GPTTrace: Private AI Detector

Is that photo real? Was that paragraph written by a chatbot? Right-click it and find out, without handing it to anyone. GPTTrace adds two items to Chrome's right-click menu: • Check image with GPTTrace: on any image on a web page • Check text with GPTTrace: on any text you select A small window shows the verdict, an estimated chance of AI, a confidence level and the evidence behind it: which signals pushed the score toward AI and which toward human or authentic. The extension runs the same detection engine as gpttrace.com, so you get the same verdict the site gives for the same input. "Full report on gpttrace.com" opens the site's detector for the views that need more room. PRIVATE BY DESIGN The check runs on your computer. The image or text you check is never uploaded: not to GPTTrace, not to anyone. There is no account and no analytics in the extension. The only downloads are the detection models, fetched once from gpttrace.com and cached. IMAGES • Content Credentials (C2PA), IPTC "made with generative AI" tags and embedded generator settings, which settle the question when present • Camera metadata and typical generator image sizes • Two neural AI-image classifiers (the Community Forensics ViT, trained on images from 4,803 generators, and our fine-tune of it on newer generators; 46 MB together, downloaded once) • Pixel forensics: error-level analysis and frequency-spectrum checks On our labelled test set of 635 images the image detector scores a cross-validated AUC of 0.98, and catches 84% of AI images while wrongly flagging about 5% of real photos. TEXT • The standard check looks for the patterns language models over-use, sentence rhythm and vocabulary statistics, and compression similarity to chatbot prose. It is instant and needs no download. • The optional deep scan adds GPTTrace's own neural text model (Ettin-150M fine-tuned on the RAID and MAGE benchmarks). It is a 144 MB download, made only when you choose to run it, then cached. On our test set of 2,467 passages, cross-validated, the standard check scores an AUC of 0.80 and the deep scan 0.96. Both are tuned so that around 5% or fewer of human texts cross the "Likely AI" line. LIMITS A detector gives evidence, not proof. Short texts (under about 150 words) are unreliable for every detector. Screenshots and social-media re-uploads strip metadata, and no detector catches every generator. Never use a score alone to accuse someone. How it is measured: https://gpttrace.com/methodology/ Privacy: https://gpttrace.com/privacy/