Check an image for signs of AI
An image can be photographed, generated or edited in many ways. This tool brings together available evidence and explains what it supports, without promising certainty about its origin.
How to analyze an image
- Upload the original JPG, PNG or WebP file. Screenshots and recompression can remove evidence.
- Use a still image up to 12 megapixels, between 32 and 8192 pixels per side.
- Start analysis and compare the evidence from each method. Each image uses one allowance unit.
- Download the private report and review its limitations before sharing a conclusion.
Is the percentage a certainty?
No. The three scores show the relative preference of an experimental classifier among generation, manipulation and camera capture. They are not calibrated probabilities of authenticity. Accuracy on User PDF users’ images has not been measured. Results can include false positives and false negatives.
What do C2PA credentials establish?
When present, the system checks the integrity of embedded records and signature trust against a local authority list. A signed declaration of AI use is provenance evidence, not proof that the content is true. Online revocation and remote manifests are not queried. Missing credentials do not establish human origin.
Which invisible watermarks are checked?
The reader looks only for the legacy StableDiffusionV1 signal using the dwtDct method. It can be removed, damaged or inserted by third parties and has no cryptographic authentication. SynthID and proprietary marks from other generators are not checked. No watermark found does not mean no AI was used.
Can generation be distinguished from editing?
Credentials may declare generation or composition with AI. The model also has a manipulation class, but it does not cover every editor or prove that an edit used AI. Weak results and small or transparent images are treated as inconclusive. The tool does not locate edited regions and does not replace forensic examination.
Analysis privacy and free use
Analysis runs on the User PDF server without sending your image to an external AI provider. Files and reports stay private to your session or account and expire. Five free uses per IP/day are shared with the other tools. The report does not publish prompts, GPS or personal metadata.
Is the detector percentage the certainty that an image is AI generated?
No. A score is a model output, not a validated probability of correctness for that image. Detectors can make both kinds of mistakes. No detected signals does not prove human origin, and a high score does not replace checking the source.
Can I remove AI indicators from image metadata?
Remove image metadata deletes descriptive fields that may include creation software, prompts, authorship and location. It creates a copy and a report of removed and retained fields. This does not erase invisible pixel watermarks or statistical patterns and does not make an image undetectable.