Two years ago you could spot a generated image from across the room. Six fingers, melted ears, gibberish text on every sign. That era is over. The current generation of image models renders hands correctly, writes legible text and reproduces skin texture well enough that the obvious checklist no longer works.
What has not changed is that generated images are built differently from photographs. A camera records light through a lens onto a sensor and writes a specific, boring set of technical facts into the file. A generator paints pixels from noise and writes nothing of the sort. Those structural differences survive even when the picture looks perfect. Here are the seven that still hold in 2026.
1. The metadata is missing or impossibly thin
Open any photo taken on a phone and you will find a dense record: camera model, lens, focal length, exposure time, ISO, orientation, often GPS coordinates and a capture timestamp accurate to the second. Generated images carry none of it, or carry a suspiciously small set of fields written by whatever software last saved the file.
The catch: social platforms strip metadata on upload. A photo downloaded from Instagram is also bare. So missing metadata is not proof of anything on its own — it only matters when someone hands you what they claim is an original straight off a camera. Then a stripped file is a real question to ask.
2. Compression is too uniform
A JPEG from a camera compresses different parts of the frame at slightly different rates, and every subsequent save changes that pattern in a way that depends on what is in each region. Error level analysis makes that pattern visible: you re-save the image at a known quality and map where the difference is unexpectedly large.
A fully generated image is often eerily even — it never went through a lens or a sensor, so it has no native compression history. A real photo with an inserted region shows the opposite: one bright, sharply bounded patch against a calm background. That patch is where something was added. The checker on this page draws boxes around exactly those regions.
3. Light does not obey a single source
Generators compose a scene from learned patterns rather than from physics. Look for shadows that fall in two directions, a catchlight in one eye that is missing from the other, a reflection in a window that does not match what is in front of it, or a subject lit warmly while the background is lit coolly with no visible reason.
4. Repetition in the background
Foregrounds get the model's attention; backgrounds get filled in. Crowds are the classic giveaway — the same face appears twice, a person has three arms at the edge of frame, a fence post skips a beat, a pattern in a carpet resets. Zoom to 200% and scan the edges of the frame rather than the subject.
5. Text that is nearly right
Modern models write real words. What they still struggle with is text that has to be consistent: a price that appears twice in the same image, a licence plate that follows a country's format, a logo whose kerning matches the real brand, a document whose font weight is identical in every line. Nearly right is the signature. Wholly wrong belonged to 2023.
6. Texture that has no grain
Every sensor produces noise, and that noise is consistent across the frame because it comes from the same silicon. Generated images have texture rather than grain: skin is smooth in the wrong places, fabric weave disappears at the shadow line, hair renders as ribbons instead of strands. Look at a flat, dark region — a wall in shadow, a black jacket. A real photo there is faintly noisy. A generated one is often perfectly clean, or noisy in a pattern that does not match the rest of the image.
7. Content Credentials, watermarks and generator strings
This is the only sign that can be definitive. Several major tools now embed a signed provenance record or an invisible watermark declaring how the file was made. When one is present and intact, you are not guessing any more.
Absence proves nothing — a screenshot removes it instantly — but presence is close to conclusive. We wrote a full explainer on what C2PA Content Credentials actually prove and where they break down.
Putting it together: never trust one signal
Each sign above has a false positive. A heavily edited real photo trips the compression test. A studio portrait trips the lighting test. A screenshot trips the metadata test. The only reliable method is to run several independent checks and look at where they agree:
- Provenance — is there a signed manifest, watermark or generator string?
- Physical forensics — does compression, noise and metadata behave like a camera file?
- Model review — does a vision model see anatomical, lighting or physics errors?
- Custody — can you prove the file you analysed is the file you received?
When three of the four agree, you have an answer you can defend. When they disagree, you have a file that deserves a human look — which is a far more useful outcome than a single confident percentage that turns out to be wrong.
What to do when the answer matters
If you are approving an insurance payout, publishing a news photo or accepting a marketplace listing, a screenshot of a detector result is not evidence. Hash the file the moment it arrives, keep an append-only record of every action taken on it, and export a certificate that anyone can re-verify from the hash alone. That is the difference between an opinion and something that survives a dispute.
You can run the same four-layer check on your own photo below, or go straight to the AI image detector.