SI Detector Guide

How to spot SI (AI) images

SI (Super Intelligence) is the new name for advanced AI: models such as ChatGPT, Midjourney, Gemini or Stable Diffusion. Generators improve every few months, so no single tell is definitive any more. But SI images, still widely called AI images, fail in predictable places, and a two-minute check of the right details catches most of them. Here is what to look at, roughly in order of usefulness, and how to confirm with the file's metadata.

1. Hands, fingers, teeth and ears

Hands remain the classic failure. Count the fingers, then check the joints: SI hands often have the right number of fingers but the wrong number of knuckles, fingers that merge where they cross, or a thumb on the wrong side. Look at what the hand is holding: cups without handles that still get gripped, phones with no edges, straps that go nowhere. Teeth are another weak point, either too many, too uniform or fused into a single white band. Ears and earrings frequently mismatch between the left and right side, and glasses may have a bridge that does not connect the two lenses.

Recent models (Midjourney v6 and later, Flux, Imagen, GPT-image) get hands right most of the time, so a perfect hand proves nothing. A wrong hand, however, is still a strong signal.

2. Text, signs, logos and keyboards

Zoom in on every piece of writing: street signs, shop names, book spines, T-shirt prints, newspaper headlines, screens in the background. Older generators produce letter-shaped gibberish. Newer ones write real words but stumble on longer sentences, repeat a word, mix alphabets or misspell a brand. Logos are often "almost right": correct colours and shape, slightly wrong letterforms. Keyboards, calendars, clock faces and rulers are great tests because their structure is rigid and the model rarely reproduces it exactly (keys in odd rows, clock hands that do not meet the centre, a dial with 13 hours).

3. Reflections, shadows and light

Physics is hard to fake consistently. Check that mirrors, windows, water and sunglasses reflect what is actually in front of them, at the correct angle. Look for a single light direction: every shadow should fall the same way and have a length consistent with the others. Generated scenes often mix a hard shadow under one object with a soft, missing or reversed shadow under the next. Reflections in eyes (catchlights) should match between both eyes and match the visible light sources.

4. Skin, hair, fabric and surfaces

AI portraits tend towards a smooth, evenly lit, slightly waxy skin with perfectly symmetric pores, or the opposite: an over-sharpened, hyper-detailed look. Hair is a giveaway when strands melt into the background, into the skin, or into each other with no individual fibres at the edges. Fabric patterns (stripes, checks, knits) should continue logically across folds and seams; generators lose track and the pattern jumps. Look at repeated materials such as bricks, tiles or wood grain for spots where the texture becomes mushy or suddenly changes scale.

5. Symmetry, repetition and the "too perfect" look

Diffusion models love balance. Faces are often more symmetric than real faces, crowds contain near-duplicate people, and patterns repeat with suspicious regularity. Conversely, objects that should be symmetric (a car's headlights, a building's windows, a pair of shoes) come out subtly different from one side to the other. Compare left and right explicitly; it is faster than staring at the whole picture.

Composition itself is a clue. Stock-photo framing, a shallow depth of field on everything, cinematic colour grading and a subject staring straight into the lens with a flawless expression are all typical of prompted images. Real snapshots are messier.

6. Backgrounds and secondary objects

The model spends its effort on the subject. Behind it, buildings lean, stairs lead nowhere, railings change count, chairs have three legs, and people in the distance are blobs with no faces. Follow lines: a road, a fence or a shelf should stay continuous when it passes behind the subject. Check that objects that must connect actually do (a leash to a dog, a strap to a bag, a cable to a lamp). Bokeh is also revealing: real lens blur has consistent circular highlights; AI blur can be uneven, or sharp objects can sit in a blurred plane.

7. Metadata and Content Credentials (C2PA)

When the file has not been screenshotted or re-saved, its metadata is the most reliable evidence available, which is exactly what our SI image detector reads for you.

8. Typical signatures by generator

9. A two-minute checking routine

  1. Run the picture through the SI image detector (and any caption or article through the SI text detector) and read the clues, not just the number.
  2. Reverse image search it (Google Lens, TinEye, Bing). A real photo usually has an origin: a news agency, a photographer, an older post.
  3. Zoom to 200% on hands, text, eyes, jewellery and anything in the background.
  4. Check the light: one direction, coherent shadows, matching reflections.
  5. Ask where it came from. An image that appears only on anonymous accounts, with no photographer credit, during a breaking event, deserves suspicion regardless of what any tool says.

Remember the asymmetry: a clear failure in any of these checks is meaningful; passing all of them only means "no evidence found". Detectors, including our free SI detector, are one more clue in that list, not a judge. For writing rather than pictures, see our companion guide on how to spot SI text, and to understand what a detector measures, read how SI detectors work.

Check an image nowFree SI (AI) image detector with clues and metadata. Play SI or Not?Train your eye on ten images.