AI image detector: how to check if an image is AI-generated
AI images can no longer be spotted reliably by eye. Here is the order professionals use: what you can check yourself, what the file tells you, and what an AI image detector actually does.
Updated 2026-09-11 · Veriflied
Start with context and common sense
Before looking at pixels: where does the image come from, and why did you receive it? A reverse image search (Google Lens, TinEye) shows whether the picture exists elsewhere in a different context. A large share of 'new' images are old images with a new story, and exposing that requires no technology.
Does the image match its own claim? Weather, season, licence plates, signs in the right language. It sounds trivial, but most exposures start here.
What you can see in the image
Text in the background
AI models still write flawed text on signs, labels and plates in the background. Zoom in and read everything.
The logic of light
Shadows pointing different ways, reflections in windows and eyes that do not match the scene. Physics is still the hardest part for the models.
Repeating patterns
Tiles, bricks, fabric and crowds reveal copied regions and impossible repetitions.
Too perfect
Uniform sharpness everywhere, skin without pores, perfect symmetry. Real photos contain mess, noise and small flaws.
Edges and transitions
Hair against background, fingers around objects, glasses against a face. Transitions are still where it most often breaks.
What the file tells you
A real photo from a phone or camera carries metadata: make, model, capture time, often GPS and manufacturer-specific fields. An AI image typically has none of it, or has metadata that does not match the image's technical properties. Missing metadata does not prove AI (sharing through apps strips it too), but the pattern of what is missing and what remains is a trace. See the EXIF guide for details.
What an AI detector does
A detector does not look at the subject but at pixel statistics: the microscopic patterns a generator leaves behind, which survive compression and sharing. A good detector is trained on millions of real and generated images, is retrained when new generators appear, and shows a heatmap of the regions that carried the most weight, so you see why and not just what.
No detector is infallible, and a serious vendor says so openly. A forensic analysis therefore combines the detector with metadata, compression traces and the context of the case, so no single source of error decides the conclusion alone.
Check an image now
You can test it yourself: create a free VeriFly account and analyse your first 10 images at no cost. You get the AI probability, the heatmap and the full forensic breakdown of the file, not just a yes/no.
Frequently asked questions
Is there a free AI image detector?
Yes. VeriFly analyses your first 10 images free, with a full report rather than a bare score. Most free tools give a single number with no explanation of why.
How accurate is an AI image detector?
Good detectors catch the vast majority of fully generated images, but none hit 100 percent, and the numbers depend on generator and compression. Treat the result as strong evidence to be read together with metadata and context.
Can an image be partially AI?
Yes, and it is the most common manipulation today: a real photo where one region was altered with AI. That requires a detector working at region level that shows where. See the guide on AI-edited images.
Does social media destroy the evidence?
Sharing strips metadata and compresses the image, but the generative pixel patterns largely survive. Detection therefore also works on images from messaging apps and social platforms.
Try it on your own document
Create an account and analyse your first 10 images or documents for free. You get a report with the concrete findings, not just a score. All processing in the EU.