Twitch uses a machine learning model combining image analysis and text processing to automatically review custom emote submissions, instantly approving compliant emotes and flagging potential Community Guidelines violations for human review.

Detailed in a June 2022 Twitch Blog post by Applied Scientist Linda Liu, the system uses MobileNetV2 (pre-trained on ImageNet via transfer learning) for image embeddings and a GRU-based character-level model for emote code text embeddings. These are concatenated and fed through dense layers for multi-class violation classification. Training data includes hundreds of thousands of violating emotes and millions of approved emotes tracked since Q1 2020. The model uses LIME (Local Interpretable Model-agnostic Explanations) for interpretability. It automatically approves a large portion of static emotes, reducing specialist workload and enabling instant emote availability for streamers.