Scripting with AI Without Sounding AI

A script the AI wrote and you never touched will underperform. That's not a moral position, it's just what the numbers do, and four separate people teaching this independently landed on the same instruction: proofread everything, then put your own creativity back into it until it sounds like you and not like a model.

The question is what "sounds like you" means at the sentence level, because you can't fix a script with vibes. So here's the formula that actually describes it. High perplexity, high burstiness, high complexity. Content that scores high on all three reads as human. Content that misses on any of them reads as generated, to a detector and to a viewer.

Perplexity is how surprising your next word is. A prediction engine reads your sentence and guesses what comes next, and the more often it guesses right, the more machine-written the text looks. So you feed it things it can't predict. Specific dates. Specific events. Your own opinions, stated as opinions, because a model has no reason to expect what you personally think about anything.

The strongest version of this is sensory detail. Take any generalization in the draft and swap it for a concrete physical thing you'd only know if you'd been there. "His apartment was a mess" is exactly what the predictor expects. Three empty Red Bull cans next to a dead plant is not. That one substitution does more than a paragraph of rewriting.

Same idea with word choice, and this is where people overcorrect. You're not reaching for fancier words. You're reaching for less obvious ones. The predictable word and the impressive word are usually both wrong. Pick the one that's slightly off the expected path and still completely normal to say out loud.

Burstiness is rhythm. AI writes sentences that are all roughly the same length, one after another, in that even article cadence, and it's the single loudest tell in a narrated script because you can actually hear it. So vary it wildly. A long sentence that winds through a couple of clauses and takes its time getting where it's going. Then four words. Then something medium. Uneven on purpose.

Complexity is the last one, and it's about vocabulary density. Models reach for the generic version of every noun. Your job is to replace those with specific, concrete, slightly random detail wherever you can. Not "a car." The make. Not "a large amount." The number.

Part of that is talking like someone who's actually in the niche. Every field has terminology real practitioners use with each other, and models tend to explain around it in general-audience language instead. Use the insider word. It signals to a viewer that whoever wrote this has been inside the thing, and it's a word combination no generic prompt would have produced.

One thing you can do upstream of all this: train the tool before you ask it for anything. Give ChatGPT standing personality instructions about how it should write for you, and the drafts arrive less templated. You'll still rewrite them. You'll just rewrite less.

Never ship a draft you haven't argued with. Read it, cut what's generic, put a real detail and a real opinion into every section, and let the rhythm go uneven.

The AI writes it. You make it yours.