do not care about, for example Fillers. Click a highlight, or a phrase in the list, to see plain alternatives and the source. Use Prev and Next to jump between the places it appears.
- 3Rewrite in your own words, then paste the new version to check again. Copy findings gives you the list as plain text, for example to send with feedback on a draft.
Frequently asked questions
Does using the word "delve" mean a text was written by AI?
No. Delve is an ordinary English word, and people used it long before chatbots. Studies show it became far more common in some kinds of writing after ChatGPT arrived, but that is a pattern across millions of texts, not a test for one text. A single delve, or ten, tells you nothing certain about who wrote something.
Can this tool tell if ChatGPT wrote a text?
No, and it does not try to. It is a style checker. It shows words and phrases that chatbots tend to overuse so you can decide whether your writing leans on them. People use every one of these phrases too, and a chatbot answer that was edited well can have none of them. Do not use the result to accuse anyone of using AI.
Why does ChatGPT use words like delve so much?
Nobody knows for sure. Juzek and Ward (COLING 2025) tested several explanations. In their tests, model design and training data did not explain it, and their results were consistent with a role for the step where models are tuned on people's ratings of answers, known as learning from human feedback. They call this a lead, not a final answer.
What is a good score?
There is no official scale. The score counts the findings per 100 words. As our own rule of thumb, under 1 reads plain, 1 to 3 has some stock wording, and above 3 is heavy. Fillers like very and really push the number up in anyone's writing, so switch that group off if you only care about chatbot style. Short texts swing a lot, because one phrase changes the number by several points.
Will removing these words make a text pass an AI detector?
This tool is not made for that and we make no such claim. Detectors look at more than a word list, and they are known to make mistakes in both directions. Use the tool to make your writing clearer for your reader. If you used AI for a school or work task, follow the rules that apply there and say so where it is required.
Why are normal words like crucial and comprehensive on the list?
Because research measured big jumps in exactly these words. Kobak and colleagues found crucial, comprehensive, additionally and notably among the ordinary words with the strongest excess use in 2024 biomedical abstracts. Using them is not wrong. They are on the list so you can see when they pile up.
Is my text uploaded anywhere?
No. The check runs in your browser. Your text is not sent to a server, not stored and not shared. Copy link saves only which groups are switched on, never the text.