The short answer
You can tell a script is AI-generated by reading the opening 30 seconds out loud and watching for a cluster of tells, not a single one. Any one sign can show up in human writing too. What gives AI away is several appearing together: a throat-clearing intro, sentences that are all roughly the same length, hedging language, list-of-three constructions everywhere, and zero specific detail that only a real person would know.
I review AI-drafted scripts most working days, and the fastest test I have found is the 'so what' read. Go line by line and ask whether the sentence tells the viewer something a human who lived the experience would say. AI fills space confidently without ever saying anything only the author could know. That hollow confidence is the real tell - the surface signs below are just how it shows up.
The 8 tells, with a fix for each
Here are the eight signs I check for, why each one happens, and the one-pass edit that removes it. Run down the table first, then read the notes underneath for the ones that trip people up.
| Tell | Why it happens | Fix |
|---|---|---|
| Filler transitions ('in today's video', 'let's dive in', 'without further ado') | Models learn these phrases from millions of generic videos and default to them when no hook is specified | Cut the intro entirely. Open on the most surprising line in the script |
| Uniform sentence rhythm (every sentence 12 to 18 words) | Models regress to the mean sentence length they were trained on | Break one in three sentences into a two- or three-word fragment. Read aloud to feel the rhythm |
| Hedging ('arguably', 'it's worth noting', 'in many cases', 'often') | Safety tuning rewards qualified, non-committal statements | Delete the hedge and commit to the claim, or add the specific case that earns the qualifier |
| Lists of three everywhere ('faster, cheaper, and easier') | The rule-of-three is statistically overrepresented in training text | Replace some triples with a single sharp claim plus a concrete example |
| No first-hand detail (no numbers, names, dates, or 'I tried') | The model has no lived experience to draw on | Add one real number, one real example, or one thing you personally saw happen |
| Symmetrical structure (every section the same shape and length) | Models optimize for tidy, balanced output | Let one section run long and another be a single line. Mirror how people actually talk |
| Empty signposting ('Now, let's talk about', 'But here's the thing') | Filler connectors pad the response to feel complete | Replace with the actual point. If the signpost can be deleted with no loss, delete it |
| Polished-but-generic tone (sounds fine, says nothing) | Fluency without grounding produces confident filler | Run the 'so what' read. Cut any line a stranger could have written about any product |
The fastest 60-second check
When I am triaging a stack of scripts, I do not read all eight tells in order. I run three fast passes that catch the majority of AI drafts.
First, read the opening three sentences aloud. If the script clears its throat before it says anything - 'In today's video we're going to be talking about' - that is the single most common AI giveaway, and it shows up in the first five seconds.
Second, scan the left margin. Human writing starts sentences in wildly different ways: a name, a number, a one-word fragment, a question. AI writing tends to start sentences with the same handful of openers ('This', 'It', 'When', 'By', 'Whether'). A column of near-identical sentence starts is a strong tell.
Third, hunt for one real detail. Search the whole script for a number, a proper noun, a date, or the word 'I'. If there is not a single concrete, checkable detail in 200 words, you are almost certainly looking at unedited AI output.
- Read the first three sentences out loud - listen for throat-clearing.
- Scan sentence starts - look for repeated openers and uniform length.
- Search for one checkable detail - a number, name, date, or first-hand 'I'.
Why these tells exist (and why they are fixable)
None of these tells mean AI writing is bad. They mean it is unedited. A language model is trained to produce the most probable next word, averaged across enormous amounts of generic internet text. The average of all writing is smooth, balanced, hedged, and free of specific lived detail - because specific lived detail is exactly the part that differs from person to person and cannot be averaged.
That is why the fixes are mostly about adding back what averaging removed: a real number, an uneven rhythm, a committed claim, a first-hand moment. The model gives you a competent scaffold. Your job in editing is to put the fingerprints back on it.
It also explains why detection tools are unreliable. They score statistical smoothness, but a human who writes in clean, balanced prose can score as 'AI', and a lightly edited AI draft can score as 'human'. The tells above are more reliable than any single detector because you are reading for meaning, not measuring perplexity.
How tooling changes the picture
Some of these tells are baked in by the writing tool's defaults, not the model itself. A generic chatbot prompt produces generic chatbot prose. A script tool that constrains the output toward how people actually speak removes several tells before you start editing.
ContentIQ, for example, generates scripts at a 2nd-to-3rd-grade reading level with a hard cap of about 12 words per sentence and a visual prompt attached to each line. Those constraints alone knock out the two most common tells - long uniform sentences and empty signposting - because a 12-word ceiling forces short, varied, spoken-sounding lines and the per-line visual prompt forces every line to carry concrete content rather than connective filler. You still edit for first-hand detail, but you start several steps ahead of a raw chatbot draft.
The takeaway: a script that sounds like AI is usually a prompting and editing problem, not a fundamental limit. Spot the tells, run the fixes, and add the one thing the model cannot - your own specific experience.