Explaining AI-Image Detection: What the Heatmap Actually Shows

AI image detectors are highly vulnerable to technical artifacts: The study found that a detector achieving 99.99% accuracy on one test dropped to 72.54% accuracy when images were re-encoded into a standard format, revealing that the model was relying on compression artifacts rather than actual signs of AI generation.

Explanation maps (heatmaps) may not show what they claim to show: When researchers tested whether the visual explanations provided by detectors actually identified real AI-generation features, many failed basic causal tests—meaning the highlighted regions didn't actually influence the detector's decision, raising questions about the reliability of "evidence" shown to users.

Current public AI detectors have serious consistency problems: Five publicly available detectors showed minimal performance changes (at most 0.07 difference) when tested on re-encoded images, suggesting they may all be exploiting similar technical shortcuts rather than genuinely detecting AI-generated content.

Fixing one problem can hide rather than solve it: When researchers aligned the technical format of real and AI-generated images, the detector simply shifted its reliance to different artifacts instead of learning to detect actual synthesis—the model's accuracy on native files dropped to near-zero (0.0004 probability of detecting AI), indicating the underlying detection capability remained weak.

Marketplace implications are serious: Because AI-generated images can now be created at zero cost, and current detectors are unreliable, platforms approving refunds based on these detections face significant fraud risks.