The obviously-AI tell: pacing, imperfection, proof

Module 2 §1 covered the tell in text — stacked adjectives, the tricolon, not just X but Y, question openers, symmetry nobody speaks in. That list is there; this section doesn't repeat it.

This is about the tell in a finished piece, which is a different problem. A script can pass every vocabulary check and the video can still announce itself. Three places it shows: pacing, imperfection, and proof.

1The general shape of it

The tell is almost never bad. It's even.

Bad is easy to spot and easy to fix. Even is what a machine produces by default, because evenness is what you get when nothing in the process had a preference. Every sentence weighted the same. Every cut the same length. Every beat given its fair share of time.

Real work is lopsided. Somebody cared more about one part.

2Pacing

Automated editing removes dead air, and dead air is often the point.

Module 2 §4 covers this mechanically — filler-and-silence removal will strip the pauses you meant along with the ones you didn't, and the undo pass exists to catch it. Here's the voice-level version: pacing is part of a brand's register, and it belongs in BV-01.

A brand whose party-behaviour answer is "answers the question you asked, then stops" has a pacing implication. It should have gaps. It should not fill silence. If the finished cut has no air in it, it doesn't sound like that brand, however correct the words are.

Things worth recording in BV-1 that people don't:

• Do they leave silence? Some brands' whole register is comfort with a pause.
• Do they finish sentences? Real speech trails off. Machine speech completes.
• How fast do they get to the point? A brand that takes its time and a brand that leads with the answer are different brands, and this survives translation into video better than any adjective.

The practical test: watch your cut with the sound off, then listen with your eyes shut. If the rhythm is identical throughout — same beat length, same emphasis, no moment that holds — it will read as machine-made even if every word is yours.

3Imperfection

The strongest single move available, and the one people are most reluctant to make: leave something in.

A stumble. A repair — the — the thing. A breath in an awkward place. A take where you said it slightly worse but meant it more.

The instinct, especially when a client is paying, is to deliver the clean one. But cleanliness is now free and therefore worthless as a signal. What is expensive, and therefore legible, is evidence that a person was there.

Two cautions, because this is easy to do badly:

Performed imperfection is worse than none. A deliberate stumble reads as a technique within about two seconds. The rule is keep rather than add — you are choosing not to remove something that happened, not manufacturing something that didn't.

Check it against BV-1 first. Not every brand can carry it. A brand whose never-sound-like list includes anything sloppy may have a real reason. Ask.

The clearest example in this guide's own material is a script note: hold two seconds on the empty pan, no music. Nothing generated that. Someone watched a rough cut and noticed the line landed harder when nothing followed it. That decision is what the piece is actually selling.

4Proof

The third tell isn't a quality of the piece. It's whether anything backs it.

Audiences start more sceptical of AI-assisted content than of anything else you will make. That scepticism is not defeated by the piece being good — a very polished piece raises it. It's defeated by evidence.

Concretely, in a UGC piece:

• The specific detail that only a user would have. The eggs came off clean, first try beats any adjective, and it cannot be produced by someone who didn't cook the eggs.
• The limit. One honest reservation, where the brand permits it, does more for credibility than three claims. Module 2's testimonial template covers when this is available to you and when a client's barred list takes it off the table.
• The provenance, where you're showing generated output. What made it, when, from what, and what a person changed afterwards.

Module 4 is entirely about this and this section defers to it. What belongs here is the connection: proof is a voice decision before it's an ethics decision. A brand that never concedes anything sounds like an advert, and BV-1 is where you record whether this one can.

5Where the tell actually gets fixed

Upstream. Almost none of this is repairable in the edit.

If the brief didn't name the off-limits words, the script has them and no amount of cutting removes them. If BV-1 didn't record that the brand leaves silence, the performance won't. If nobody wrote down what actually happened with the product, there is no specific detail available to include, and you will reach for an adjective instead.

That is the argument for this whole module. The tell is a symptom of a thin intake, and by the time you can hear it you are three steps downstream of the fix.

6A short checklist

Before a piece ships:

1. Read the script aloud. Where do you stumble? That's where a viewer will.
2. Is there one moment that holds longer than the others? If not, find one.
3. Is there one thing in it that could only have come from someone who used the product?
4. Is there one honest limit — and does BV-3 allow it?
5. Watch it silent, then listen blind. Same rhythm throughout, or does it move?

If you answer no to 3, stop. The others are craft. That one is whether the piece has anything in it.