The words that make writing sound like AI

You read your draft back and nothing is wrong with it. Every sentence is fine. Together they sound like every other post in the feed: a warm up opener, a big verb, a hedge and a closing question nobody asked. For the whole workflow, from brief to approval, see our practical guide to AI writing for X.

Those are the AI words to avoid. This list sorts them into six types, gives each a plain rewrite, and says what the research does and does not show. One rule first: no word on this list proves who wrote a post. They are signs of padding, and readers skim padding whoever produced it. Plenty of padding is human too; see filler words to cut from your X posts.

Why do some words make writing sound like AI

Language models lean on some words far more than people do, and readers have started to notice. The clearest evidence comes from science writing, where researchers could count.

A team led by Dmitry Kobak looked at more than 15 million biomedical abstracts from 2010 to 2024. For every word they compared 2024 use with what the 2021 and 2022 trend predicted. After chatbots arrived, a set of style words rose far above that line.

The kind of word changed, not only the number. The excess in the Covid years was things, like "respiratory". The excess in 2024 was almost entirely style words, and 66% of those 379 words were verbs.

Two limits apply. The study covers PubMed abstracts, not X posts, and it counts words across millions of texts. It cannot say how any one text was written. We use it as proof that the pattern is real, not as a ranking of words for X.

Nobody has a settled answer for why it happens. A 2024 preprint by Tom Juzek and Zina Ward tested model design, algorithms and training data and found no evidence for any of them. Its results were consistent with a part for human feedback training, but not conclusive.

Do these words prove a post was written by AI

No. Two facts hold at once: detectors fail, and readers still notice.

• Detectors misfire. A 2023 study ran 7 detectors on 91 essays by non native English speakers. On average they wrongly flagged 61.3% as AI generated. The authors' explanation: detectors reward unpredictable wording, so plain vocabulary scores as machine like.
• A maker gave up. As of 20 July 2023, OpenAI's own AI text classifier was no longer available. OpenAI said the reason was its low rate of accuracy.
• Readers still notice. In a 2025 study, people who often use chatbots for writing read 300 non fiction articles. The majority vote of 5 of them misjudged just 1. They leaned on "AI vocabulary" and on formality, originality and clarity.

The last study used long articles, not short posts, so read it as a reason to care and not as a test you can run on a post. Words like these cost you with readers. They never prove a thing about the writer.

The six types of AI words to avoid

Our AI phrase checker sorts its more than a hundred entries into six types. This list uses the same six, so the two always agree. The types differ in where they sit in a post and in how you fix them.

Read a draft in that order: first line, last line, then the sentences between. The first two are the quickest wins. The checker marks words in red and sentence patterns in amber, so you can see both kinds at once.

Filler openers to cut

Cut any opener that announces the post instead of starting it. Delete it and the first real sentence usually works as the opening.

It matters more on X than in an article. A post holds up to 280 characters, so a warm up is room you do not get back. Nielsen Norman Group's eye tracking work on lists found people typically see about two words of each item, so the first words carry the most weight. That research is about lists and links, not posts, but the habit it describes is the same.

First lines get their own guide: how to write X hooks covers seven structures and the rewrites that go with them.

Inflated verbs to swap for plain ones

Swap the big verb for the small one: use, show, help, improve. Then ask how. The big verb is usually standing in for a fact you have not given yet.

Verbs are where the research is clearest. In the 2025 study, delves was used about 28 times more than the trend predicted, underscores about 14 times and showcasing about 11 times. Words as ordinary as crucial rose too.

Swapping the verb is half the job. Streamline how? Elevate by how much? Answer the question and the sentence earns its place.

Hype words that claim instead of show

Hype words declare an impact the reader cannot check. Replace each one with the fact it stands for, and if there is no fact, cut the sentence.

• Claims no one can check. World class, best in class, unparalleled. Name the standard or the customer instead.
• Intensity with no unit. Game changer, groundbreaking, mind blowing. Show the before and after.
• Mood words. Vibrant, tapestry, realm. Describe what you actually saw.

Wikipedia's field guide lists crucial, pivotal and vibrant among the words that AI text overuses. The 2025 study found crucial used well beyond the trend. For tapestry and realm we know of no study measuring a rise, so we make no claim about them. They are on our checker's list because they almost never fit a post about work.

Hedges and padding that dilute the point

Our checker puts it well: one hedge is caution, three is fog. Keep a hedge only when you really are unsure, and say what you are unsure about.

Most hedges we see are not doubt, though. They are habit, and they spend characters. The Federal Register's clear writing guide gives the same swaps: in many cases becomes often, and because of the fact that becomes since or because.

The third row is a real rule from X's developer documentation. Notice that the plain version is not only shorter. It reads as more certain, because the writer stopped asking permission to say it.

Cliché closers and engagement bait

End on the last useful sentence. If you want replies, ask a question only your readers can answer, or ask nothing.

Closers are easy to copy. What do you think? could follow a post about anything, so it gives no one a reason to answer this post in particular. A specific question gives people something to answer.

• What do you think? becomes a choice: Which would you pick, A or B?
• Follow for more needs a reason: say what you post and when.
• Let me know in the comments becomes the actual question.
• Let that sink in and Food for thought can simply go.
• Stay tuned needs a date, or nothing.

Sentence shapes that give a draft away

Some tells are shapes, not words. Each is fine once. Used by habit, they are a common sign of a template, because a reader sees the same frame in post after post.

• The contrast frame. It's not just a checklist, it's a whole new way to start. State the second half directly: what the checklist does.
• The two step reveal. Not just X, but Y. Lead with the more important half.
• The self answered question. The result? Merge the answer into the sentence.
• The pointer. Here's why: Cut it and start the explanation.
• The label. Spoiler:, Pro tip:, Hot take:. Let the sentence stand alone.
• The slogan. No X. No Y. Just Z. Say what it does.
• The stage direction. Read that again, Let me explain. Cut it and explain.

Three more are visual: the spool and pointing finger markers at the end of a thread opener, stacks of rocket and fire emoji, and two em dashes wrapped around an aside. The checker flags all three. Wikipedia's guide lists heavy em dash use too, and plenty of careful writers use them, so decide for your own voice. A comma, a bracket or a full stop does the same job.

When a flagged word is fine to keep

Keep a word when you mean it and no plainer word is as exact. A checker finds patterns, not meaning, so the last decision is yours.

The same word can be filler in one place and exact in another. Robust in a launch post claims nothing. In a test report it can describe how a result held up under changed conditions. Three questions sort most flagged words: did you mean it, is there a plainer word, does the sentence need it.

How to edit a draft for AI words in ten minutes

Edit in a fixed order so you do not rewrite the whole post. Start at the edges, then work through the marks.

1. Read the first and last line

   Cut any opener or closer from the lists above. Start on the point and end on the last useful sentence.

2. Run the checker

   Paste the draft into the AI phrase checker. It runs in your browser, so nothing is uploaded.

3. Decide on each mark

   Keep, swap or cut. Swap when a plain twin exists. Cut when the sentence works without it.

4. Add one detail where you cut a claim

   A number, a name, a before and after. Only add what is true, and say it the way you would say it to a colleague.

5. Read it aloud

   If you would not say a line, change it. Then check the length in the character counter.

If you draft with an AI model, put your never list in the brief. In Kairo Post, each account keeps its own list of phrases to avoid, and AI drafts follow it. You can see how that fits together on the Brand brain page.

Mistakes we see when people try to sound human

The most common mistake is fixing the word and keeping the empty claim. Delve into becomes dive deep into, and the sentence says the same nothing in a fresh coat of paint. Wikipedia's editors make the same point about their own list: do not merely treat the signs as problems to be fixed, because that can just make detection harder.

• Adding typos or slang to fake a voice. It reads as fake, and it makes the post harder to read. Write the way the account really sounds.
• Chasing zero marks. A draft with no marks can still be vague. The marks per 100 words figure in the checker is for comparing two drafts of the same post, not for grading quality.
• Banning words, not habits. If leverage is blocked and harness appears, the habit is the problem. Block the pattern, then add the detail.
• Accusing someone. A mark is a prompt to reread. It is not a finding about the writer.

If you write for a brand, turn the marks you keep catching into a short never list for that account. We cover how a voice sheet works in our guide to managing X for clients.

Before your next post: a checklist

Run this on the next draft. It takes about two minutes once you have done it a few times.

Paste the draft into the AI phrase checker for a second pair of eyes, then decide mark by mark. When the first line still feels weak, grade it with the hook grader. A thread needs the same pass on every post, and how to write an X thread shows where each one does its job.