The giveaways are not a vibe, they are measurable. Below is the list of the actual tells, each with why it reads as AI and what to write instead, then a copy-paste pass you run on any draft to clean them out. The real fix, though, is not deleting words, it is teaching the tool your voice, and there is a section on that too.
AI writing gives itself away with a small set of overused words (delve, underscore, pivotal, showcasing), puffery, the "it is not just X, it is Y" rhythm, a flood of em dashes and bold, warm-up openers and a tacked-on summary. The durable fix is teaching the tool your own voice, not just banning the words.
First, the proof it is real
This is measured, not a hunch. A 2025 Science Advances study of fifteen million research abstracts found words like delve and underscore spiked right after ChatGPT, enough to estimate one in ten had been AI-assisted. Wikipedia keeps a living list of the tells too.
The tells also move faster than any fixed list, which is the part most guides miss. One independent analysis of around sixteen thousand Substack posts found the giveaway is no longer the "it is not just X, it is Y" line on its own, it is repeating that same shape within one piece: of the posts that used it at all, a quarter used it three or more times. So the thing to watch for is not a single word, it is a habit that keeps recurring. The full analysis is here, and it is the kind of thing you only find by digging past the first page of results.
The tells, and what to write instead
These are the ones that give it away most. Scan for them in anything you or an AI wrote.
| The tell | Why it reads as AI | Write instead |
|---|---|---|
| Delve, delve into | The single most outed AI word. Its use exploded in real writing after ChatGPT. | Look at, get into, or just cut it. |
| Underscore, pivotal, showcasing, intricate | The "excess vocabulary" the study measured spiking. | Plain words: shows, key, detailed. |
| A testament to, rich tapestry, the realm of | Puffery that inflates everything into a big deal. | Say the actual thing, plainly. |
| "It is not just X, it is Y" | A rhythm the model loves for fake profundity. | Make the point once and move on. |
| Em dashes everywhere | AI reaches for the em dash constantly. It is such a strong tell that this whole site bans them. | A comma or a full stop does the job. |
| Bolding half the sentence | The model bolds like it is marking a textbook. | Bold rarely, or not at all. |
| "In today's fast-paced world..." | A warm-up opener that says nothing. | Start on the actual point. |
| The tidy rule of three | Three neat items every single time reads generated. | Vary it. Use two, or four, or one. |
| "Experts say", "studies show" | Vague authority with no name, often invented. | Name the source, or drop the claim. |
| A summary tacked on the end | AI wraps up neatly when nobody asked it to. | Stop when you are done. |
Tells drawn from Wikipedia's Signs of AI writing and the Science Advances 2025 excess-vocabulary study, both linked above.
Here is what they look like stacked into one sentence, then the same thing cleaned:
Before: In today's fast-paced world, it is not just a tool, it is a testament to how AI can delve into the intricate realm of your writing.
After: AI can help you write faster. Here is how.
The real fix is your voice, not a banned-word list
Here is the honest bit most tells guides skip. The list above is a moving target. As soon as everyone learns that "delve" is a giveaway, the models drift and the readers recalibrate, and a fresh set of tells takes over. If all you do is find and replace the words, you are always a step behind. The durable fix is to teach the tool what you actually sound like, so it writes in your voice from the start instead of the neutral AI one you then have to scrub. Two ways to do that: paste two or three real things you have written into the pass below, or use Claude's built-in Styles, where you upload samples of your writing and it matches them. The pass is the quick clean-up, the voice is the long game.
The de-AI pass
Paste this into Claude or ChatGPT with your text underneath. It finds the tells, rewrites them out, and tells you what it changed so you learn them for next time. Add a couple of your own writing samples and it will keep your voice while it cleans.
Do a de-AI pass on the text below, in three steps, and show your work.
Step 1, find the tells. Read it and list every giveaway, quoting the exact words:
- AI words: delve, underscore, pivotal, showcasing, intricate, testament, tapestry, realm, and their cousins.
- Puffery that inflates everything into a big deal.
- The "it is not just X, it is Y" line, and any shape or rhythm that repeats across the piece (the same construction, or a rule of three, used again and again). The repetition is the bigger tell than the phrase.
- Em dashes, half-sentence bolding, and warm-up openers like "in today's world".
- Vague authority ("experts say", "studies show") with no real name.
- A summary tacked on the end that nobody asked for.
Step 2, rewrite it, fixing every tell you found:
- Plain words a real person would say out loud. No em dashes, use a comma or a full stop.
- Say the actual thing instead of praising it. Cut the warm-up and the tacked-on summary.
- Keep every real point and fact from the original. Change how it is said, not what it says.
- Vary the sentence length so it reads natural, not choppy or robotic.
- Never invent a fact, a name or a number to fill a gap. Leave a clear gap in brackets for me instead.
- If I pasted samples of my writing below, match that voice. If not, keep it plain and natural, do not invent a personality.
Step 3, report. Give me a short list of what you changed and why, so I learn the tells for next time.
[PASTE YOUR TEXT HERE]
[Optional, so it sounds like me: paste two or three things I actually wrote, and match that voice.]
See the tells light up in your own writing
The fastest way to learn them is to watch them appear in something you wrote. Simon Willison, an engineer who writes closely about AI, built a free browser tool that does exactly that: paste your text into his LLM cliché highlighter and it marks every sentence that matches a known tell, no signup, nothing to install. It beats an AI detector, which only hands you a score and gets it wrong in both directions. This shows you the actual phrases to fix, which is the part that teaches you to spot them yourself.
If you live in Claude Code and want a reusable version, there are free open-source humanizer skills that score and rewrite your text on your own machine, like this one. For everyone else, the highlighter plus the pass above is plenty.
Where the line is
- The honest bit: this list will age. The tells are real today and backed by the research above, but the models and the readers both move. Treat the words as a starting point and lean on the voice fix, which does not go stale.
- A detector score is not proof. It is a signal that gets things wrong in both directions. Do not use one to accuse a student, a colleague or yourself of anything. The tells you can see with your own eyes are worth more than a number.
- If the rewrite comes back stiff or choppy, you stripped too hard. A pass with no samples flattens everything to a neutral, clipped tone. Paste more of your own writing and tell it to keep your rhythm, not just cut words.
- Not everything needs this. Run the pass when your name is on it or it goes public. A two-line message to a mate is fine as it is, so do not police yourself into stiffness.
Run the pass on one thing today
Take the last thing you wrote with AI, an email, a caption, a post, and run the pass on it with two of your own real messages pasted underneath. Read the before and after side by side. That one comparison teaches you your own tells faster than any list, and after a few passes you start catching them as you write.
Want the output to sound like you from the first draft, not after a clean-up? That is the whole job of the Voice File.
A couple of quick questions
What are the biggest tells that writing was made by AI?
Overused words like delve, underscore, pivotal and showcasing, puffery that calls everything a testament or a rich tapestry, the "it is not just X, it is Y" rhythm, a flood of em dashes and bold, warm-up openers, and a tidy summary tacked on the end.
How do I make AI writing not sound like AI?
Do not just delete the tell words. Teach the tool your voice with real samples of your own writing, then run a de-AI pass that cuts the tells and keeps your rhythm. Voice-matching is the fix that lasts.
Can AI detectors tell if something was written by AI?
Roughly, not reliably. A detector gives you a signal, not a verdict, and it gets it wrong both ways. Use one to sanity-check your own draft, never to accuse anyone.