How to Make AI Video Not Look AI Generated: The 3 Checks That Matter

You have already tried the fixes. Softer lighting, a little grain, less saturation, a slower camera move. The clip looks better than it did, and it still lands wrong.

There is a reason the craft fixes have a ceiling, and one experiment shows it cleanly. Researchers at the Nuremberg Institute for Market Decisions split participants into two groups and showed both the same six advertisements. Nothing about the images differed between the groups. One group was told the ads were AI-generated, the other that they were human-generated. The group who believed they were looking at AI rated the ads as less natural and less useful, took a more negative view of them, and was less willing to research or buy the product (NIM, Transparency Without Trust, Experiment II, German participants only, retrieved 2026-08-07).

Identical images, different response. Part of what you are fighting is not in the pixels at all. It is whether the viewer believes what they are looking at.

Key Takeaways

• Polishing texture has a ceiling. Shown the same ads, the group told they were AI-generated rated them less natural and less useful and was less willing to buy (NIM experiment, German participants, 2025).
• The tells that cost you a sale are usually truth failures, not texture failures: the product is subtly wrong, the presenter's face changes between cuts, the setting is a place you do not have.
• Your own eye is the wrong instrument. Advertisers overestimate how positively buyers receive AI-generated ads by 37 percentage points (IAB, 2026).
• Run three checks before posting, on a paused frame, on a phone: product, person, place. Reject the scene, not the video.

Why your AI video still reads as AI

The tells split into two groups, and almost every guide covers only one.

Craft tells are the surface: floaty motion, over-smoothed skin, light that does not match between shots, physics that does not obey mass or gravity. These are real, they are worth fixing, and the standard advice works on them. Specify a lens and a shutter. Pull the saturation down. Add grain. Keep the camera move short and motivated. Fix these and your clip stops looking like a render.

Truth tells are underneath, and they are the ones that lose the sale. The label on the jar is not quite your label. The presenter's jaw is narrower in scene three than in scene one. The shop in the background is a shop you do not own. These are not aesthetic problems. A viewer cannot always name what is wrong, but they register that the video is making a claim it cannot keep, and the feeling that follows is the one the NIM experiment measured.

The reason this is hard to catch is that you are the worst-placed person to catch it. When IAB surveyed 505 US Gen Z and Millennial consumers alongside 104 advertising executives, executives estimated that 82% of those consumers felt positive about AI-generated ads. Among the consumers themselves, 45% did. The gap has widened from 32 points to 37 since 2024, which means the people making this content are getting less accurate about how it lands, not more.

So do not trust the impression you get watching your own video at full speed on a laptop. Check it deliberately instead.

The three checks to run before you post

Every check runs the same way: pause the frame, view it at full screen on a phone at full brightness, and compare it against something real. The whole pass takes about four minutes on a six-scene video.

1. Is that actually your product?

Pause on every frame where the product is the largest thing on screen. On a short promo that is usually two to four frames. Hold the real item next to the phone.

Read the label, do not glance at it. Generative models reconstruct text as texture rather than as writing, so packaging tends to look right at a glance and fall apart when you stop. Check the brand name letter by letter, then the weight or volume, then the smaller line underneath. If it is not legibly yours, the scene is not usable, however good the lighting is.

Count things. Buttons, stitching rows, straps, lid threads, the holes in a shoe. Models are reliable about the impression of an object and unreliable about its inventory. A jar with the wrong number of ridges reads as a different jar.

Check the finish. Matte drifts to glossy, brushed metal drifts to chrome, a knit drifts to something plastic. Material is what a repeat customer knows by hand, and it is the tell they notice fastest, because they are not looking at your ad the way a stranger does.

2. Is that the same person in every cut?

Drift happens between shots, not within them. A single continuous clip usually holds together. The break comes at the cut, where the model derives the face again from scratch.

Pull the first frame of every scene and view them side by side. A five-scene video gives you five stills. Look at them as a group, not in sequence. Watching in sequence hides drift, because each step is small enough to accept and the accumulated change is not.

What moves first: jaw width, nose shape, the spacing between the eyes, the hairline, and the tooth line when the person speaks. Skin tone shifts under changed lighting in a way real skin does not. Then check continuity that has nothing to do with the face: a necklace that appears in scene two, a mole that swaps sides, sleeves that change length.

Check the hands doing the holding, not just the hands. Everyone counts fingers now. Fewer people check that the grip is the same across cuts, or that the product is the same size relative to the palm in every shot. Inconsistent scale is what makes a demonstration feel staged.

3. Is that a place you actually have?

A generated backdrop is fine when it reads as generic. It becomes a problem the moment it reads as your premises.

The specific risk is a scene that a viewer could take as a picture of your business: a storefront, a service counter, stocked shelves, staff in uniform, a queue at the door. That is not set dressing, it is a statement about what you have. Watch also for invented signage, price boards with numbers you do not charge, and awards or certifications on a wall.

The test is simple. If a customer could plausibly walk in expecting that room, either it is your room or the scene needs reframing. Reframing usually means a tighter crop, a plain surface, or a neutral background that carries no claim.

Two physical tells give away a generated space quickly: shadows falling in more than one direction with a single visible light source, and reflections in glass or a screen that do not contain the person standing in front of them.

Reject the scene, not the video

The instinct that ships a bad clip is wanting to keep five good scenes and forgive the sixth. Do not. Regenerating one scene costs almost nothing, and the one you forgave is the one your buyer will pause on.

Build the check in instead of bolting it on

Everything above is a manual pass over output that already exists, which means you find the drift after paying for it. The better place for the check is before the scene is ever shown to you. That is the idea behind Ugcly: it holds the product, the presenter and the place to the photos you actually upload, which it calls the Triple Lock, then checks every generated scene against those locks and regenerates what drifts before you see it. Its Realism Engine tunes for social-native footage from the start rather than correcting toward it afterwards. That narrows the tells; it does not claim to remove them, and the three checks above are still worth running on anything you are about to spend money promoting.

Frequently asked questions

Why does my AI video feel floaty or uncanny?

Usually because the motion is not obeying weight. Video models learn what movement looks like without modelling mass, gravity or momentum, so objects settle too slowly, fabric falls without resistance, and hands pass through surfaces they should meet. Human perception is unusually sensitive to physics that is slightly wrong, which is why the clip feels off before you can say what is off. Shorter shots and slower, motivated camera moves reduce how much the model has to invent.

Do I have to tell people the video was made with AI?

Often yes, and the rules differ by platform and by region, so check the current policy of wherever you are posting rather than relying on a summary. One practical note that removes the usual objection: YouTube states that disclosing altered or synthetic content "won't limit a video's audience or impact its eligibility to earn money" (YouTube Help, retrieved 2026-08-07). Non-disclosure carries more risk than disclosure does.

Is AI video good enough to run as an ad?

For a lot of scenes, yes. The limiting factor is rarely polish now. It is whether the clip is accurate about the thing you are selling, because that is what a buyer is evaluating and what advertising rules actually govern. A technically rougher clip showing your real product will usually outperform a beautiful one showing something adjacent to it.

The short version

Craft fixes are worth doing and they are not the hard part. Grain, lens choice and slower camera moves will get you past the obvious render look, and every guide on this topic will tell you how. What they will not tell you is that the tells costing you sales are the ones where the video quietly stops being true: a label that is not yours, a face that changes between cuts, a shop you do not have.

Those are catchable. Pause the frame, check product, person and place against something real, and throw out the scene rather than the video. Four minutes, before you post.