Why AI ads look fake (and the 7 fixes that make them usable)

AI ads look fake for a specific, fixable set of reasons: waxy over-smoothed skin, a honeyed "glamour" color grade, garbled label text, physics that no object obeys, lips that don't match the audio, and a stock-photo perfection no phone has ever captured. None of these are inherent to AI video — they're defaults you have to actively prompt and pipeline your way out of.

We route Seedance 2.0, Veo 3.1, Kling 3.0, and Google's newest any-input video model every day at Hermoso, and the gap between an unusable render and one you'd actually put spend behind almost never comes down to the model. It comes down to seven craft decisions made before and after the model runs.

This post names the tells honestly, then walks through the seven fixes practitioners actually use — the same ones we've baked into our own pipeline.

What makes an AI ad look fake?

Seven tells account for almost every AI ad that gets clocked in the first second. If you've been burned by glossy, uncanny output, it was some combination of these:

• Waxy skin. Models are trained heavily on retouched, beauty-lit imagery, so they default to poreless, softly glowing faces that no phone camera produces.
• The golden glamour grade. Warm, honeyed, cinematic lighting on everything — a kombucha ad lit like a perfume commercial.
• Garbled label text. The product name melts into near-letters. Captions render as alphabet soup. This is the fastest single giveaway.
• Impossible physics. Products float a beat too long, liquid pours defy gravity, a tossed bottle hangs weightless. Mass and momentum are still hard for these models.
• Dubbed lips. TTS audio laid over a generated mouth produces the uncanny lip-flap everyone recognizes from badly dubbed movies.
• Model-painted captions. Asking the model to render the caption bar or CTA button inside the frame — it can't, and the smeared result poisons an otherwise good shot.
• Stock-photo perfection. Perfectly centered framing, perfectly still hands, nobody blinks, nothing is slightly wrong. Real content is a little wrong everywhere.

Each of these has a specific countermeasure. Here are the seven, in the order of how much they matter.

Fix 1: Prompt for matte skin and flat, phone-camera capture

The anti-gloss fix is to explicitly prompt the model away from its beauty-lit default: matte skin with visible texture, neutral lighting, and the capture character of a phone camera shooting a flat SDR picture profile rather than a graded cinema rig. Left unprompted, every major video model drifts toward glow.

In practice this means writing the realism layer into every UGC-style prompt: unretouched skin, no rim light, no bloom, handheld framing, the slight noise and dynamic-range limits of a front camera. The output stops looking like a commercial and starts looking like something a person filmed — which is the entire point of the UGC format. This is the fix with the highest effort-to-impact ratio, because skin is where viewers' fake-detection is sharpest.

Fix 2: Composite all text, logos, and captions in post — never let the model paint them

Any text a viewer needs to read — logo, label, caption, CTA, end card — should be composited over the finished footage, not generated inside it. Diffusion models produce pixels that resemble text statistically; they don't write glyphs. The result is melted letters and invented near-words, and a garbled label undoes an otherwise flawless render.

The working pipeline is: render the scene clean, then overlay the real brand assets in a compositing pass — actual logo file, actual typeface, pixel-perfect captions. This is also how you get an end card with a crisp logo and a readable CTA. It's more pipeline work than typing "add the logo" into a prompt, which is exactly why most fake-looking AI ads skip it.

Fix 3: Ground every render on a real product photo

If the model doesn't have your actual product photo as a reference, it will invent a plausible-looking cousin of your product — right category, wrong bottle shape, hallucinated label, slightly-off color. Customers who know the product notice instantly, and so do you.

The fix is to feed a real product photo into every generation and composite the genuine article into the scene rather than describing it in words. Text prompts describe categories; reference images pin identity. Combined with Fix 2 (composited label text), this is what makes the product in the ad the product on the shelf — and it's the difference between a lifestyle image and a finished ad you can actually run.

Fix 4: Write micro-gestures and one imperfect beat into every performance

Fake-looking AI performances are usually under-directed, not under-rendered. A prompt like "woman holds product and smiles" gets you a mannequin: locked eyes, frozen grip, symmetrical smile. Real people fidget.

Direct the performance at the gesture level: she glances down at the label mid-sentence, readjusts her grip, tucks hair behind an ear, half-laughs at her own phrasing. Then include exactly one imperfect beat — the camera bumps, she talks over herself for a syllable, the framing drifts. One flaw reads as authentic; five read as chaos. This is scriptwriting, not model tuning, and it's the fix most teams never think to make because they're busy blaming the model.

Fix 5: Native model speech for visible mouths, TTS only for voiceover cuts

The rule is simple: if the mouth is on camera, the speech must be generated natively by the video model in the same pass, so the lips and the audio come from one process. If the shot is product B-roll with narration over it, TTS is fine — there's no mouth to betray it.

Where teams go wrong is generating silent footage of a person talking, then dubbing TTS on top. The mismatch between generic mouth movement and specific phonemes is something human perception is ruthlessly good at catching — it's the dubbed-movie effect, and no amount of grading hides it. Current models generate genuinely lip-synced speech when you let them; the fake look comes from bolting audio on afterward.

Fix 6: Keep heavy objects out of ballistic flight

Don't ask the model to toss, drop, flip, or catch anything with mass. Ballistic motion — an object in free flight under gravity — is where video models' physics still visibly break: the bottle hangs a beat too long, tumbles at a constant rate, lands without weight.

Choreograph around it. Products get held, turned, set down, picked up, poured into — motions where a hand carries the object and the model never has to solve gravity on its own. Pours and splashes are riskier than holds but far safer than throws. If the concept truly needs the product airborne, that's a shot to cut around, not a shot to generate. Most "impossible physics" fails we see are self-inflicted: the storyboard asked for a hero toss the model was never going to land.

Fix 7: Cool, neutral lighting over honeyed grades

Prompt for neutral or slightly cool white balance and even, unglamorous light — overcast daylight, a kitchen at noon, office fluorescents — instead of accepting the warm golden grade models reach for by default. That honeyed look is the visual signature of "AI ad" in 2026, and viewers have learned it.

Neutral light does double duty: it kills the glamour tell, and it makes Fixes 1 and 3 land harder, because matte skin and a real composited product both read most convincingly under ordinary light. The goal for performance creative isn't beautiful footage — it's credible footage. Save the golden hour for shots where the brand genuinely calls for it, chosen deliberately rather than inherited from the model's defaults.

How Hermoso builds these fixes into its pipeline

The heaviest-lift fixes on this list are built directly into Hermoso's pipeline, so you get the output of a team that does this daily without doing the pipeline work yourself:

• An anti-gloss realism layer writes matte skin, neutral lighting, and phone-camera capture character into UGC renders (Fixes 1 and 7).
• All text, logos, and captions are composited in post — never model-painted — including an end card with your real logo and CTA (Fix 2).
• Renders are grounded on your real product photo, composited into the scene, with your brand palette applied (Fix 3).
• Native model speech is used whenever a mouth is on camera, with TTS reserved for voiceover-only cuts (Fix 5).

Underneath, it routes each job across Seedance 2.0, Veo 3.1, Kling 3.0, and Google's newest any-input video model — plus Flux, GPT Image, and Google's latest imaging models for stills — picking per job so you never choose. And because it starts by researching what's already winning in your market across the Meta, Google, and LinkedIn ad libraries plus organic TikTok, the concepts are grounded in what's already winning before the pixels are rendered. The free plan includes 250+ earnable free credits with no card required; image ads cost a few credits and video starts around 25. Try it on your own product — the fastest way to test this post's argument is to see whether the output passes your own fake-detector.

Frequently asked

Why does AI-generated text and logos come out garbled?

Video and image models generate pixels that statistically resemble text rather than actual glyphs, so labels and captions come out with melted letters, extra strokes, and near-words. The reliable fix is to render the scene without text and composite the real logo, captions, and CTA over the footage in post.

Can AI video do convincing lip sync?

Yes, but only when speech is generated natively by the video model in the same pass as the footage. Dubbing separately generated TTS audio over a rendered mouth is where the uncanny lip-flap effect comes from — reserve TTS for cuts where no mouth is visible.

Which AI video model is best for ads?

There is no single best — Seedance 2.0, Veo 3.1, and Kling 3.0 each have different strengths depending on the shot, and the ranking shifts with every model release. Hermoso routes each job to the model best suited for it, so users never have to pick.

Can I put my real product in an AI-generated ad?

Yes, and you should — grounding every render on a real product photo is one of the highest-impact fixes on this list. Hermoso composites your actual product photo into generated scenes instead of letting the model hallucinate a lookalike.

How much cheaper is AI ad creative than a shoot?

Agencies commonly quote four to five figures for a single produced ad, and creator marketplaces commonly list UGC videos in the low hundreds each. On Hermoso, an image ad costs a few credits and a video starts around 25 credits, with paid plans at $19, $49, and $149 per month.

Will viewers reject an ad just because it's AI-generated?

In our experience the tells are what get rejected, not the origin. Footage with matte skin, neutral light, real label text, and one imperfect human beat reads as ordinary content; footage with waxy skin and golden gloss reads as fake regardless of how it was made.