How to tell if an image is AI-generated

The easy tells are disappearing

For a couple of years you could spot a generated image by counting fingers or reading the sign in the background. Current image models have largely fixed both. What has not changed is that generation and photography produce different kinds of detail, and that difference is still visible if you know where to look. None of the checks below is conclusive on its own. Treat them as a stack: one hit means little, four hits pointing the same way means a lot.

1. Zoom into fine texture

Cameras capture noise. Sensors are imperfect, light is uneven, and real surfaces have grain that varies across the frame. Generated images often render texture as an average of what that surface 'should' look like, which reads as slightly too even. Zoom to 200–400% on skin, fabric, foliage, brick or hair. In a photograph the detail stays messy and irregular. In a generated image it often smooths out, repeats, or dissolves into a soft wash when you get close.

2. Follow the edges where two things meet

Boundaries are hard. Look at where hair meets background, where a sleeve meets a wrist, where a chair leg meets the floor. Photographs have consistent edge behaviour: the same lens produces the same kind of transition everywhere in the frame. Generated images often have edges that are crisp in one place and soft in another for no optical reason, or strands of hair that merge into the background instead of ending.

3. Check whether the light agrees with itself

Pick the brightest object and work out where the light is coming from. Then check every other object in the scene against that. Shadows should fall in compatible directions, and reflective surfaces should reflect things that are actually present. This is one of the more durable tells, because it requires a model to keep a consistent 3D scene in mind rather than produce locally plausible pixels. Mismatched shadow directions and reflections that show nothing in the frame are both strong signals.

4. Read the background, not the subject

Attention goes to the subject, and so does the model's effort. Backgrounds are where things fall apart: architecture that does not structurally connect, patterns that repeat at odd intervals, a crowd where faces are mush, text on a distant sign that is almost-but-not-quite letters. Deliberately ignore whatever the image wants you to look at and audit the last 20% of the depth of field.

5. Check the metadata — but do not rely on it

EXIF data can carry a camera make and model, a lens, exposure settings and sometimes a C2PA content credential marking AI generation. When it is present and consistent, that is meaningful evidence. The catch is that it is almost always absent. Every major social platform strips EXIF on upload, screenshots carry none, and metadata is trivial to forge. Missing metadata tells you nothing at all — it is the normal state of any image that has passed through the internet.

6. Reverse image search it

Before analysing pixels, find out where the image has been. Google Lens, TinEye and Bing Visual Search will often surface an earlier posting, a stock listing, or the news story the image was lifted from. This is the highest-value check per second spent, and it answers a different and often more important question than 'is this AI': is this image being used to represent something it is not?

7. Run a detector — and read what it actually says

An image detector analyses statistical properties invisible to you: frequency-domain artefacts, compression inconsistency, texture distribution. It compresses all of that into one number. That number is a likelihood, not a verdict. A well-built detector will tell you which visual dimensions drove the score — texture, edge transitions, lighting, background consistency — so you can check its reasoning against your own eyes. If a tool gives you a confident yes-or-no with no explanation, be more suspicious of the tool than of the image.

1. 1Get the highest-resolution copy you can find — compression destroys the signals detectors rely on
2. 2Avoid screenshots and re-saved versions; re-encoding shifts the score
3. 3Run the scan and read the summary, not just the number
4. 4Combine the result with what reverse search and your own inspection turned up

What moves a score for the wrong reasons

Detectors respond to processing, not just to generation. Heavy retouching, aggressive beauty filters, upscaling, denoising, low-resolution sources and repeated re-sharing all push scores upward on genuine photographs. A professionally retouched portrait can score higher than a casually generated image. This is why an elevated score on a single image should never be the end of an inquiry. It is a reason to look harder, not a conclusion.

The honest bottom line

There is no reliable single test, and anyone selling you one is overselling. What works is convergence: reverse search shows no earlier origin, the background geometry does not resolve, the lighting disagrees with itself, and a detector scores it high with texture and edge signals. Four independent things pointing the same direction is worth something. One is not. And when the stakes are real — a news claim, a legal matter, an accusation about a person — the answer is a human review that considers the source and the context, with detection as one input among several.