Twitch will use machine learning to detect people evading bans

The image is by Alex Castro.

A new tool that uses machine learning to detect people who may be attempting to evade bans is being used by twitch to reduce harassment. It is the company's latest addition to combat streamers' chats being overrun with troll messages.

Suspicious User Detection can identify users who have evaded bans from a streamer. The tool uses machine learning to identify potential evaders by analyzing their behavior and characteristics, and compares that information against accounts that have been banned from a streamer's channel.

Messages from likely evaders won't be sent to chat, but streamers and their mod can see them. Streamers and mod can choose to monitor a likely ban evader, which adds that user to a monitoring list and puts a message next to a user's name, or ban them. The messages of possible evaders can be seen in chat, but streamers can choose to have those messages blocked from chat as well.

Suspicious User Detection will be turned on by default, but streamers can turn it off if they want. Streamers and mod can manually monitor users who are suspicious.

The tool was inspired by feedback from the community about the need for better ways to curb ban evaders. It can be hard to distinguish between a user who chats something that violates their channel's norm and a newer viewer who hasn't learned that channel's customs yet. We designed this tool to give creators more information about potential ban evaders so they could make more efficient and informed decisions.

If Suspicious User Detection is used in conjunction with the recently introduced controls that let a streamer require phone or email verification for accounts participating in chat, it could make a difference. It's not known how effective Suspicious User Detection will be in practice or if ban evaders can find ways to get around it.