How to tell if an image is AI-generated

Image models got very good, very fast. There is no single trick that catches every fake anymore. What still works is a habit: look at the places where models cut corners, and ask whether the photo could have been taken by a real camera in a real place. This guide lists the nine tells that show up most often in the images on AI or Real.

The 30-second checklist

  1. Hands and fingers: Hands are still the most common giveaway. Count fingers, then look at how they hold things.
  2. Text and signs: Background text that almost reads as words is a strong sign of a generated image.
  3. Reflections and mirrors: A reflection has to agree with the scene. Generated images often get the geometry wrong.
  4. Light and shadows: Every shadow in a real photo points away from a real light. Check that they agree.
  5. Skin and faces: Generated faces often have airbrushed skin, perfect symmetry and hair that dissolves at the edges.
  6. Background details: Models focus on the subject. The background is where they cut corners.
  7. Physics and logic: Ask whether the scene could exist. Generated images break rules a camera cannot.
  8. Repetition and symmetry: Nature repeats with variation. Generated images sometimes repeat without it.
  9. Texture and noise: Real cameras add noise and imperfections. Generated images are often too clean.

1. Hands and fingers

Image models learn from millions of photos where hands are small, blurred or partly hidden. They know roughly what a hand looks like, but not how it is built. That is why errors cluster around joints, nails and the places where a hand touches something.

  • Count the fingers on every visible hand, including people in the background.
  • Follow each finger from knuckle to nail. Does it end where it should?
  • Check where the hand meets an object: cup handles, phones, steering wheels, other hands.
  • Look for rings, watches and bracelets that blend into skin or change shape.

False alarm: Motion blur, wide-angle lenses and odd poses make real hands look strange too. One weird hand is a hint, not proof. More on hands and fingers.

2. Text and signs

Language and pixels are learned differently. Recent models can write a short headline correctly when the prompt asks for it, but the small text nobody asked for, like shop signs, price tags, book spines and license plates, often comes out as letter-shaped noise.

  • Zoom into the smallest text you can find, not the largest.
  • Check whether the same word is spelled the same way twice.
  • Look at license plates, menus, keyboards and clock faces.
  • Ask whether a logo belongs to a brand you could name.

False alarm: Foreign scripts, stylised fonts and motion blur can make real text unreadable. Gibberish in a language you can read is the stronger signal. More on text and signs.

3. Reflections and mirrors

A real reflection is a strict copy of the world, flipped and distorted by the surface. A model draws something that looks like a reflection without computing it, so the mirrored version often disagrees with the original.

  • Compare the object with its reflection. Same pose, same colours, same count?
  • Check catchlights in both eyes for matching shape and position.
  • Look for reflections that are missing where a shiny surface should show one.

False alarm: Curved glass, two-way mirrors and wet surfaces distort real reflections in ways that look wrong at first glance. More on reflections and mirrors.

4. Light and shadows

Physics forces a photo to be consistent: one sun means shadows fall the same way. Image models blend many lighting setups they have seen and sometimes put two suns in one scene.

  • Find two or more shadows and check that they point the same direction.
  • Ask where the light comes from. Can you see or infer the source?
  • Notice light that is too perfect for the situation, like studio glow in a garage.
  • Look for blown highlights and noisy shadows, which point to a real sensor.

False alarm: Multiple lamps, reflectors and flash create real scenes with several shadow directions. More on light and shadows.

5. Skin and faces

Faces are what image models are best at, which makes them hard. The giveaways are rarely in the face itself. Look at the edges: where hair meets background, where glasses sit on ears, where earrings hang.

  • Follow the hairline and loose strands against the background.
  • Check that earrings, glasses and collars match on both sides.
  • Look for pores and blemishes. Is the skin too clean for the lighting?
  • Count teeth and check that they have edges.

False alarm: Beauty filters and heavy retouching make real portraits look generated. Professional headshots are the hardest real photos to judge. More on skin and faces.

6. Background details

The subject of a generated image gets the most attention, both from the model and from you. Move your eyes to the edges. People in a crowd may lack faces or merge into each other. Windows on a building may not line up floor by floor.

  • Look at people in the background. Do they have faces, hands and feet?
  • Follow straight lines like railings, tiles and window rows. Do they stay straight?
  • Ask what each background object is for.
  • Notice clutter. Real places are rarely tidy.

False alarm: Shallow depth of field blurs real backgrounds into shapes that look melted. More on background details.

7. Physics and logic

A camera only records what is there. A model can produce a fork with four handles, steam that flows down, or a cup that is both in front of and behind a plate.

  • Find every place two objects touch and check that it makes sense.
  • Compare sizes against things you know, like doors, cups and phones.
  • Look at liquids, smoke and fabric. Do they move with gravity?

False alarm: Forced perspective and long lenses make real scenes look impossible. Street photographers do this on purpose. More on physics and logic.

8. Repetition and symmetry

Models generate textures by learning patterns, and they sometimes repeat them too exactly: identical leaves, cloned clouds, the same face twice in a crowd, a row of fruit where every piece is the same size.

  • Look for clones in crowds, leaves, bricks and food.
  • Check whether paired things like earrings and shoes are mirror copies.
  • Ask whether the composition is too neat for the situation.

False alarm: Mass-produced objects are identical in real life too. Supermarket shelves are full of clones. More on repetition and symmetry.

9. Texture and noise

Every real sensor adds noise, especially in dim light. A night scene with no grain at all is suspicious. Generated images often have an even, polished surface where every material, from wood to skin to fabric, has the same smooth finish.

  • Look at dark areas for grain. Is there any?
  • Zoom into hair, grass and fabric for swirls or painterly strokes.
  • Check whether different materials have different textures.

False alarm: Phones apply heavy noise reduction and sharpening, which can make real photos look painted. More on texture and noise.

What does not work

Metadata is unreliable: social networks strip it, and anyone can edit it. Content credentials (C2PA) help when they are present, but most images online carry none. AI detector tools give a probability, not a verdict, and they fail on compressed or cropped images. Treat them as one more hint.

The most reliable check is the oldest one: find the source. A reverse image search that leads to a named photographer, a news agency or an older upload is strong evidence that a photo is real.

Practice

Reading about tells helps. Seeing them helps more. Play the daily game, then switch to Spot the AI to compare four photos at once.

Questions

Can you always tell if an image is AI-generated?

No. The best current models fool most people on some photos.

What is the most reliable sign of an AI image?

Garbled background text and broken hands are still the most common giveaways. Neither appears in every fake, so check several tells.

Are AI image detectors accurate?

They help as a second opinion but produce false positives on real photos and miss edited fakes. Use them together with a source check.