Aims
All posts
Production8 min read

Why Your AI Ads Look Like AI (and How to Fix It)

The problem was never the model. It was what you handed it.

Henry Sedgwick

Henry SedgwickProduct marketing

Two printed ad posters on a white wall: the same bottle centred on a flat peach gradient, and art directed on a sunlit ledge with a long shadow

Everyone blames the model. You typed a careful prompt, the generator gave you a centred bottle floating on a peach gradient with a soft glow behind it, and it looked exactly like the last four things you made. So you tried a longer prompt and got the same ad in a different colour.

AI ads look like AI for one reason, and it is not model quality. A prompt is a handful of words, so the model fills every gap you left with the average of everything it has seen. Average lighting, average composition, average colour. The fix is not a better prompt, it is giving the model a real piece of human work to follow, so the gaps get filled with someone's actual decisions instead of a statistical middle.

01

The look is an averaging artefact

Unspecified detail defaults to the middle of the training data. That middle is the AI aesthetic.

02

Prompts underspecify by design

Thirty words cannot encode a layout, a light source, a crop and a type hierarchy. A picture can.

03

Readers penalise the look, not the method

Research points the same way: AI creative does fine until people can see it was AI.

04

Labels are now automatic

Meta and Google attach AI disclosures themselves. Hiding it is no longer a strategy.

05

A reference beats a longer prompt

Hand the model work a human already got right and the taste comes with it.

01

The tells that make an AI ad look like AI

When someone says an ad looks AI, they are usually reacting to five or six specific things at once. It helps to name them, because each one has a different cause and a different fix.

Nine printed ad cards laid out in a grid, each one a different product centred on the same soft peach gradient with the same glow behind it
Fig. 1 Nine different products, one ad. Nobody asked for the centred crop, the glow or the peach wash. That is the model filling in everything the prompt left blank.
01

Dead-centre product

No crop, no angle, no reason for the framing. Nobody art directs this way.

02

The gradient nobody chose

A soft warm wash with a glow behind the product, on every ad, in every category.

03

Unmotivated light

Shadows that imply two or three light sources, none of them in the frame.

04

Over-smoothed surfaces

Skin, fabric and packaging lose their fine texture and read as plastic.

05

Turned-up saturation

Colour pushed past where any brand would take it. This one is measurable.

06

Nothing to look at second

No hierarchy, no joke, no hook. A picture of a product, not an ad.

Six tells, six different causes. Each one is fixable on its own.
"I ran 60 variants for a candle launch and my designer picked them out of a lineup in about four seconds. She said they all had the same glow behind the jar.", Dan, home fragrance brand

02

Why the model defaults to that look

A prompt is not a specification. It is a few dozen words describing an image that needs millions of decisions, so the model resolves everything you did not mention toward what it saw most often. And what it saw most often is commercial stock and retouched studio photography, where the lighting is soft, the surfaces are clean and the product sits in the middle.

Unsaid means averaged

Every gap in the prompt is filled with the centre of the distribution. Averages have no taste.

The training set was already retouched

Models learned from imagery where texture had been smoothed away, so they treat smooth as good.

Everyone drew from the same well

Overlapping training data across the major models is why outputs from rival tools look like siblings.

03

What it costs when an ad reads as AI

This is the part that turns an aesthetic complaint into a media-buying one. The research that exists points in a consistent direction: AI-made ads can beat human-made ads on click-through, but only while viewers cannot tell they were AI-made. The moment the tells are visible, the advantage goes. The Harvard and Oxford working paper "AI in Disguise" found exactly that, and singled out heavy colour saturation as one of the cues doing the damage.

7%

Consumers who trust a brand more for visible AI content (Klaviyo, 2026)

31%

Who trust it less

39%

Who say heavy AI use reduces trust, up from 20% in 2025

37 pts

Gap between how positive execs think buyers feel and how they do (IAB)

04

How to stop your AI ads looking like AI

  1. 1

    Start from a real ad, not a blank prompt

    Pick one piece of creative that already works in your category. That is your specification.

  2. 2

    Name the light and the lens

    One source, one direction, one focal length. Unmotivated lighting is the tell you can kill in a sentence.

  3. 3

    Feed the model your actual product

    A product photo as a reference keeps the packaging, the proportions and the logo honest. A description will not.

  4. 4

    Pull the saturation and the smoothness back

    Grade every variant toward your real brand colours and leave texture in. Both are measured tells.

  5. 5

    Break the symmetry deliberately

    Off-centre crop, something entering frame, a hand in shot. Give the eye a second thing to do.

  6. 6

    Check it against the reference, not the prompt

    The question is never "did it follow my words". It is "does this hold up next to the ad I started from".

"Naming the light was the one that fixed it for me. One window, from the left, nothing else. That single line did more than the paragraph of adjectives I had before it.", Iwan, outdoor gear brand

05

What a reference carries that a prompt cannot

This is the whole argument, and it is worth being precise about it. A prompt encodes intent as adjectives. A reference encodes decisions as pixels. When a designer made that ad they chose a crop, a hierarchy, a moment of wit, a place for the eye to land. None of that survives being compressed into the word "premium".

Tracing paper over a printed ad, with the crop line, the light direction and the type hierarchy drawn on top in pencil
Fig. 2 What a reference actually hands over. None of these decisions were written down anywhere, and all of them are readable off the finished piece.

Two ways of asking for the same ad

Prompt only

Reference led

What you hand overThirty words of intentA finished ad that works, plus your product
Crop and framingAveraged to dead centreCarried from the reference
Light directionWhatever the model saw mostOne source, readable off the image
Type hierarchyRarely specified, rarely rightOrder and weight both survive
Product fidelityA plausible bottle, not yoursYour packaging, from your photo
The ideaNone. There was nowhere to put oneSomeone else's, applied to your brand
Fig. 3 Reference-conditioned generation is a standard capability in current image models, not a trick. The difference is what you put in the top of the funnel.
The first month I was writing prompts and getting the same shiny bottle back. Now I start from an ad I like, and my last three winners came out of that. Same model, I just stopped asking it to guess.
Marta, supplements brand, team of 4

06

Where AIMS fits

You now know what to feed the model. The awkward part is doing it: finding work worth referencing, then getting the model to actually read it rather than glance at it. That is the job AIMS does. It holds a searchable library of real ads, emails, videos and social posts from other brands, analyses the one you pick to work out what the designer actually did, and hands that analysis to the model along with your product, your palette and your claims. The output carries the structure and the wit, with your product in it correctly. Our AI static ads generator and AI product video ads pages walk through it.

  • Ads from Meta and TikTok, kept as references rather than screenshots
  • Competitor email newsletters, including the long campaign builds nobody can prompt from scratch
  • Video and UGC, where the reference carries pacing and shot order
  • Pinterest, Etsy and Amazon listings, for product photography and packaging cues
The hero section of a protein shot brand campaign email, used as the inspiration
Five bags of popped lentil chips photographed on a white background
The AIMS result: a snack brand campaign email hero in the brand palette
Fig. 4 Inspiration plus product, with AIMS in the middle. The badge row, the stacked headline and the button placement carry across. The palette, the copy and the product are the snack brand's own.

The point of the picture is not that the two look alike. It is that a person decided the badges should sit under the subhead and the product should be the only thing in focus, and that decision survived the trip to a completely different brand. No prompt I could write would have specified it, because I would not have thought to.

07

FAQs

Why do all AI images look the same?

Because they are all resolving toward the average of heavily overlapping training data. The major image models learned from similar pools of retouched commercial photography, so when a prompt leaves lighting, crop and colour unspecified, they each land in roughly the same place.

Why does AI-generated skin look plastic?

The training imagery was mostly already retouched, so the model learned that smooth equals high quality and strips the fine texture that makes skin read as real. Adding a reference photograph with real texture, and resisting the urge to upscale twice, is the practical fix.

Do AI ads perform worse than human-made ads?

Not automatically. The Harvard and Oxford "AI in Disguise" paper found AI-generated ads can beat human-made ones on click-through, but the advantage depends on viewers not being able to see that they were AI-generated. Visible tells, particularly heavy saturation, are what costs you.

Do I have to disclose that an ad was made with AI?

Increasingly it is not your call. Meta labels ads made with its own generative tools automatically and reads C2PA metadata from third-party tools. Google Ads introduced an AI disclosure requirement on 9 July 2026 with self-attestation for outside tools, and the EU AI Act's content marking obligation applies from 2 December 2026.

Will a longer prompt fix it?

It helps a little and then stops. Every extra clause specifies one more variable out of thousands, and the ones you never think to mention are exactly the ones a designer got right by instinct. A reference image specifies all of them at once.

Is using a competitor ad as a reference legal?

Taking structure, format and creative approach is normal practice and the platforms publish these ads deliberately. Taking their logo, photography, footage or copy is not. Keep the line at decisions, never assets.