GAS supports two types of miners that work together in an adversarial loop:
Miners submit classifiers that detect AI-generated content across image, video, and audio modalities. Models are evaluated against diverse benchmark datasets and scored using the sn34_score metric (accuracy + calibration). Each hotkey gets one counted submission (any modality); see Discriminative Mining.
Miners create synthetic media (images and videos) that challenges the discriminators. They generate increasingly realistic content to test and improve detection capabilities, and are rewarded based on validation pass rate and adversarial performance.
Choose your path above to get started with mining on GAS.