Guide

How to tell if an image is AI-generated

“Count the fingers” stopped working years ago. These are the checks that still catch AI images from current generators — starting with the ones that give certain answers.

Free · no sign-up · your file never leaves your device

Drop the image to run checks 1–4 automatically, click to choose, or paste

JPEG · PNG · WebP · AVIF · HEIC (Safari) — checked on your device

Checks that give certain answers

1. Content Credentials

Look for a signed C2PA manifest. OpenAI, Adobe, Microsoft and others record AI generation in it. The tool above reads it automatically; the C2PA viewer shows every detail.

2. The IPTC “made by AI” field

Google and Meta, among others, mark generated images with the digital source type trainedAlgorithmicMedia in XMP metadata.

3. Generator settings

Stable Diffusion tools store the prompt, seed and model in PNG text chunks; ComfyUI stores its whole workflow; NovelAI hides its settings in the alpha channel.

Checks that give strong hints

4. Camera metadata — or its absence

Real photos straight from a phone carry the make, model, lens, exposure and time. A full set that matches the claimed time and device supports authenticity. None at all is normal for shared images, so only count it as weak evidence of AI.

5. Size and shape

Generators favour fixed sizes: 1024×1024, 1536×1024, 832×1216. Phone cameras produce 4032×3024 and similar. Cropping and resizing hide this.

6. The first appearance

Reverse-search the image and sort by date. AI images often trace back to an AI art community or an account that posts nothing but generated pictures.

7. Corroboration

A real event is captured by many people. A dramatic image that exists in only one version, from one account, with no other coverage, deserves suspicion.

Visual checks that still catch many fakes

8. Small text

Headlines are often perfect now; read the fine print, price tags, licence plates and background signs, which still turn into near-words.

9. Things that must match

Earrings, glasses arms, shirt buttons, shoelaces, window frames and fence posts should be consistent from one side to the other and continuous behind objects.

10. Physics

Check that shadows fall away from the light source together, that reflections in mirrors, water and sunglasses show the right scene, and that objects rest on surfaces rather than float.

11. Too-even quality

Cameras blur things at different distances and add noise in dark areas. Generated images often render everything — skin, background, fabric — to the same smooth finish.

12. The story behind it

Ask who benefits from you believing the image, and whether the original poster can provide the camera file. A real photographer usually can; someone sharing a generated image usually can’t.

No single item on this list is conclusive except the first three. When four or five of the others point the same way, you can be fairly confident — and the automatic check above will usually agree.

How accurate is it? Our measured numbers

We test GPTTrace on labelled image samples and publish the results, including where it does badly. It is tuned to keep false accusations rare, so it misses some AI content rather than flag real work.

image check: AUC 0.938 (cross-validated)

679 labelled samples (399 AI, 280 human), run 2026-10-08. At the “Likely AI” line it caught 64% of AI samples and wrongly flagged 5% of human ones.

SourceTruthSamplesResult at “Likely AI”
gemini-nano-bananaAI4040% caught
midjourney-v6AI4068% caught
midjourney-v5AI4073% caught
flux-devAI4013% caught
flux-schnellAI4048% caught
sdxlAI40100% caught
gpt-imageAI4030% caught
klingAI3997% caught
leonardo-stablecogAI4098% caught
bitmind-imagine-mixAI4080% caught
fullsize-photosHuman4010% wrongly flagged
open-images-photosHuman405% wrongly flagged
lfw-facesHuman400% wrongly flagged
caltech-objectsHuman403% wrongly flagged
coco-photosHuman400% wrongly flagged
ffhq-facesHuman400% wrongly flagged
celeba-facesHuman4018% wrongly flagged

Data sources and method: methodology & accuracy.

Frequently asked questions

Do hands and fingers still give AI images away?
Rarely. Generators from 2024 onwards draw hands correctly most of the time. Hands holding or interacting with objects — a cup handle, a phone, interlaced fingers — still fail more often than hands at rest, but it is no longer a dependable test.
Is there one test that always works?
Only provenance: a signed Content Credential, an IPTC AI label or embedded generator settings. When the file says how it was made, you know. Everything else is a set of probabilities that become convincing when several agree.
Why do AI images look so polished?
Generators are trained and tuned on images people rated highly, which skews toward beautiful lighting, ideal composition and flawless skin. Real snapshots are messier. Many AI users now prompt for amateur, low-quality looks to counter this.
Are AI detectors better than people at this?
On average, yes — studies have repeatedly found people perform near chance on high-quality AI faces and photos. But detectors fail in different ways from people, so a careful human check of context and details plus an automatic check beats either alone.