What makes an AI photograph believable

A photograph makes a claim that a drawing does not. It says this existed, light came off it, and somebody stood in front of it with a camera. Viewers read that claim first. They have believed it since the daguerreotype was announced in Paris in 1839, and they do not stop believing it because a picture came from a model rather than a lens. Believability is the word for whether a generated image survives being read that way, and a viewer settles it at a glance, usually without being able to say what settled it.

A generated image fails in a predictable direction, because a model returns something close to the middle of the millions of pictures it learned from. The light arrives from everywhere at once. Skin is even and edges are crisp, and the background falls into one uniform blur that stops exactly at the subject's outline. Nothing carries grain. Each of those is a thing a camera would not do, and people who have never held a camera still feel it, which is why the usual verdict on a weak generated image is simply that it looks fake.

So the first habit a professional builds is to describe the photograph before the subject. That means deciding where the camera stands, how high it sits, how far back it is, which lens it carries and what format it shoots. A prompt that asks for a premium bottle of olive oil in beautiful light gets the average of every image ever captioned that way. A prompt that puts the camera at table height, a metre back, on a 90mm lens, with window light from the left, asks for one particular picture, and the model understands those words well because photographers have always described their own work in them. Height and distance also carry meaning, which the next lesson takes up.

The second habit is to get the light right before worrying about detail. A real light source is one thing in one place, so it leaves one family of highlights on shiny surfaces and it falls off with distance according to the inverse-square law. Models blend thousands of lighting set-ups, which is why a generated bottle often carries three highlights that disagree about where the window is. Sharpness does not fix this. Turning up clarity or asking for hyper-detail makes it worse, and gives the waxy skin and crunchy edges people mean when they call an image plastic. A soft frame with one convincing light reads as a photograph.

The third habit is to put back what a camera would have recorded and the model left out. Real depth of field is continuous, so sharpness falls away gradually with distance, while a generated background is often one flat blur behind a cut-out edge. Real skin has pores, fine hair and uneven colour. Real pictures carry grain. When a person is in the frame, four checks in a fixed order catch most problems, so look at the hand where it touches something, the catchlights in both eyes, the colour cast on the skin, and whether the eye line goes where you asked.

Believable and true are separate questions, and a professional keeps them apart. A generated world behind a real product is a composite, and composites have been ordinary retouching practice for decades. A generated product is different. If the label came out of the model, the picture is making a statement about an object a customer is about to pay for, and models are weak at exactly the details that customer will check, such as small type and exact colour. So the working route starts from photographs of the real thing, taken on a phone by a window, which give the model something true to hold. The product is then composited into the generated scene rather than drawn by it.

The last habit is honesty. An image that looks like a photograph makes the photograph's claim whether or not a caption says otherwise, and on real events that claim is judged against every detail a witness would know. For a background or a mood, nothing has been misrepresented to anyone. For a product, a real person or a news event, a false photograph is a false statement, and a label added afterwards does not always repair it. Content Credentials, the provenance standard published by the C2PA, can record how an image was made, though some platforms still strip that information on upload. Decide which side of the line each job sits on before you start.