A recent paper — “Guess the Unified Model: How Much Can We Recover from Generated Images?” — showed you can train a classifier on AI-generated images and then tell, almost perfectly, which model produced a given image. Nano Banana (Google’s Gemini image model) and ChatGPT’s image generator are both in this paper. Earlier work covered the diffusion models behind tools like Midjourney and Seedream. I think this has real implications for anyone using these tools for client work. But this isn’t “coming soon” — browsers are already moving on it. Google is building C2PA and SynthID detection straight into Chrome, and disclosure rules are landing too, with the EU AI Act’s transparency requirements kicking in this August.

Your (or client’s) eye might be fooled. A trained classifier won’t be.

(I found this paper from my Daily Arxiv podcast, the episode on May 26, 2026 )