Home Blog How to Tell If an Image Is AI-Generated

How to Tell If an Image Is AI-Generated

Updated 2026-07-30 · 8 min read · by the FakeSec team

Most advice on this subject is two years out of date. Counting fingers stopped being reliable around 2024, and \u201clook for garbled text\u201d died when generators learned typography. Here is what actually works in 2026, in the order you should try it \u2014 including the point where looking at the image stops helping entirely.

Start with the metadata, not the pixels

Before you squint at hands, check what the file says about itself. It takes ten seconds and sometimes ends the investigation immediately.

Several major generators sign their output. DALL·E 3, Sora, Adobe Firefly and Google Imagen embed C2PA Content Credentials — a signed manifest declaring the image as AI-generated. Local Stable Diffusion setups running through AUTOMATIC1111 or ComfyUI frequently store the entire prompt, seed and model in the file, because most people never realise it is there.

The catch is that metadata dies easily. Screenshot an image, send it through WhatsApp, upload it to Instagram — the manifest is gone. So a finding is meaningful, but finding nothing means nothing at all. Empty metadata is the normal state of almost every image on the internet.

What still gives generated images away visually

When metadata is absent, you are left with the picture. These are the tells that still hold up in 2026 — note that they are all about consistency rather than quality.

Ears, teeth and jewellery. Not hands any more — modern models handle hands well. Ears are rarely a training focus and are as individual as fingerprints, so generated ones often differ from each other. Earrings appear on one side only. Teeth merge into a single white block.

Backgrounds that mean nothing. The subject is sharp and coherent; the world behind them is not. Architecture that does not line up, patterns that shift mid-repeat, objects that merge into each other.

Light without a source. Shadows falling in directions that contradict the highlights, reflections that do not correspond to anything in the scene, catchlights in the eyes that disagree between left and right.

Detail that never decays. Real optics produce falloff — detail drops away from the focal point. Generated images often keep the same density everywhere, which reads as subtly unreal even when you cannot name why.

Skin without texture. No pores, no fine lines, no sensor grain. Retouching produces something similar, so this is a hint rather than proof.

What stopped working

Three pieces of advice you will still find in most articles, all obsolete:

Counting fingers. Current models get hands right most of the time. A correct hand is no longer evidence of anything.

Garbled text. Ideogram was built specifically for typography, and others have caught up. Clean, readable text in an image no longer means a human made it.

\u201cIt looks too perfect.\u201d Generators now deliberately add imperfection: grain, chromatic aberration, motion blur, even fake lens dust. Polish is not a signal in either direction.

When looking is not enough

Here is the uncomfortable part. Photorealistic output from current models routinely passes human inspection. Studies consistently show people perform close to chance on modern generated faces — and being told to look carefully does not improve the result much.

That is not a failure of attention. The remaining evidence lives in places the eye has no access to: the frequency domain, the noise floor, compression history, statistical distribution of colour. You cannot see a frequency profile by squinting harder.

This is where a detector earns its place — not because it is smarter than you, but because it measures things you physically cannot.

How to check properly, in order

1. Read the metadata. Free, instant, occasionally decisive — our EXIF viewer needs no account.

2. Reverse image search. Catches the most common case of all: a real photo, recycled with a false story attached. No forensic tool can catch that, because nothing is wrong with the file.

3. Look for the visual tells above — ears, background, light, falloff.

4. Run a forensic analysis for what the eye cannot reach — our AI image detector runs twelve methods and shows every score.

5. Weigh the source. Who posted it first, and where did they get it? This is often more decisive than anything in the file.

Key points

  • Check metadata first — it is free and sometimes ends the question immediately.
  • Finding no metadata proves nothing: every messenger strips it.
  • Fingers and garbled text are obsolete tells. Ears, backgrounds and lighting are not.
  • Photorealistic output passes human inspection — that is expected, not a failure.
  • Reverse image search catches recycled real photos, which forensics cannot.
  • No method is conclusive alone. Agreement across several is what matters.

Frequently asked questions

Can AI images be detected 100% of the time?
No, and anyone claiming otherwise is selling something. Generators improve specifically to remove the traces detectors look for. Treat any result as strong evidence rather than proof.
What is the single most reliable check?
There isn't one, which is precisely the point. Metadata is decisive when present and useless when absent; visual tells depend on the subject; forensic analysis weakens under compression. Agreement across methods is the signal.
Does zooming in help?
Sometimes — ears, teeth, jewellery and text reward close inspection. But for modern photorealistic output, the decisive evidence is statistical and not visible at any magnification.
Are screenshots harder to check?
Considerably. A screenshot is a fresh image of an image: metadata gone, compression history rewritten, noise floor replaced. Always work from the original file when you can get it.

Check an image yourself

Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.

More from the blog

Try it on your own image

Free account, 3 checks included, no card required.