If it is a P-image, AI can actually detect most of it, but if it is an AI retouched image, whether it can be caught or not is uncertain .. I spent 10 minutes Vibe a P-image detector and deployed it directly, and the effect was actually quite good! I took several normal images that I had doubts about for detection, and the accuracy was pretty good (I need to use my own API). The prompt used in it is also publicly available in the comments and can be directly detected on ChatGPT or Gemini app. The traditional manual retouching (Photoshop/Meitu Xiuxiu) logic is "collage and overlay". No matter how skilled the technique is, the modified area will still have a slight trace on the data level. Nowadays, AI models can do some subtle detection in this part. But when it comes to generative AI, it is basically undetectable because the logic of generative AI is "in painting", not pasting words on it, but understanding light, shadow, and texture to "regrow" the image at the pixel level. The result is a perfect fusion of noise and coherent data features, which looks like an untouched original image to current AI. That's also why the industry is now promoting standards such as C2PA (Digital Content Authentication) or Google's SynthID (Invisible Watermark) for anti-counterfeiting. However, I'm still waiting on the waitlist for SynthID. When it's really released, I believe it will be a very important part of verifying the authenticity of content in the future. However, AI detection is still very powerful for common poor P-images on the market. AI is making the world increasingly untrustworthy, and what is urgently needed is AI that can counter AI