How to tell if an image is AI generated

Look for repeated texture, wrong reflections and text that falls apart under zoom. Then check the pixels with a detector, because the newest generators have fixed most of the tells people were taught to look for.

What still gives an AI image away

Generation models build an image from statistical patterns rather than from a scene that existed. That leaves marks. These are the ones you can still find by looking, ordered by how often they appear.

Text falls apart. Signage, labels, book spines and number plates turn into letter-shaped marks that spell nothing. Zoom in on any writing in the frame.
Repeated texture. Brickwork, foliage, crowd faces and fabric weave repeat with a regularity that a lens never produces.
Reflections disagree. A mirror, a window or an eye shows something that does not match the scene, or shows nothing at all.
Jewellery and straps break. A necklace passes behind a neck and comes out at the wrong height. Watch bands and glasses arms end mid-air.
Backgrounds melt. Objects behind the subject lose structure: a chair with five legs, a door frame that bends.
Skin is too even. Pores, fine hair and blemishes are smoothed into a plastic surface, especially on cheeks and foreheads.
Lighting has no single source. Shadows fall in two directions, or a face is lit from the front while the scene is lit from behind.

Still works in 2026

Text and signage under zoom
Repeated texture in fabric and foliage
Reflections that contradict the scene
Jewellery and straps that break continuity

Stopped working

Counting fingers
Glassy or waxy skin
Warped ears and teeth
Obvious anatomical errors

Every item on the left used to be reliable. Most of them were fixed between 2023 and 2025.

Why your eyes stopped being a reliable test

The tells people were taught to look for are exactly the ones model authors prioritised. Hands were the joke of 2023, so hands were fixed. Text was next. What remains are failures of physics and continuity rather than failures of anatomy, and those are harder to spot because they need you to reason about the scene rather than scan it.

There is a second problem. A generated image that has been screenshotted, posted to a platform and re-saved loses the fine detail that carried the tell. By the time a picture reaches you through a group chat, the evidence you were going to look for has been compressed away.

How to check an image properly

Get the original file
Ask the sender for the file rather than a screenshot. Compression destroys the signal that both your eyes and a detector rely on.
Look at the frame at full size
Zoom to 100 percent on any text, reflection, or point where two objects overlap. That is where continuity breaks.
Run a pixel-level check
A detector reads the statistical texture of the image rather than its content. It catches what survives a screenshot, which your eyes do not. Check an image here — it runs in your browser and the file is not uploaded.
Read the region map, not just the score
One hot region in an otherwise clean frame means an edit, not a generated picture. Those are different problems with different consequences.
Check where the image came from
Run a reverse image search for the earliest copy. A picture online since 2019 is not diffusion output, whatever the pixels suggest.

What a detector sees that you cannot

A camera sensor and a diffusion model produce different relationships between neighbouring pixels. The difference is invisible at any zoom level, and it survives a screenshot. A detector reads that, scores the whole frame, then scores the frame again as a grid of regions.

The grid is the part that matters most, because it separates two cases a single number cannot. A fully generated image scores high everywhere. A real photograph with one object added scores low everywhere except the tile containing the edit.

12 9 14 11 87 16 8 13 10

Tiles at or above the decision line of 65 are filled. One hot tile in a cold frame points to a local edit.

A real photograph with an inserted object. Eight tiles sit below the decision line; the ninth does not.

How to read the score you get back

A score is a likelihood, not a verdict. What makes it usable is knowing where the boundaries sit before you look at the number, so the same result always reads the same way.

No AI signal 0–20
Probably real 20–45
Uncertain 45–65
Likely AI 65–90
Very likely AI 90–100

The five bands used here, with the decision line at 65. Detectors that do not publish their bands are asking you to trust a number with no scale behind it.

Scores in the middle band are the honest answer to a hard image, not a failure. Heavy compression, small crops and screenshots all push results toward the middle. If you land there, the image is telling you it cannot carry a confident answer.

When to stop and get a professional

None of this is forensic examination. If the answer decides an insurance claim, a disciplinary hearing, an immigration case or anything that reaches a court, a score from a free browser tool is not the right instrument. Instruct a qualified digital forensic examiner and keep the original file untouched.

Questions people ask

Can you tell if an image is AI generated just by looking?

Sometimes, and less often every year. Text, repeated texture and broken reflections still give away many images. Current top-end generators clear all of those checks, and any image that has been screenshotted or recompressed loses the detail your eyes would use. Looking is a useful first pass and a poor final answer.

What is the fastest way to check one image?

Drop it into a pixel-level detector and read both the score and the region map. That takes a few seconds and catches what survives a screenshot. Follow it with a reverse image search if the answer matters, because an early publication date rules out generation more firmly than any score.

Does zooming in still help?

Yes, on specific things. Zoom to full size on any writing, any reflective surface, and any point where two objects overlap or pass behind each other. Those are continuity problems rather than anatomy problems, and continuity is what generators still handle badly.

Do AI images always have a watermark or metadata?

No. Some generators embed Content Credentials or an invisible watermark, but most images arrive with neither. Metadata is stripped by almost every platform on upload, so an absent marker proves nothing at all. That is why pixel analysis is the only method that still works on an image that has travelled.

Can a real photo be mistaken for AI?

Yes, and it is the failure that causes real harm. Heavy retouching, upscaling, aggressive noise reduction and phone night modes all push genuine photographs upward. If a picture you know is real scores high, re-check the original file rather than a screenshot of it.