How the ai image scanner works
After you start a scan, the browser downloads and caches a quantized Vision Transformer model, decodes the selected image locally, resizes and normalizes it for the model, and compares its output scores for the REAL and FAKE classes. A dedicated Web Worker runs the inference away from the main interface. The result is a classifier estimate based on learned pixel patterns, not verified authorship or provenance.
verdict = threshold(model AI-class score)Scores of 70% or more are labeled likely AI-generated, 30% or less likely real, and the middle range inconclusive.
Scan a compressed image downloaded from social media
Choose the image and start the scan. If the AI score is 58% and the real-image score is 42%, the tool reports Inconclusive rather than turning a narrow model difference into a confident claim. Compression and resizing may have removed signals the classifier learned during training.
Using the processed image
Keep the exported dimensions, format, transparency, and compression level appropriate for where the image will be used.
const verdict = aiScore >= 0.70
? 'likely-ai'
: aiScore <= 0.30
? 'likely-real'
: 'inconclusive';<p>AI detection is an estimate, not proof of how an image was created.</p>Using the result accurately
Use the original, highest-quality image when possible. Treat the scanner as one signal alongside Content Credentials, source history, reverse-image search, contextual verification, and human review. Never accuse a creator, reject evidence, or make a consequential moderation decision from this score alone.
AI-image detectors can misclassify real photos, human-made digital art, edited images, screenshots, compressed files, and outputs from generators not represented in training. Metadata removal or the absence of Content Credentials does not prove an image is real or synthetic. The displayed scores are model outputs, not calibrated probabilities of authorship.
References: Hugging Face ONNX Community: distilled AI image detection model.