Guide

Can AI images be detected?

Sometimes, with error. A detector is a test built for a particular kind of evidence. Two tests built differently will disagree, including on images a person finds obvious.

17 October 2026 · Check AI Free

The useful form of the question is narrower than “is detection solved?”. Embedded text can show that a file claims to be a Stable Diffusion export. A signed Content Credential can show that a trusted signer attested to a set of actions. A pixel classifier can estimate whether an image resembles the generated images it was trained to recognize. Each of those can fail, and none of them is a property of the scene in front of the lens.

This page explains why AI detectors disagree and how to read a disagreement. It is not a ranking of products, and the numbers below are other researchers’ results, cited so you can open them. They are not this site’s accuracy.

Three tests that get called detection

A parameters chunk, a ComfyUI workflow field or a camera EXIF block is a statement stored in the file. It can be true, stale, forged or missing. Reading it is reliable as a reading of bytes. It is not reliable as a biography of the picture. The file checklist puts that reading first because it is specific when it is present.

Content Credentials add a signature and a binding between a claim and the bytes. That only helps after a validator checks them. This site does not verify signatures. Seeing the letters “C2PA” in a header is an observation of text.

A pixel model looks at the image after much of that text is gone. On this site, the optional check may send a smaller JPEG with metadata removed. A completed score at or below 0.35 is displayed as likely camera-origin, and at or above 0.70 as likely AI-generated. The middle is inconclusive. The method page states that these are display thresholds, not a validated calibration. We do not publish an overall accuracy figure.

Why the numbers diverge

Published detectors are trained on particular generators, compressions and label definitions. A model that learned an older GAN artifact can miss a newer diffusion image, and a model tuned to one lab’s JPEG settings can move when a messenger recompresses the file. Thresholds differ too: the same internal score can be shown as a percentage, a label or an “inconclusive” band.

Ren, Zhou and colleagues report a zero-shot comparison of open detection methods in arXiv:2602.07814. Their abstract says there was no universal winner. Across the detectors they summarized, mean accuracy ran from 37.5% to 75.0%, and rankings changed sharply from one dataset to another. They also report that images from several recent commercial generators were detected at only 18–30% average accuracy in that study. Those figures describe their benchmark. They are not a forecast for the next file you upload here.

Li and colleagues, in the AIGIBench preprint arXiv:2505.12335, start from the opposite publicity: detectors often report high accuracy in controlled settings. Their abstract says the detectors they evaluated suffered significant performance drops on real-world data. Together, the two papers support a plain conclusion. A leaderboard from one dataset does not travel. Disagreement is an expected result when the generators, the compression and the decision threshold change.

A watermark such as SynthID is a third family again. It is evidence only for images that passed through a system that embedded it, and only for a reader that knows how to look. This site has no SynthID decoder. A general classifier will not reliably “see” a watermark it was not built to read, and a missing watermark is the default for cameras and for many generators.

What to do when two tools disagree

1. Step 1

   Put the two results side by side

   Record the product, the date, the file you submitted and the exact label. “AI” from one tool and “uncertain” from another are not the same disagreement as “AI” and “camera.” Include any statement about whether a model actually ran.

2. Step 2

   Check that both tools saw the same bytes

   A screenshot, a social download and the original export are different inputs. If one of them was stripped or recompressed, compare that history with the metadata guide before you treat the gap as a conflict about the scene.

3. Step 3

   Separate a stored claim from a pixel score

   If one tool is showing a PNG parameter block and the other is showing a probability, they answered different questions. Keep both. Do not average them into a single percentage.

4. Step 4

   Look for an earlier publication

   Reverse image search can show that the picture is older than the caption, which settles a different dispute than “were the pixels generated?”.

5. Step 5

   Match the generator only when the file says so

   Style is not an identifier. Use the Midjourney, Flux, DALL·E or Stable Diffusion page for the context you were given, and keep the conclusion at “a general AI-image estimate,” unless a field in the file names a tool.

6. Step 6

   Prefer the inconclusive reading when the decision is costly

   Hiring, news, school discipline and criminal allegations need the original file and a source that can explain it. A disputed score is a reason to ask for that material. It is a poor reason to announce a verdict.

A practical answer

AI images can be detected when the file still holds a specific claim, when a signature validates, or when a pixel model happens to have learned artifacts that this image still carries. They can also pass every one of those tests. Newer generators, heavy recompression and ordinary unusual photographs are routine failure cases in the research above. Use the checker to collect the evidence it actually returns, then keep the uncertainty in the sentence you tell someone else.