AI Images: Where They Help, Where They Embarrass You

Generative image tools crossed a line in the last couple of years. They stopped being a novelty and became a production tool — fast enough, cheap enough and good enough that most marketing teams are already using them, usually without a single written rule about when not to. That is the actual risk. Not that AI images look fake, but that they look real enough to make a promise your business cannot keep. This is an honest map of where generated imagery earns its place in commercial work, and where it will embarrass you.

AI images are a production tool, not a shortcut past reality

Treat generative imagery the way you treat every other part of the machine — a component with a specific job, not a replacement for a whole department. It is very good at making things that never needed to be true: a gradient, a texture, a mood, a shape sitting behind your headline. It is very bad at standing in for something a customer will later see with their own eyes. Almost every AI image disaster in marketing comes from confusing those two categories, and almost none of them come from the tool being weak.

Where it genuinely earns its place

There is real, unglamorous value here, and it clusters in five places:

• Backgrounds, textures and abstract surfaces — the layer behind the message, where nothing is being claimed.
• Concept and mood direction — showing a client the feeling of a campaign before anyone books a crew, a location or a day of light.
• Extending or reframing a shot you already own — turning a wide landscape frame into a vertical story without recomposing the whole thing.
• Social graphics and typographic posts — where the image is decoration around words that carry the actual meaning.
• Cleaning up a real photograph that is usable but flawed — a dust spot, an ugly sign, a cable across the sky, a crop that came out slightly too tight.

Notice what those have in common. In every case the customer is not being shown a thing they will later inspect. The image supports a claim; it is not the claim. That single test is more useful than any list of rules about prompts, and it holds as the tools keep improving.

The moment a customer can compare it to reality, you have a problem

The failure mode is specific. A generated image of a real property, a real dish, a real hotel room, a real product — anything the buyer will eventually stand in front of — sets an expectation that reality then has to meet. Sometimes it cannot. The balcony is narrower. The pool is smaller. The room faces the car park. The customer does not think interesting use of AI; they think they were misled, and they are not entirely wrong.

This is why the Matarma Beach Residence gallery runs to a hundred and twenty-nine real photographs. Generated interiors would have filled that grid faster and cheaper. But somebody books a room on the strength of those images and then walks into it, and the gap between the two is the only thing that matters. A beautiful generated bedroom would have become a liability the day the first guest arrived.

Faces, hands and text are still where the seams show

Two things break in generated images more often than anything else, and both are things people look at closely. Human faces at close crop go wrong in ways viewers cannot articulate but always feel — eyes that do not quite focus together, teeth that repeat, skin with no pores. And text inside the image is still unreliable: a sign, a label, a price, a logo, an Arabic word rendered by a model that guessed at the letterforms. In a bilingual market that second problem matters more than most teams expect. Arabic typography does not survive being hallucinated, and your audience will spot it instantly.

Both have the same fix. If a face must be close and human, shoot it or licence it. If the image needs words, set real type in the layout on top of the image — never inside the generation. Generate the parts that carry no information, and control the parts that do.

In real estate and hospitality this is not a style choice

For most categories a bad AI image is an aesthetic embarrassment. In property and hospitality it is closer to a claim about something that does not exist yet, made to a person who is about to sign a contract. A render is normal in real estate — buyers have understood renders for decades, and developers rely on them. A photorealistic generated image presented as a photograph of a finished space is a different thing entirely, and the whole difference is whether the buyer knows what they are looking at.

So label renders as renders, keep photography for anything finished, and never let the two blur inside the same gallery. When we build a project site — the same discipline that runs through our Websites and Automation work — the image library gets sorted by that question before a single page goes live: is this a photograph, a render or a mood image, and does the page make that obvious to somebody skimming on a phone?

Nobody endorsed you, and generated proof is worse than no proof

The fastest route from a clever image to a legal letter is generating something that implies a real person, a real customer or a real brand stood behind you. A face that resembles a public figure. A testimonial portrait of a client who does not exist. A partner logo dropped into a scene because it filled the space. None of these are edge cases now; they are the most common way a small brand talks itself into trouble with generative tools. The AI made it has never worked as a defence, and it will not start.

Put the rule in the workflow, not in somebody's head

Good intentions do not survive a deadline. The only version of this that holds is a rule that lives inside the process rather than inside a person: every generated image is reviewed against reality before it publishes, and anything depicting a real place, product or person needs a real source. We use AI to draft, never to publish unread — and the review step is where the value actually sits. It is the same principle that runs through our Content Creation system. The tool accelerates the making; a human still owns the claim.

Three questions handle almost every case:

• Can the customer compare this image to a real thing they will see? If yes, use a photograph.
• Does it contain text, a logo, or a face at close crop? If yes, control that element in the layout, not in the prompt.
• Would you be comfortable if the customer knew exactly how this image was made? If not, you already have your answer.

An AI image is only dangerous when it starts making promises. Keep it decorating the message and it pays for itself; let it describe the product and it becomes a bill you settle later.

None of this is an argument against the tools. Inside the right boundary they take real cost and real time out of a content system, which is why they sit inside our content and video work rather than off to one side of it. The boundary is simply narrower than most people assume: generate the world around the product, photograph the product. If you are not sure which side of that line your current image library sits on, an outside read of the whole site — the kind our Growth Scan gives you in a couple of minutes — is usually the quickest way to find out.