When people talk about the “giveaways” of AI-generated images, the discussion often turns immediately to what is represented: impossible hands, strange objects, wrong reflections, fake text, distorted anatomy. Those things matter, of course, but there is another question that is more purely photographic: not what the image shows, but how the image appears to have been made.

A real photograph normally carries traces of an acquisition chain. Light passes through a lens, reaches a sensor or film, is affected by focus, aperture, exposure, noise, grain, demosaicing, sharpening, compression, scanning, or other processing steps. These traces are often subtle, but they form a kind of technical history embedded in the image.

AI-generated images can imitate the look of a photograph, but they do not naturally inherit this full image-formation process. This can leave technical clues.

One of the most obvious is noise. Real sensor noise is not just a uniform sprinkling of random speckles. It usually depends on exposure, tonal values, ISO, sensor behaviour, and processing. Shadows often behave differently from highlights. Smooth areas behave differently from detailed ones. Film grain, too, has its own material character: it is not just texture laid over an image.

AI images often have a suspicious relationship with noise. They may be too clean, too smooth, too uniformly detailed, or oddly denoised while still showing a lot of artificial micro-detail. Sometimes the detail is present, but it does not feel acquired. It feels invented.

This is especially visible at high magnification. A real photograph tends to show a coherent relationship between sharpness, grain, blur, compression, and texture. AI images may show sharp detail in places where optical softness would be expected, or smooth blur without a convincing photographic noise structure. Edges can look too clean, too painterly, or inconsistent. Fine textures such as fabric, hair, grass, skin, paper, or stone may contain high-frequency information that resembles detail, but does not resolve into meaningful structure.

Then there are deeper technical traces. Digital photographs often contain demosaicing behaviour from the camera’s colour filter array, sharpening halos, chromatic noise, compression artifacts, lens softness, and sometimes even sensor-specific fingerprints. AI images usually do not have these traces unless they are simulated or introduced later. Adding JPEG compression or grain may make the image look more photographic at first glance, but it does not recreate the complete chain of a real capture.

This leads to an interesting question: what happens if we add “film grain” to an AI-generated image?

The answer is: it helps, but only superficially.

Adding grain can hide the overly clean look of an AI image. It can make the image more pleasant, more textured, and less sterile. But it does not automatically make the image photographic. A simple grain layer is usually just that: a layer. It does not fix synthetic texture underneath. It does not create real camera noise, real film response, real demosaicing traces, real optical sampling, or a coherent compression history.

The giveaway then shifts. The problem is no longer simply “there is no noise.” The problem becomes: does the noise belong to the image?

Fake grain can look suspicious when it has the same strength everywhere: in shadows and highlights, on skin and sky, over sharp objects and blurred backgrounds. Real grain or noise is usually tied to the physical or digital process that produced the image. It interacts with tone, exposure, scan, compression, and sharpening. A uniform overlay may make the image look more textured, but not necessarily more acquired.

So the most useful way to think about this is not: “Does the image have grain?”

It is: “Does the image have a believable photographic history?”

A convincing photograph is not only made of subject, composition, and lighting. It also has a technical provenance. Its imperfections are not random decorations; they are consequences of an imaging process. AI images can imitate many of these consequences, and they are getting better at it. But adding film grain alone does not reconstruct the missing chain.

Film grain can make an AI image look less clean.

It does not, by itself, make it look truly photographed.