Explainer

How to recognise badly generated AI images

5 things to look out for when you generate content with AI.

Most advice on this is out of date. Six fingers and melted text were the 2024 tells, and the models have largely engineered them out. We tried four times to generate a six-fingered hand for this piece and were corrected every time. These five still catch bad output, and each one is shown twice: generated badly, then properly.

Reflections that do not match

A mirror, a window, a puddle. A model paints what a reflection usually looks like instead of working out what this one should contain, so the copy drifts away from the original.

Wrong A woman in a red jumper holding a green mug, whose mirror reflection wears a blue jumper and holds nothing
Red jumper, green mug. In the mirror: blue jumper, empty hands.
Right The same scene where the reflection matches the woman exactly
Same jumper, same mug, correctly reversed.

What we do Anything reflective gets held against what is actually in the frame.

Skin with no texture

Generated skin tends to come back uniform: no pores, no fine hair, no colour variation, perfectly symmetrical. It reads as beautiful for half a second and as plastic after that.

Wrong Beauty close-up with completely smooth, poreless skin
Poreless and uniform. Nothing on this face varies.
Right Beauty close-up with visible pores, fine hair and natural asymmetry
Pores, peach fuzz, a mole, redness that is not symmetrical.

What we do We put texture back in on purpose, and reject faces that come back airbrushed.

A product that is nearly yours

This is the expensive one, and you can only see it if you have the real thing to compare against. So here it is first: a matte green bottle, brass cap, a terracotta band across the label and a small embossed mark below it. Now watch what happens when it is placed into a scene.

Studio packshot of a sage-green serum bottle with a brass cap and a terracotta band on the label
The product Matte glass, brass cap, terracotta band, embossed mark.
Wrong The bottle in a scene, but glossy with a silver cap and a blank label
Glossy instead of matte, silver instead of brass, no band, no mark.
Right The same scene with the real bottle reproduced correctly
The real one, attached as a reference. Every detail survives.

What we do Your product enters the work as an image, never as a sentence. One hero product per frame.

Small text, anywhere in the frame

Big headline text is mostly solved. Small and background text is not: menu boards, shelf labels, packaging small print, a book spine. From across the room it reads as writing. Lean in and it dissolves.

Wrong A café interior where the menu board and every shelf label are garbled
Convincing from here. Every label in frame is invented letterforms.
Enlarged Enlargement of the bottom right corner of the menu board showing nonsense words
The bottom right corner of that board: "VRYNA", "GRLANÖLA BLËND", and a last line that is not letters at all.

What we do Type that has to be read is set in code and laid over the picture, never generated.

A composition nobody chose

The quietest failure and the most common. With nothing steering it, a model does not just go plain, it goes awkward: subject dead centre and clipped, horizon off level, clutter competing behind, objects growing out of the subject, flat frontal light.

Wrong A bowl of pasta shot dead centre and clipped, with a cluttered background and a chair growing out of it
In focus, correctly exposed, and framed by nobody.
Right The same bowl of pasta art-directed with raking light and a considered palette
One light, a palette, negative space, a point of view.

What we do Art direction is decided before anything is generated, and held against your brand after.

All five, on everything we deliver

None of this is automatic. There is no setting for it: it is somebody opening the file, checking the five things above, and sending back what fails. That is most of the job, and it is the difference between content that looks made and content that looks generated.

Request a proposal →