Most guides on detecting AI images are from 2022 and have aged like dinosaur milk. “AI can’t draw hands.” “AI can’t draw straight lines.” “AI can’t spell words.”

We now live in an age of photorealistic fake media. It is no longer true that AI images have such obvious mistakes.

However, there are still some signs of AI imagery—many are strangely getting worse as the technology advances—but often they’re not errors so much as they’re “conceptual tension.” At a high level, an AI image has several different goals (fulfill the user’s prompt, look coherent/attractive, satisfy a moderation policy, etc.) and if the goals clash (ie, the user prompts for something ugly or incoherent), the image can get subtly pulled in different directions. I show many examples of what, exactly, to look for.

These are my personal heuristics only. There is currently no foolproof way to identify an AI image. Be careful out there.