Words and phrases AI overuses, and what to write instead

AI in 2026 loves words like "genuinely", "roughly", "honestly" and "rather than". It overused them in both of our studies: when it rewrote human text, and when it wrote from scratch. To polish, AI also uses "plenty of", "essentially" and "for instance". To write from scratch, it also uses "here's" and "whether you're". The words it used in 2023, including "delve", "tapestry" and "leverage", are now rarely used by AI. To not sound like AI, avoid filler words, provide real numbers instead of "roughly" and vary the rhythm of your writing. Changing words alone won't do it.

What words does AI overuse in 2026?

Current AI models use two main kinds of words far more than people do: useless intensifiers ("genuinely", "truly", "simply") and vague approximations ("roughly", "plenty of"). The most overused was 12.5 times more common in AI rewrites than in the human originals, and "honestly" was 8.3 times more common in drafts AI wrote from scratch. These words don't make the writing wrong; humans just use them a lot less, and readers notice the pattern even when they can't name it.

Here is the list, in order of how much more common each word was when AI rewrote human writing, with the from-scratch figure next to it and what to use instead.

Does AI use different words when it writes from scratch?

Yes and no. The words "honestly", "genuinely", "roughly" and "rather than" were overused by AI in both of our studies, so they're the safest ones to look out for. The other words depend on whether you asked AI to polish your text or to write from scratch. When polishing, AI adds "for instance", "plenty of" and "essentially". When it writes from scratch, it uses "for instance" and "essentially" less than people do, and "plenty of" only slightly more. Instead, it uses other words, including "here's" (3.9 times the human rate) and "whether you're" (7.1 times, as in "whether you're a beginner or a pro"). So if you asked AI to write from scratch using some notes you provided, watch out for the words it uses to write from scratch. If you asked it to polish, however, watch out for the words it uses to rewrite.

• Both rewriting and drafting habits: honestly, genuinely, roughly, rather than.
• Mostly rewriting habits: for instance, plenty of, essentially.
• Drafting habits: here's, whether you're.

Why does "just cut it" come up so often?

Because most of these words don't help the reader understand anything. "This is genuinely useful" tells your readers the same thing as "this is useful", only with more insistence. When every claim gets an intensifier, none of them land, and the writing starts to sound like it's trying to convince you.

Also, often you can replace these words with specific data. Instead of saying "roughly", you can say how much. Instead of saying "plenty of", you can say how many. If you can't say how much or how many, saying "about" is fine, but don't do it too often. AI uses a lot of these words because that's the safe thing to do, but readers see through it.

Note that "for instance" is a bit of an exception: while the bloggers in this set almost never used it, AI used it 0.25 times per 1,000 words, which is about once per 4,000 words. That makes it by far the most one-sided phrase on this list when rewriting. It's a rewriting habit, though: when the models wrote from scratch, they used it less than people did.

Are "delve", "tapestry" and "leverage" still AI words?

Not in our 2026 data, whether the models were rewriting or writing from scratch. Back in 2023, there were lists going around of words that AI models tended to use, such as "delve", "tapestry" and "leverage". At the time, those lists were accurate. Since then, the models seem to have been tuned to avoid those words. Just to give you some idea, here's how often some of them appear in our data:

• delve: zero, or close to it, in both the human originals and the AI rewrites.
• tapestry and testament: close to zero, again on both sides.
• leverage: 0.05 per 1,000 words for humans and 0.04 for AI when rewriting. No difference. When writing from scratch it was close to zero on both sides.
• moreover, furthermore, additionally: 0.01 per 1,000 words for humans and 0.02 for AI when rewriting, and about the same rate on both sides (1.0x) when writing from scratch. These words are so rare that comparing them doesn't tell you much.

There's a more extensive list of AI words in Wikipedia's guide to signs of AI writing. Note that it has different lists for different eras: "delve" was famously overused in 2023 and early 2024, then dropped off sharply in 2025. Its most recent list (for mid-2026 onward) includes words like "leverage", "foster" and "multifaceted". We didn't see "leverage" become more common in either study, when models polished human writing or when they wrote from scratch. But if you're still scanning drafts for "delve", you're behind the curve: it's a problem that has mostly gone away.

Do different AI models overuse different words?

Yes. A few examples from our rewriting data: the model that uses "plenty of" the most is Claude Opus 5.5, at 0.68 times per 1,000 words. It also leans on "isn't", "roughly" and "genuinely". The models that use "genuinely" the most are GLM-5.3 and MiniMax M3, at 1.25 and 1.09 times per 1,000 words. The model that uses "actually" the most is also MiniMax M3, at 2.43 per 1,000 words (humans use it 0.65 times).

If you're only using one AI model, read five pieces of writing it's produced for you and see if any words appear in all of them.

Will swapping these words make AI text sound human?

Not by itself. If you stop the model from using "genuinely", it will just use "really" instead. And if you stop it from using "really", it'll use "truly". The model is going to use some kind of intensifier, because that's what it thinks the sentence needs.

More important than specific words are the patterns of rhythm and structure in AI writing. In a held-out sample of 471 pairs from our rewriting study, we found that human writers varied sentence lengths much more than the models did (the coefficient of variation in sentence length was 0.65 for humans, 0.59 for the OpenAI models and 0.56 for Claude). We also found that the four bloggers started more of their sentences with "And", "But" or "So" (12.0% of sentences for the bloggers, 9.7% for the OpenAI models and 5.0% for Claude).

The gap was smaller when it wrote from scratch. While there was slightly more variation in the length of sentences in the originals (a coefficient of variation of 0.50 in the originals vs 0.47 in the drafts; the most uniform rhythm was with GPT-5.4 mini at 0.43), the percentage of sentences starting with "And", "But" or "So" was about the same (3.8% for people vs 3.4% for AI), which suggests that some of the 12.0% came from the specific habits of the four bloggers. Either way, when AI writes, it smooths things out, and that's a large part of what makes it sound like AI.

And there are structural signs too, like three items in every list, the tidy summary at the end, or the "it's not X, it's Y" reveal. You can see all of them on our list of signs of AI writing, and you can learn how to fix them in our guide to making AI writing sound human.

Can I give this list to ChatGPT or Claude?

Yes, just copy and paste this into your custom instructions, project prompt or agent rules file. It won't fix rhythm, but it stops the most common filler before you have to edit it out.

When writing for me, avoid these words and phrases unless they carry real meaning:
genuinely, genuine, truly, honestly, essentially, simply, actually,
roughly, plenty of, for instance, rather than, crucial, navigate,
here's, whether you're.

- Cut intensifiers instead of swapping in a synonym (no "really", "deeply", "incredibly").
- Replace vague amounts ("roughly", "plenty of") with the real number, or "about" once.
- Use "for example" or just give the example.
- Use "instead of" or two short sentences instead of "rather than".
- Vary sentence length. Mix short sentences with longer ones.
- It's fine to start a sentence with And, But or So.
- No em dashes.

This page will be updated as new models are released. The words on it today won't be the ones that matter a year from now, which is the main lesson of the 2023 lists.

If you'd rather not police word lists by hand, an easier way is to use Zavus to rewrite the entire finished draft. Instead of just avoiding certain words, Zavus uses a separate rewriting model that restructures sentences and chooses different words. You can use it within Claude, ChatGPT, Cursor and many other agents. Our Zavus page for marketers shows how teams use it on posts and emails.