AI Art Detection

CNN classifier that distinguishes AI-generated images from human-made art; improved accuracy from 66% to 98% through model iterations during COSMOS @ UC San Diego.

Launch demo

Final project from the COSMOS Machine Learning Cluster at UC San Diego. The model classifies whether a given image is AI-generated or human-made — a fast-moving target since the generative landscape shifts every few months.

Starting baseline was 66% accuracy (logistic regression on image statistics). After iterating through KNN, MLPs, and finally a fine-tuned CNN with augmentation, the final model reached 98% accuracy on the held-out test set.

The bigger lesson was about the model-iteration loop itself: each architecture revealed a different failure mode (color statistics fooled the early models; texture features fooled later ones), and the final accuracy came from understanding those failures, not from picking a fancier model.