Short answer: writing sounds like AI when stock phrases pile up and the sentences settle into one length. Typical examples are “it is important to note”, “plays a crucial role”, “a wide range of”, and chains of “Furthermore”, “Moreover” and “Additionally”. To check your own text, search for these phrases, look at how close together they sit, and measure how much your sentence lengths vary. None of this shows who wrote the text. It shows where the writing could be more specific.
There’s some evidence behind the impression that these words spread. A study of biomedical abstracts by Kobak and colleagues tracked the excess use of certain style words after language models appeared, and estimated a lower bound of 13.5% for the share of 2024 abstracts processed with them. That describes a body of published text. It can’t tell you anything about one person’s paragraph.
If you want a longer dictionary of clichés with suggested replacements, see why AI text sounds robotic. This guide is about checking a text you already have.
The phrases to look for
These groups cover the patterns you’ll meet most often. All of them are ordinary English, so what matters is how often they appear and what they contribute.
| Group | Examples | Question to ask |
|---|---|---|
| Scene-setting openers | “In today’s digital age”, “in the ever-evolving”, “in the realm of” | Does the point start in the next clause? Then cut this one. |
| Announcements | “It is important to note”, “it is worth noting”, “needless to say” | Does the sentence lose anything without them? |
| Empty importance | “Plays a crucial role”, “a testament to”, “pivotal”, “invaluable” | What exactly does it do, and for whom? |
| Stacked transitions | “Furthermore”, “moreover”, “additionally”, “ultimately” | Is the logical link between the sentences actually clear? |
| Inflated verbs | “Leverage”, “utilize”, “facilitate”, “delve”, “underscore” | Would “use”, “help” or “show” say the same thing? |
| Hype | “Game-changer”, “cutting-edge”, “groundbreaking”, “seamlessly” | What’s the evidence or the concrete detail? |
How to check your text by hand
- Search for the phrases. Use Ctrl+F or Cmd+F on the examples above, and search for stems such as “facilitat” and “leverag” to catch every form.
- Look for clusters. Highlight each hit. One phrase in a page means little. Five in one paragraph tell you that paragraph needs attention.
- Read the first word of every sentence. Several transitions in a row, or the same structure repeated, stand out quickly when you read only the openings.
- Count the words in a few sentences. Pick a paragraph and write down the length of each sentence. If the numbers stay close together, the rhythm will feel flat, whoever wrote it.
- Ask what’s specific. Underline every number, name, example and source. A paragraph with no underlining is often generic, and that’s a bigger problem than any single word.
When these words are fine
A phrase list can’t read for meaning, so expect false alarms. Keep the word when it’s doing a precise job:
- “Significantly” in a results section can report a statistical test. Replacing it with “greatly” would change the claim.
- “Robust” has technical uses in statistics and engineering, as in “robust standard errors”.
- “Facilitate” can be a term of art. “Social facilitation” is a concept in psychology, and a “facilitator” may be someone’s job.
- “Consequently” is correct when one thing really does follow from another.
- “In conclusion” and “comprehensive” are plain and sometimes required, as in a “comprehensive exam”.
These words also appear in plenty of text written without any AI, so never treat them as evidence against a person. Our review of whether AI detectors work explains why even a formal detector score is an estimate, not proof of authorship.
A worked example
Here’s a short invented draft, followed by a version edited by hand. We ran both through the same calculation that humai’s naturalness report uses.
Before: “In today’s digital age, remote work plays a crucial role in how companies hire. Furthermore, it offers a wide range of benefits for employees. Additionally, it facilitates better work-life balance. Moreover, companies can leverage a larger talent pool. Ultimately, remote work is a game-changer for the modern workplace.”
After: “Remote work has changed how companies hire. Employers can recruit from a larger pool of candidates, because location no longer limits the search. Employees get something too. Without a daily commute, work can be easier to fit around the rest of life. Ultimately, the change affects both sides.”
| Check | Result |
|---|---|
| Phrases found in the original | 10 |
| Removed | 9: “In today’s digital age”, “plays a crucial role”, “Furthermore”, “a wide range of”, “Additionally”, “facilitates”, “Moreover”, “leverage”, “game-changer” |
| Still in the text | 1: “Ultimately” |
| Sentence lengths in words, before | 14, 10, 6, 8, 10 (variety score 28) |
| Sentence lengths in words, after | 7, 16, 4, 15, 6 (variety score 51) |
Those numbers need careful reading. The lengths are more varied after the edit, and nine stock phrases are gone. But the numbers can’t tell you whether the new text is true. The edit also changed the content: “plays a crucial role in” became “has changed”, and it added a reason (location no longer limits the search) and a mechanism (no daily commute) that the original never gave. Those additions make the paragraph more concrete. They’re only acceptable if the writer means them and can support them. “Ultimately” stayed because it closes the paragraph naturally, and that’s a reasonable call.
The naturalness report in humai
When the humanizer changes your text, it shows a naturalness report. Everyone gets it, it’s the same on Free and Pro, and it’s calculated by fixed rules without an AI model. It contains three parts:
- Phrases typical of AI writing. The phrases found in your original, up to 20, each marked “removed” or “still in the text”.
- Variety in sentence length. A before and after number. A higher number means your sentence lengths vary more. With fewer than four sentences, there’s too little to measure and the report says so.
- Copy-and-paste cleanup. How many long dashes, markdown marks and invisible characters were cleaned up.
The report describes writing style. It doesn’t give a probability of AI authorship and it doesn’t measure detection in any form. It also has limits worth knowing. It starts from phrases in your original, so a stock phrase the rewrite introduced won’t appear on the list, which means you should still read the result. And variety can get worse after a rewrite. The report shows the numbers as they came out.
humai’s separate AI detector gives an overall estimated style score with the features behind it. Use it to find passages to revise. It isn’t a verdict on who wrote them.
What to fix first
Work in this order, since the first items change the most:
- Replace generic claims with specifics. Add the real number, name or example, if you have one. If you don’t, consider whether the claim should stay.
- Show the logic between sentences. Swap “Moreover” for a sentence that makes the connection clear, or cut the transition if the connection is obvious.
- Delete announcements. Remove “it is worth noting that” and start with the thing worth noting.
- Vary length where the content allows it. Join two related short sentences, or give an important point its own short one.
- Keep precise terms. Leave technical words alone, and see our guide on keeping keywords when rewriting if a tool keeps replacing them.
If this text is for a course, removing stock phrases doesn’t alter what the rules require. Follow your university’s policy on AI and disclose your use where it asks.