Skip to main content

20,000 toxic CS:GO players banned in six weeks by FACEIT and Google's new chat AI

(Image credit: Valve)

A new AI built to combat toxicity in online gaming has banned 20,000 Counter-Strike: Global Offensive players within its first six weeks, solely by analyzing messages in the game's text chat.

The AI is called Minerva, and it's built by a team at online gaming platform FACEIT—which organised 2018's CS:GO London Major—in collaboration with Google Cloud and Jigsaw, a Google tech incubator. Minerva started examining CS:GO chat messages in late August, and in the first month-and-a-half marked 7,000,000 messages as toxic, issued 90,000 warnings, and banned 20,000 players. 

The AI, trained through machine learning, first issued a warning for verbal abuse if it perceived a toxic message, while also flagging spam messages. Within a few seconds of a match finishing, Minerva sent notifications of either a warning or a ban to the offending player, and punishments grew harsher for repeat offenders.

The number of toxic messages reduced from by 20% between August and September while the AI was in use, and the number of unique players sending toxic messages dropped by 8%. 

The trial started after "months" of eliminating false positives, and it's only the first step in rolling out Minerva to online games. "In-game chat detection is only the first and most simplistic of the applications of Minerva and more of a case study that serves as a first step toward our vision for this AI," FACEIT said in a blog post. "We’re really excited about this foundation as it represents a strong base that will allow us to improve Minerva until we finally detect and address all kinds of abusive behaviors in real-time."

"In the coming weeks we will announce new systems that will support Minerva in her training."

Thanks, PCGamesN.

Samuel is a long-time PC Gamer freelancer who loves RPGs and making long lists of games he'll never have time to play. You can find him on the floor, struggling under the weight of his Steam backlog.