AI image detector

Checks a JPEG or PNG for signs of AI generation. First it scans for solid metadata markers: C2PA Content Credentials (JUMBF in JPEG/PNG — the cross-industry standard pushed by Adobe, Microsoft, BBC, NYT, Sony) pointing at a specific model; the EXIF Software field (DALL-E, Midjourney, Firefly, Stable Diffusion, Gemini, Leonardo.Ai, Recraft, Ideogram, NovelAI, ComfyUI, A1111 — all leave their signature); the dedicated IPTC DigitalSourceType marker with values like trainedAlgorithmicMedia or compositeSynthetic that were standardised specifically for AI content; PNG tEXt chunks with parameters/workflow fields where Stable Diffusion and ComfyUI write the full prompt and seed. If no direct markers are found, two heuristics kick in: absence of a camera EXIF block (Make/Model/GPS/Exposure — AI images almost never carry these), and a noise analysis (the σ of pixel-to-neighbour differences in flat regions — on real sensors this is ≥ 1.5 due to photon noise, on diffusion models it often drops below 0.5). The final verdict scales from “AI” (90+ points), “likely AI” (60-89), “uncertain” (30-59), to “likely a real photo” (<30). This is a heuristic, not a verdict: metadata is easy to strip, and retouching or re-saves can mask an AI image. A high score is a reason to verify further (reverse image search, C2PA key validation), not a proof. No vendors, plain JavaScript. Processing happens in the browser, the file never leaves your device.