How to Make AI Ads That Don't Look Like AI

By Atharv Sankpal

This is the question every brand asks before they run their first AI ad, and almost nobody answers it straight: will people be able to tell, and does it matter? The short version is that "looks AI" has become shorthand for "low effort", and low effort costs you trust before it costs you a conversion. The fix is not a better model. It is a set of specific choices.

Why "looks AI" costs money

A viewer who suspects the person on screen is fabricated stops evaluating the product and starts evaluating you. That is fatal on a cold ad, because you had about two seconds of goodwill and you just spent it. The damage is not caused by AI being used. It is caused by an ad that almost passes and does not.

The tells people actually notice

Lip-sync that drifts, especially at the end of longer sentences.
No micro-expressions: nobody blinks wrong, glances away, or reacts to their own words.
Skin and lighting that are too perfect. Real phone footage has noise, uneven light and a bad white balance.
Flat delivery, every sentence at the same pitch and pace, with no emphasis on the word that matters.
Robot script: complete sentences, no contractions, tidy transitions, marketing adjectives.
Continuity breaks: hands, jewellery, background objects and labels changing between cuts.
A room that does not exist. Perfect symmetry, no clutter, no personality.

The fixes

Write it spoken, not written

Read your script aloud. If you would not say it to a friend, rewrite it. Contractions, a false start, one incomplete thought, a specific number instead of a rounded one. Specificity is the strongest authenticity signal there is: "it took eleven days" reads as true in a way that "results in under two weeks" never will.

Cast believable, not beautiful

Overly polished avatars with studio lighting are the fastest way to trip the detector. Ordinary-looking faces in ordinary rooms outperform them consistently. Test several faces against one script before you decide the script is the problem.

Match the platform's native feel

TikTok wants raw, handheld, slightly badly framed. Instagram tolerates elevated but still real. Facebook is somewhere in between. An ad that looks like the content around it is judged as content; an ad that looks produced is judged as an ad.

Keep clips short and cut often

Sync and expression errors accumulate with duration. Short takes with real cuts hide the weaknesses and match how people film themselves anyway.

Add human artefacts

Room tone, an imperfect crop, a caption typed the way a person types, a hand entering the frame. Small mess signals a real moment. Perfection signals a render.

The reframe: sometimes stop hiding it

The counterintuitive move is to make the ad obviously not real. Pixar-style and claymation ads perform precisely because no one expects a cartoon to be a person. The uncanny valley only exists on the road to photoreal. Choosing an honestly stylised format skips the entire problem, and in a feed of near-real AI it stands out more, not less.

The compliance floor

Never stage a testimonial with a person who does not exist, and never imply an endorsement you do not have. Beyond being a regulatory problem in most major markets, it is the single fastest way to turn "this ad looks AI" into "this brand lies". Use synthetic presenters for hooks, demos and explainers, and keep proof real. More on avatar ads and the line.

Volume from AI, direction from humans

The teams getting real results all work the same way: generate widely, then apply a human review step before anything publishes. Models are good at producing options and bad at emotional judgement, which is exactly the judgement an ad needs. AI decides who can afford to play. Craft still decides who wins.

FAQ

Why do AI ads look fake?

Usually a handful of tells: lip-sync that drifts, missing micro-expressions, skin and lighting that are too perfect, flat vocal delivery, and scripts written in tidy sentences no one speaks aloud. Each one is fixable, and most of them come from casting and script choices, not the model.

Do AI ads still convert if people can tell?

Yes, when the format is honestly stylised. Animated and claymation ads perform well precisely because no one expects them to be real. The conversion damage comes from realistic ads that almost pass, because near-real reads as deceptive.

The honest part

This whole post is my pitch, stated as a guide. The reason the ads I make do not read as AI is not a secret model. It is that a person writes them, casts them, cuts them and throws away the nine takes that felt slightly wrong.

See the AI ad maker service, or browse the formats I make.

Want the ads made for you?

I make AI video ads for ecommerce brands end to end: the angle, the hook, the script, the actor and the edit. Not a tool to learn, a person who ships creative you can run.