Writing YouTube scripts with AI without the AI tells

Most AI-written YouTube scripts get clocked in the first 30 seconds. Not because viewers are reading frame-by-frame for telltale phrases, but because the rhythm is off. The sentence lengths cluster. The transitions all sound like an essay. And the same eight or nine phrases show up across thousands of videos at the same time.

Here is the actual list of patterns that gives an AI script away, in the order we've seen them cost retention on our own channels. Each one is checkable with a regex, which means a deterministic post-generation scan catches them before the script ever hits a teleprompter.

1. The em dash

The single highest-confidence AI tell. Models reach for em dashes because they were trained on essay-grade prose where em dashes are legitimate punctuation. Spoken YouTube scripts almost never use them, because narrators don't pause that way.

If your script has any em dashes (—), the script is AI, full stop. Replace each with either a comma plus continuation, a period plus a new sentence, or restructure to remove the aside entirely.

2. "But here's the thing."

The most overused AI transition in 2026. Models use it as a rhythm break before a payoff sentence. The problem is that viewers have been hearing it across every AI script for two years now and it has become a deboost signal in their heads.

Same family: "here's the kicker," "what's fascinating is," "the truth is," "buckle up."

Strip all of them. Cut directly to the next sentence. If the next sentence needs a transition to make sense, the previous sentence wasn't doing its job.

3. The "X isn't Y. It's Z" pattern

Watch for this exact shape: short negative declarative, period, short positive declarative starting with "It's".

Example: "This isn't a tech story. It's a power story."

Once you notice it, you cannot un-notice it. Sonnet 4 leans on it constantly because it's a clean contrastive frame that compresses meaning. It also makes every script sound identical. Ban it in your style constraints and the script reads twice as varied immediately.

A simple regex catches it: /isn'?t .{0,40}\. it'?s/i.

4. Triple anaphora

Three or more sentences in a row starting with the same word. Models love this because it sounds rhetorical and confident. Viewers register it as "this person memorized a TED talk."

"They didn't see it coming. They didn't plan for it. They didn't survive it."

One repetition is fine. Two is occasionally powerful. Three or more is a tell. The fix is to break the rhythm by inverting the middle sentence: subject-verb to verb-subject, or splitting one of the three into two shorter sentences.

5. Staccato fragment chains

Three or more consecutive sentences under 6 words. Often shows up in the cold open because the model is trying to be punchy.

"The deal collapsed. The CEO vanished. The investors lost everything."

It feels punchy at first read. By the third one in a row it feels mechanical. Cap staccato at two in a row and the rest of the paragraph lifts.

6. The "In this video" opener

"In this video, we're going to look at..." or "Today, we'll explore..."

Both are dead. Viewers who clicked already know what the video is about, that's why they clicked. Open with the hook the title implied, not a meta-statement about the video itself.

7. Sub-N-word sentences in narration mode

If your channel runs in explanatory or documentary mode (think mid-pace voiceover, 30-50 audience), sentences under 8 words start to sound like a chatbot answer. The minimum-words-per-sentence floor varies by channel voice. For an explanatory documentary channel, 8 is a reasonable floor. For a fast first-person internet channel, 4 is fine.

This is the one tell that is style-dependent, not universal. The point is to set a floor that matches your channel's voice and enforce it with a linter, not to copy someone else's floor.

8. Generic chapter titles

If your script has chapters and the AI named one "The Reality," "The Aftermath," or "What Happens Next," rewrite it. Chapter titles should be a specific noun phrase that hints at the content, not a generic abstraction.

Bad: "The Beginning" Good: "The 1987 Reebok memo that started it"

The specific version improves CTR on the chapter thumbnail in YouTube's player and improves the searchability of the chapter in the description.

9. The 90-second drift

This is a structural tell, not a phrase tell. AI scripts often have a hook, then drift for 60-90 seconds before the next narrative pivot. Viewers leave during the drift.

The fix is a re-hook every 60 to 90 seconds. A re-hook is a single sentence that re-states the stakes in fresh language. "The number we're about to put on screen is the reason this company stopped existing." If the script doesn't have re-hooks, the script structure was never edited, only the prose.

10. Documentary voice in an internet voice channel

If your channel voice is first-person internet (18-30 audience, casual, "we" and "I" throughout), and the script starts saying "observers noted" or "industry analysts have suggested," the model defaulted to documentary voice mid-paragraph. This happens more often than people realize because models pattern-match to whatever genre cues are densest in the topic.

Pin the voice in a single line at the top of your system prompt and the drift goes away.

11. The overconfident close

"And that is the story of how..." or "And so we see that..."

These read as essay conclusions. Spoken video closes with a question, a callback, or a reframe, not a summary. The summary should already be in the viewer's head if the script did its job.

---

The deterministic check

Every one of the eleven patterns above is checkable with code. You don't need an AI judge for any of them. We run a regex pass on every script our pipeline produces, before the script ever reaches a human reviewer. The pass takes about 8 milliseconds per script and surfaces a list of inline highlights with the rule that fired.

If the count of flagged patterns is above a threshold, we automatically re-prompt the model with the specific tells highlighted and a fix instruction. The second pass is clean about 85% of the time.

This is the part that almost no AI script tool ships. They'll generate the draft, then either trust it blindly or hand it back to a separate "humanize" model that just shuffles the rhythm. Neither approach catches the specific phrases viewers have already been trained to filter out.

What good looks like

A clean script reads like the narrator wrote it after sketching it twice on paper. Sentences vary in length without obvious rhythm patterns. No em dashes. No essay transitions. One contrastive frame at most, used at a high-stakes moment. Voice locked to a single mode. Re-hooks every 60-90 seconds. Chapter titles that are specific noun phrases.

You can write one yourself with the AI doing the heavy lifting if you keep the eleven tells in front of you while editing. Or you can run a pipeline that enforces them before the script lands in your hands. Either works. What doesn't work is pasting the model output straight into a teleprompter.