A Simple Checklist for Removing AI Tells From Any Draft

Most AI-generated text fails on feel, not facts. The information is often fine. What gives it away is the rhythm: every sentence roughly the same length, the same phrasing habits repeated across paragraphs, a conclusion that wraps everything up with hollow warmth. Readers feel it before they can name it.

This checklist walks through the specific patterns that flag AI copy, with concrete edits to fix each one. Work through it top-to-bottom after your first AI draft, before you publish anything.

The checklist

1. Sentence length: break the drone

Read your draft aloud. If every sentence takes about the same breath, that's the AI pattern. Good prose mixes long and short. Short sentences hit.

Open a paragraph and count the word lengths of each sentence. If the variance is low (most sentences between 18 and 26 words), start cutting. Chop one long sentence into two. Then combine two short ones. The goal is variety, not any particular length.

For more on this, see varying sentence rhythm to break the AI pattern.

2. The opener problem: kill your first sentence

AI almost always starts with context-setting. "In today's fast-paced environment..." or "When it comes to X..." or a restatement of the topic. Cut the first sentence and see if the second one can open the piece instead. Often it can, and the draft immediately reads sharper.

3. Banned vocabulary list

Run a find-and-replace pass for these words. If any of them appear, rewrite the sentence from scratch. Patching around a banned word rarely works.

• delve
• tapestry
• testament (as in "a testament to...")
• seamless
• robust
• realm
• vibrant
• pivotal
• crucial
• underscore (as a verb)
• foster
• garner
• leverage (as a verb)
• elevate
• unlock
• intricate
• multifaceted

These words aren't wrong in isolation. The problem is that AI returns to them constantly, so their presence signals machine authorship. Replace them with plain language or a more specific word. "Crucial" becomes "important" or gets cut entirely because the surrounding sentence should already imply urgency.

4. The "serves as" trap

AI uses "serves as," "stands as," and "boasts" as filler connectors. "This tool serves as a bridge between X and Y." Cut it. Write what the thing actually does: "This tool connects X to Y." Shorter, clearer, and it doesn't sound like a product brochure.

5. Rule-of-three filler

AI loves tidy triplets. "Faster, cleaner, and more reliable." "Simple, effective, and easy to use." When you see three adjectives or three parallel phrases, cut to one or two. Ask which of the three actually matters most. Use that one.

6. Em dash overuse

Em dashes aren't wrong — but AI stacks them. If you find more than two in a draft, replace the extras with commas, parentheses, or a new sentence. Two is fine. Six is a tell.

7. The -ing tail clause

Look for sentences that end with a participial phrase: "..., highlighting its value," "..., demonstrating the importance of consistency," "..., showcasing what's possible." These phrases add length without adding meaning. Cut them or rewrite the sentence so the point lands at the end of the main clause, not in an afterthought.

8. Negative parallelism

"It's not just a tool, it's a solution." "Not only does it save time, but it also improves quality." AI uses this construction to sound emphatic. It ends up sounding promotional. Rewrite as a direct statement of what the thing actually does.

9. The generic conclusion

AI conclusions usually: summarize what was just said, end with an encouraging sentence about the reader's ability to succeed, and use phrases like "by keeping these tips in mind" or "with the right approach." Write a real ending instead: a specific next step, a concrete example, or a note about what to watch out for.

10. The non-attribution attribution

"Studies show..." "Research suggests..." "Experts agree..." These land differently than you might think. Without a specific source, they read as AI filler. Either cite the actual study (publication, year, finding) or cut the phrase and just make the claim. "Most people find this approach easier" beats "research suggests that individuals prefer..."

---

Before and after

Here's the kind of sentence AI produces, followed by a rewrite.

Before (AI draft):

Leveraging the power of structured prompts can serve as a robust mechanism for fostering more authentic, human-sounding output, underscoring the pivotal role that deliberate editing plays in the content creation process.

Everything on the banned list shows up here. Let's rewrite it as a human editor would.

After:

Structured prompts help, but they won't carry you the whole way. The draft still needs an editor.

Same idea. A quarter of the words. No vocabulary tells.

---

How to run the checklist efficiently

Don't try to fix everything in one pass. Pick two or three items per edit session.

First pass: vocabulary sweep. Find-and-replace for the banned word list, then rewrite any sentence that contained one.

Second pass: rhythm. Read aloud, break up long sentences, combine short ones.

Third pass: structure. Check the opener, the conclusion, and the -ing tail clauses.

This takes 20 minutes on a 1,000-word draft. Longer pieces need longer, but the process is the same.

If you want a prompt that handles some of this mechanically before you edit by hand, the free humanizer prompt at /humanizer-prompt is a reasonable starting point. It won't replace manual editing, but it reduces the initial cleanup work.

For a deeper walkthrough of the manual editing process, see how to edit an AI draft so it reads like a human wrote it and how to rewrite a robotic AI paragraph by hand.

---

FAQ

How do I know if my draft sounds like AI?

Read it aloud. If it sounds like a company brochure or a Wikipedia summary (polished but slightly flat, with no opinion anywhere), that's the feel. Specific tells: all sentences roughly the same length, no contractions (or contractions placed artificially), a conclusion that summarizes without adding anything.

Does this work on all AI models?

The same patterns show up in output from ChatGPT, Claude, Gemini, and most other large models, because they share training data and optimization targets. The banned vocabulary list is fairly universal. Sentence rhythm and conclusion style vary a little by model, but the checklist applies to all of them.

Will editing like this fool AI detectors?

Probably, for most detectors, if you do thorough work. But that's a moving target and not the main point. The goal is to write copy that works for human readers, not to beat a classifier. Copy that reads well almost always clears detectors anyway because it actually varies from the statistical baseline the detectors are looking for.

How much editing is actually required?

More than people expect on a first pass. A 10-minute skim won't do it. A thorough edit on a 1,500-word piece takes 30-45 minutes when you're applying the whole checklist. If time is short, prioritize vocabulary and rhythm. Those two changes have the biggest impact on feel.

Can I use AI to help edit AI text?

Yes, with a caveat. Asking the same model to "make this sound more human" often produces minor surface changes while keeping the underlying pattern. You get better results with a specific instruction: "Rewrite this sentence without using any of the following words: [list]." Or: "Shorten this to one sentence." Concrete instructions outperform vague ones.