How the AI Image Authenticity Checker works
The checker looks for evidence from four different places and combines them into an AI signal of Low, Medium or High, with a plain-English reason. It never says an image is "real" or "fake".
1. Metadata
We read EXIF, XMP and ICC data, PNG text chunks and JPEG comments. We look for generator names (for example Midjourney, DALL·E, Stable Diffusion, Adobe Firefly, ComfyUI) and the IPTC "trained algorithmic media" marker. We also report dimensions, colour profile, bit depth, orientation, a JPEG quality estimate and whether camera information is present. Metadata is easy to remove or forge, so a missing or present tag is only one signal.
2. Content Credentials (C2PA)
If the file carries a Content Credentials manifest, we validate its signature and hashes. A valid manifest that declares AI generation sets the signal to High with the reason "Verified provenance says AI-generated". A valid manifest is shown as "Provenance verified". Most images do not carry credentials, and the absence of credentials is not evidence of anything.
3. Basic image forensics
In a background Web Worker we run error level analysis (ELA), per-tile noise variance, per-tile sharpness (Laplacian variance) and an 8x8 block-grid estimate for JPEG double compression. These drive the "Edges and details", "Texture consistency" and "Compression artifacts" rows. They are heuristics: they can flag unusual patterns, but they cannot identify a generator.
4. Optional classifiers
By default no machine-learning classifier runs, and the reason line says so. The site owner can enable an in-browser ONNX model or the optional server check. When detectors disagree, the result is forced to Medium.
5. From evidence to signal
Each piece of evidence has a weight and a direction. The weighted result is a 0 to 100 score: Low is under 35, Medium is 35 to 65 and High is over 65. If the evidence is weak or conflicting, we show Medium with the reason "Not enough evidence" or "Detectors disagree". Lighting and shadow checks need a visual AI model and are shown as "Not checked" rather than guessed.
Frequently asked questions
Can this tool tell me for certain if an image is AI-generated?
No. No tool can. The checker combines several weak and strong signals into an "AI signal" of Low, Medium or High. The result is probabilistic and can be wrong, especially for screenshots, heavily edited images and re-compressed images.
Is my image uploaded to a server?
Not by default. The analysis runs in your browser and the image is never stored on our server. An optional "Enhanced server check" (off by default, and only shown when the site owner has configured it) sends the image for analysis; it is processed in memory and not stored.
What does a Low, Medium or High AI signal mean?
Low means we found little evidence of AI generation. Medium means the evidence is weak, missing or conflicting. High means strong evidence, such as a generator tag in the metadata, valid Content Credentials that declare AI generation, or a detector that reports a high likelihood. None of them is proof either way.
Why does it say "Not enough evidence"?
Most images shared online have had their metadata stripped and have been re-compressed. When there is nothing reliable to measure, the honest answer is Medium with a clear reason, not a guess.
What are Content Credentials (C2PA)?
Content Credentials are a cryptographically signed record of how a file was made and edited. When a valid manifest is present we show "Provenance verified" and the claimed generator. A missing manifest says nothing about whether an image is AI-generated.
What should I do after checking an image?
Run a reverse image search to find earlier or original copies, look for the original source, and consider context. Use the reverse-search buttons in the Source check card.