AI Image Detector

What AI image detection means

Generative models build pictures from learned patterns rather than from light hitting a sensor. Detection tries to notice the difference. It analyzes the pixels for traces that are typical of synthesis and weighs them to produce an estimate, much as an art conservator weighs brushwork, pigment and paper rather than relying on one tell.

The output is a classification with a confidence level, not a certificate. Drubl sorts images into Real, AI, CGI, Filtered or Artwork, because "not a plain photo" has several very different explanations.

Common artifacts and signals in generated images

Some of these you can spot with your own eyes, especially in older or weaker generators. Treat them as prompts to look closer, not as proof.

• Anatomy and fine structure

  Fingers, teeth, ears, jewellery clasps and eyeglass frames require consistent structure. Check whether they hold together when you zoom in.

• Text and signage

  Lettering on shirts, posters and shop fronts may look plausible at a glance but dissolve into near-letters when read slowly.

• Lighting logic

  Look for shadows that fall in different directions, reflections that do not match the room, and highlights in eyes that disagree.

• Surface texture

  Skin, fabric and foliage can look uniformly polished or repeat the same pattern in a way real surfaces rarely do.

• Background coherence

  Away from the subject, architecture, crowds and objects can lose logic: windows that merge, railings that end nowhere, duplicated faces.

Why modern generated images are hard to identify by eye

The obvious glitches of a few years ago are becoming less common, and people also edit, upscale and retouch outputs to remove them. A good generated portrait can survive a careful look. Meanwhile real photographs get beauty filters, aggressive sharpening and cropping that make them look synthetic. Confidence from squinting at an image is rarely earned.

That is why a software estimate is worth having as an additional input. It can register statistical patterns that human perception skips over. It still can be wrong, particularly on small, heavily compressed or screenshotted images.

The limits of metadata

Image files can carry camera details, edit history or content credentials. When a trustworthy signed record exists, it is valuable evidence about origin, not a guarantee that a depicted event happened. Many platforms strip or rewrite metadata, and screenshots generally do not preserve the original file's camera details or credentials. Missing original metadata is common and says very little.

The reverse also holds: metadata can be edited or copied, so a camera name in a file is not a guarantee. This is a main reason pixel-level analysis exists alongside provenance checks. If the picture comes from a moving clip, see how the same questions apply to footage on the AI video detector page.

Why several signals beat one guess

Any single indicator has innocent explanations. Soft detail could be a cheap lens, odd text could be a worn sign, perfect skin could be a filter. When several independent indicators agree, the estimate becomes more meaningful, and when they disagree, a low confidence reading honestly reflects that. Drubl combines multiple signals for this reason, and reports confidence rather than pretending to certainty.

Be especially careful with edited portraits. If a face is the heart of the question, whether it was swapped or synthesized, the deepfake detector page describes the additional checks that matter.

Getting a better read from your analysis

Use the largest, least-processed version you can find. A screenshot of a screenshot has lost most of the evidence. Crop out borders and watermarks only if you must. If you are tracing where an image came from, a reverse image search can find earlier copies that are better to analyze. Remember what the result cannot tell you: who made the image, why, or whether a caption attached to it is true.