Nebula AI Detector
AI image detection · Guide

How to Tell If an Image Is AI-Generated: 7 Methods

Most of the folk wisdom floating around — count the fingers, check the ears, look for odd shadows — was fixed years ago. Two things actually work: verifying provenance and testing physical consistency. Here they are, ranked by how much you can actually trust them.

Last updated: 2026-11-11

First, a correction to a widespread belief: "you can spot an AI image at a glance" is not true. By 2026 the output quality of mainstream generators is close enough to photography that gut-feel judgements are not reliable. What works is separating what you can verify from what you can only infer.

Method 1: read the generator signature in metadata

Most AI art tools write generation details on export: software name, model version, sampler, step count, prompt fragments, seed. Common signatures include Midjourney, Stable Diffusion, DALL·E, ComfyUI, NovelAI and Flux.

The value here: a hit is conclusive, and nothing else needs checking. If the original file names the generator that made it, the image is AI-generated and there is nothing to debate.

Distinguish "no signature" from "a real photograph"

The absence of a signature does not imply the photo is genuine. Signatures get stripped, and plenty of real photographs carry no generator information either. "No signature" is a neutral fact, not reverse evidence.

Method 2: look for AIGC labelling metadata

In China, the 《Measures for Labelling AI-Generated Synthetic Content》 and the mandatory national standard GB 45438-2025 have applied since 1 September 2025. They require generators to write an implicit label into file metadata: field names containing an AIGC marker, recording generation status, the service provider, the distribution platform, a unique ID, and a digital signature or hash.

There is also an explicit label — the visible "AI generated" notice users see on the content itself. In other words, images from compliant tools in China are now machine-readable. This is why an increasing number of tools specifically look for these markers.

Method 3: read camera EXIF

Working in the opposite direction: if the image carries camera make, model, lens, aperture, shutter, ISO, GPS coordinates and capture time, then it is very likely a real photograph.

This signal gets underrated, and it is genuinely strong: generators rarely fabricate a complete set of camera parameters. The reverse is also informative — a suspiciously clean EXIF on an otherwise flawless photo can suggest the metadata was deliberately scrubbed.

Method 4: examine text and spelling

Early AI images were famous for garbled text — signage, packaging and book pages turning into illegible strokes or nonsense spellings.

But this only works in specific settings now. Since 2025, mainstream models render English text quite reliably. Chinese characters and ornate fonts still fail often enough to be a useful signal, but it is now a probabilistic cue rather than a giveaway.

Method 5: test physical consistency

This is the most dependable of the tool-free methods. The logic: generative models learn visual correlation, not physical law. They therefore fail on relationships, not on individual objects.

CheckReal photoAI-generated image
Mirror reflectionsReflection matches the objectReflection unrelated or mismatched
Shadow directionConsistent across objectsShadows of different objects disagree
Contact pointsClear occlusion where object meets groundObject floats or edges interpenetrate
Text on curved surfacesText bends with the surfaceText lies flat and ignores curvature
Pattern continuityTiles, grids and stripes stay coherent in perspectivePatterns break at corners
Distance detailFine detail falls off with distanceDistant detail stays equally sharp or turns to noise

Reflections are the easiest to verify, because they provide a self-reference: what appears in the mirror must correspond to what is outside it. This remains hard for current generation models.

Method 6: trace the source

If the image came from social media or a news post, do two things first:

  1. Reverse image search (Google Lens, Baidu, TinEye). If it first appears in an AI art gallery or a model demo, the question is settled.
  2. Find the original context. Many "viral photos" are actually real images with a false caption — the picture is genuine, the story is not. Tracing tells the two apart.

Why tracing often beats detection

A detector can only say whether an image "looks AI-generated". Tracing answers where it came from. One is an inference, the other is a fact. Where you have the choice, trace first.

Method 7: run it through a detection tool

Submit the image and look at the two signals separately — metadata clues and pixel analysis. Do not just read the 0-100 score.

Old rules that no longer work

These circulated widely in 2023 and are now mostly obsolete. Understanding why they failed matters more than the conclusion.

Combining the methods

Each method fails alone; together they are far more dependable. A sensible order:

  1. Trace first — finding the original is a matter of fact.
  2. Then check metadata — a generator signature or AIGC marker is conclusive.
  3. Then check camera EXIF — a full camera record supports a real photograph.
  4. Only then look at pixels and visual cues — these are inference, and only meaningful on files that have not been re-compressed.

If all you have is a platform-compressed, screenshotted image, you are already in inference territory. The honest answer there is "cannot tell" — not a precise-looking score.

Frequently asked questions

Can you tell if an image is AI-generated?

From a single re-encoded file, often not reliably. Generator output quality in 2026 is close to photography, so visual judgement is weak. The dependable routes are metadata signatures and AIGC markers (conclusive when present) and source tracing — none of which depend on subjective impression.

Why is my AI image detection score so low?

Usually because the image was re-processed. Screenshots, platform transcoding, JPEG re-saving, resizing and retouching all destroy the statistical features pixel-level detection relies on. Another common cause is a generator update the model has not seen. Check the metadata first: a generator signature in the original file settles the question.

Does no EXIF mean the image is AI-generated?

No. Missing EXIF is a neutral fact. Platforms strip metadata in bulk, and screenshots and most editing tools discard EXIF entirely. The meaningful test is the presence of a generator signature or AIGC marker, plus whether physical relationships hold up.

Can Midjourney images be detected?

If you have the unprocessed original export, it usually still carries the generator signature and can be confirmed. Once the image has been screenshotted or transcoded by a platform, that signature is gone and only pixel-level inference remains — with noticeably lower accuracy.

How accurate are AI image detectors?

No single number applies. Accuracy swings with generator version, whether the image was re-processed, and how the test set was built. Products quoting a fixed accuracy figure usually tested only against generators their model had already seen.

Related articles

Check an image yourself

Upload an image and see metadata clues and pixel analysis as separate results

Open the detector