How to tell if an image is AI-generated (and where every check fails)

A picture of a storm over a city lands in a group chat, and somebody asks the question that now follows half the viral images online: is it real? We make AI images all day, so we know both sides of this. There are honest ways to raise or lower your confidence in a picture. There is no way, for you or for us, to be certain from the file alone. This is a checklist in the order we would use it, with the limit of each step written next to it.

Step 1: look before you zoom

Start with the whole frame at normal size and ask whether the scene makes sense as an event. Who would have been standing there with a camera? Is the moment too perfect, with every element posed for the shot? Real events usually leave many imperfect photos. A generated scene often leaves exactly one flawless one.

The limit: this is a question about context, not pixels. A real photographer can get one perfect frame, and a careful faker can post a few variations.

Step 2: the visual tells, and why they are fading

Now zoom in. Models draw what is likely, not what is consistent, so the errors cluster where a scene has rules that must hold across the whole frame.

• Hands and joints. Count fingers, then check how a hand holds an object: a cup handle that passes through a finger, a ring that melts into skin, an arm with no clear elbow.
• Text and symbols. Signs, labels, licence plates and book spines that look like writing from a distance but turn into near-letters close up. Logos that are almost a known brand.
• Reflections and shadows. A mirror or a shop window that shows a different room, a reflection missing the person in front of it, shadows that point in two directions under one sun.
• Lighting logic. A face lit warmly from the left in a scene whose only window is on the right. Rim light on a subject with nothing behind it to cast it.
• Repeated textures. Bricks, leaves, crowds or carpet patterns that copy themselves, faces in the background that are the same face, hair that turns into a smooth smear.

The limit: these tells are getting rarer with every model generation, and they are also exactly what a careful creator fixes. We even wrote the guide for it: fixing hands and text in AI images. Real photos have their own oddities from lenses and phone processing, too. A clean result proves nothing; a single odd detail proves little.

Step 3: read the provenance data

A file can carry information about where it came from. Camera photos usually hold basic metadata such as the device and the capture time. On top of that there is an open standard for signed provenance, often shown to people as Content Credentials: a record attached to the file that can say which tool created or edited it and what changed along the way, protected so that tampering is detectable. Some cameras, editing apps and generators write it, and some platforms show it as a small clickable label.

Free online inspectors will show you whatever provenance a file still carries. If you find a signed record that says the image was generated or heavily edited, that is strong evidence. If it says a camera captured it and the signature checks out, that is strong evidence the other way.

The limit: absence means nothing. A screenshot, a re-save or an upload to most social networks strips metadata, and many tools never write it. Plain metadata fields are trivial to edit. Provenance tells you something when it is present and valid, and nothing at all when it is gone.

Step 4: run a reverse image search

Upload the picture, or a crop of its most distinctive part, to a reverse image search engine and sort the results by date. You are looking for the earliest copy. A photo of a real event usually traces back to a news outlet, a photographer or a local account with other pictures from the same place. A generated image often traces back to an AI art community, a prompt-sharing account or nothing older than the viral post itself.

The limit: a brand-new image has no history to find, whether it is real or not. Cropping, mirroring and filters can hide a picture from search, and finding an older copy only tells you the picture is old, not how it was made.

Step 5: be careful with detector apps

Automated detectors give you a percentage, which feels scientific. Treat it as one more opinion. They are trained on the models that existed when they were built, resizing or noise can fool them, and they err in both directions, including flagging real photos. Never accuse someone on a detector score alone.

Putting it together

1. 01Ask whether the scene makes sense as an event, and who else should have photographed it.
2. 02Check hands, text, reflections, light and repeated textures, especially in the background.
3. 03Inspect the file for provenance, and remember that missing data is not evidence.
4. 04Reverse search for the earliest copy and its source.
5. 05Weigh it all. Several weak signals pointing the same way are worth more than one strong-looking one.

The honest answer at the end is often "probably", not "definitely". That is fine. What you do with "probably" is the useful part: do not share it as fact, and say what you checked if you do share it.

The other side: your own renders

If you generate pictures yourself, in the studio or anywhere else, the same checklist tells you how to be a good neighbour on the internet. People cannot reliably tell, so the responsibility sits with whoever posts the picture.

• Label it. Say the image is AI-generated in the caption, and use the AI label when a platform offers one. Many platforms now require it.
• Do not pass a render off as a photo of a real event, a real product test or a real person. A fictional scene presented as fiction is fine; the same picture presented as news is not.
• Leave real people out of it unless they have agreed to it. A realistic picture of an identifiable person is exactly what people cannot check.
• Keep your records. Save the prompt, the model and the date for anything you publish, so you can show how a picture was made if anyone asks. Our post on who owns AI-generated images covers the rest of the commercial side.

Many models in the catalog produce pictures that pass a casual look. That is what makes them useful for concept art and storytelling, and why the label matters.