How to tell if an image is AI-generated
Judging an image by eye — counting fingers, hunting for garbled text — is now unreliable in both directions. A more defensible approach is to start from the original file and read the provenance evidence it still carries. AICheck365 inspects C2PA Content Credentials, EXIF/XMP fields, PNG text chunks holding generation parameters or a ComfyUI workflow, and the file name, all inside your browser. It separates cryptographically verifiable signatures from metadata anyone can edit.
Short answer: weigh layered evidence, not a single percentage
The strongest case is a valid C2PA signature whose manifest declares digitalSourceType = trainedAlgorithmicMedia. Unsigned EXIF/XMP tool fields and PNG prompt or workflow text support that reading but can be edited. File names and visual oddities are weak hints only. Finding no signals means this copy carries no readable evidence — it does not prove a human made the image.
Real samples you can reproduce on this site
| Adobe Firefly JPEG | A public sample from the Content Authenticity Initiative. digitalSourceType = trainedAlgorithmicMedia is readable and the C2PA signature is cryptographically valid, but the signing certificate is not in the current trust list — so the result reads valid rather than trusted. |
|---|---|
| ComfyUI / Stable Diffusion XL PNG | An original from the official ComfyUI examples. The PNG still holds prompt and workflow text chunks alongside unsigned generation parameters. |
| JPEG with no provenance metadata | Returns "no readable evidence". This control sample shows why an empty result cannot separate an ordinary photo from an AI image whose metadata was stripped. |
Signals to Check
| C2PA Content Credentials | The manifest, signature state, trust state, digitalSourceType, claim generator and actions are reported separately. A valid signature that declares AI generation is the strongest evidence available. |
|---|---|
| EXIF / XMP | Software, CreatorTool, Credit, DigitalSourceType and UserComment sometimes retain a platform name or an AIGC label. Unsigned fields can be edited or copied. |
| PNG text / workflow | Stable Diffusion, AUTOMATIC1111, ComfyUI and NovelAI may write prompt, negative prompt, seed, sampler, CFG, model name and a full workflow JSON. |
| File name and visual context | A platform prefix or an odd hand is worth noting, but both are trivially changed or misread, so neither stands on its own. |
Why visual inspection misfires in both directions
Current models have closed most of the early tells, so a fully synthetic image can look unremarkable. The reverse error is just as common: a real photograph pushed through heavy compression or motion blur produces smeared fingers and warped lines that read as "AI". The traits below are no longer conclusive on their own.
- Hands, teeth, jewellery and hair: recent models render these convincingly, and compression damages them in genuine photos.
- Text inside the image: garbled lettering was once a reliable tell, but text rendering has improved quickly.
- Lighting and reflections: judging physical consistency by eye is unstable and frequently mislabels ordinary photographs.
An evidence hierarchy for conflicting clues
When signals disagree, rank them by how hard each is to forge and trust the most robust layer first:
- 1. A valid C2PA signature declaring AI generation — the strongest and independently verifiable evidence.
- 2. Unsigned metadata (EXIF/XMP, PNG prompt or workflow) — informative, but editable.
- 3. The file name, such as a platform prefix — weak context, trivially changed.
- 4. Visual intuition — a starting point for investigation, never a conclusion.
Can you check an image saved from social media?
Often only partially. WhatsApp, Instagram, Facebook, X and screenshots re-encode the image, and that usually strips C2PA, XMP and PNG text chunks along the way. An empty result therefore says nothing about whether the image is synthetic. The practical move is to go back a step: ask whoever shared it for the file as it came out of the tool, or find the highest-resolution original that was published. The closer you get to that first copy, the more readable evidence survives.
Checking an image step by step
- Keep the original file — Use the file downloaded from the platform or sent by whoever generated it — not a screenshot or a copy pulled from social media.
- Run the analysis locally — Load the image into AICheck365 and read the C2PA, EXIF/XMP, PNG text and file-name summary. Everything happens in your browser; the file is never uploaded.
- Confirm the C2PA is actually valid — The presence of C2PA is not the finding. Check whether the signature validates, whether tampering errors appear, and whether the manifest genuinely declares AI generation.
- Cross-check the supporting fields — See whether tool, model, prompt and workflow agree with each other and with where the file came from. Unsigned metadata can be fabricated.
- Record provenance and limits — Note the source URL, the date, whether the file was re-exported, and the result. Leave visual oddities as a closing note rather than the verdict.
Limits
- Messaging apps, social platforms, image optimisers and screenshots routinely strip C2PA, XMP and PNG text.
- Unsigned EXIF/XMP fields and workflow JSON can be edited or copied between files — treat them as clues, not proof.
- The presence of C2PA does not by itself mean AI: you have to read what the manifest declares (digitalSourceType and actions).
- No signals found means this copy preserves no readable evidence; it does not rule out AI generation.
Start with the full detector to gather the evidence, then open C2PA or PNG workflow findings in the dedicated tool.
FAQ
Is there a free AI image detector?
Yes. AICheck365 reads C2PA, EXIF/XMP and PNG generation parameters for free, in your browser, with no account and no upload. It reports the signals actually present in the file rather than a bare verdict label.
Can any tool be 100% accurate?
No, and any tool claiming otherwise is overselling. What can be established is which provenance signals this file still carries. When the evidence has been stripped, the honest answer is that the file no longer supports a conclusion.
What does "no signals found" mean?
That this copy carries no readable provenance evidence. The metadata may never have been written, or it may have been removed in transit. It is not evidence that a human made the image.
Does C2PA always mean the image is AI?
No. C2PA records provenance in general, including for ordinary camera photos. You have to read whether the manifest declares an AI source through digitalSourceType or a generative claim generator.
Are my files uploaded anywhere?
No. Parsing runs locally in your browser through Rust compiled to WebAssembly, so the image never leaves your device.
Sources and verification records
- AICheck365 public sample records — Sample provenance, test dates, observed signals and known limits.
- Content Authenticity Initiative example assets — Public source of the Firefly C2PA sample.
- Official ComfyUI examples — Source of the SDXL PNG carrying a workflow.
- C2PA Technical Specification 2.2 — Official specification for manifests, assertions and signature verification.
- AICheck365 methodology — How evidence strength and the absence of signals are interpreted.