Both games switched to Amanda while players kept chatting.
During the parallel run, personal information was where Amanda and the previous provider agreed most closely.
More room to chat
Expression improved and blocking dropped, largely from fewer false positives.
A partner that moves fast
Tuning requests were resolved within the day.
Players talk in four places at once: game chat, trade chat, pack chat and Jam-a-Grams, private messages styled like postcards.
Changing moderation provider is a big move for an established kids’ platform. Players know the filter, and they notice quickly when it changes. WildWorks had a setup that worked: twelve years of hand written rules and risk levels with its previous provider. When that provider began winding down its service, WildWorks needed a replacement that matched that protection from day one.
The new setup had to:
- Catch personal information in every language, including names, addresses and phone number formats from outside the US
- Handle filter dodging, from repeated letters to words split across several messages
- Keep contextual pop ups accurate, since a pop up on an innocent message wrongly tells a child they broke a rule
- Give moderators a faster way to follow one conversation across four channels
WildWorks also wanted a partner who would stay engaged after launch, because a migration of this size keeps going long after cutover day.
The two teams ran the migration together and let the data decide when it was ready. WildWorks’ community, trust and safety, and engineering leads met the Aiba team weekly, with a shared Slack channel in between.
Every message in both games went to the previous provider and Amanda in parallel. Amanda was tuned on live production chat: WildWorks’ block lists were imported and refined, game terms and item names were added to the safe list, and Aiba’s data science team adjusted the AI layer for WildWorks’ audience.
WildWorks’ own comparison showed the two systems agreed on roughly nine out of ten messages. Among flagged messages, personal information was where they agreed most closely. Most of the difference came down to approach: hand tuned rules on one side, a context aware AI layer with configurable block lists on the other. Once the data supported it, WildWorks moved both games over.
The teams built a layered approach that keeps false positives away from players, and WildWorks’ moderation tool shows which layer stopped each message. Moderators now see everything two players said to each other in one place, across all channels.
Players saw zero downtime during cutover. In the first week after, contextual pop ups held at the same levels as before, and customer support is fielding fewer complaints. As WildWorks’ Trust and Safety Manager puts it: “There have been significant improvements to expression and an overall reduction in blocking that appears to be largely a removal of false positives.”
Tuning requests were resolved within the day. In the first weeks, the teams tuned gender expression terms, smoking and vaping slang, and in game item names, and improvements done in English automatically carried through to the other languages.
Since launch, Aiba has shipped filter updates that catch repeated letters, spaced out words and split messages. The AI model is retrained monthly, in step with WildWorks’ game updates, so new item names are cleared before each release.
