Rethinking Chat for Enjoyable, Gameplay, and Civility

We’re nonetheless in the beginning with rephrasing and offering extra context to customers when textual content is blocked. Over time, like several AI translation instrument, it should grow to be more and more correct and nuanced in the way it preserves a person’s authentic tone or context. Our precedence stays constructing a secure and civil option to talk on Roblox that continues to prioritize informing, educating, and guiding our customers as they be taught our requirements. 

While rephrasing reduces among the disruption in chat, our multilayered security system stays in impact for extra severe conduct. Rephrasing is obtainable solely for in-experience chat, which is obtainable solely between age-checked customers in comparable age teams and their Trusted Connections. Rephrasings are supported in all languages at present accessible by our automated translation instruments. Our purpose is to realize a degree of accuracy that enables us to cut back the ### till we are able to retire it totally, making digital communication as pure and clear as an in-person dialog.

As we developed this characteristic, we consulted with our Teen Council members to construct a system that displays how teenagers really chat. Council member Sofia advised us, “Moving toward real-time rephrasing preserves the flow of conversation while still helping keep everyone safe. It gently guides the interaction back on track and helps people feel welcome.” 

Teen Council member Jordyn agrees. “This shift teaches players to consider how they communicate and strengthens interaction by fostering friendly connections,” said Jordyn. “By switching to rephrasing, it allows a middle ground between player requests and necessary safety standards.”  

Improved Text Filters: Evolving from Patterns to Context

Our security options regularly evolve to remain true to how customers chat on Roblox. Our textual content filters are skilled to detect and block something that goes towards our Community Standards. That must occur on a world scale and in actual time, as customers are chatting.

Initially, this method had two elements: 1) A rules-based system designed to detect and block content material that breaks our guidelines inside minutes when actual world occasions happen. 2) A set of specialised fashions that had been skilled off of a bigger mannequin, primarily based on a set of simulated and actual samples which might be constantly up to date to incorporate evolving sorts of language. 

These specialised fashions are superb at particular duties, like discovering a listing of phrases we don’t permit, however they’ll’t motive very nicely, so that they aren’t good at discovering phrases which might be evolving. Extremely giant fashions don’t work in actual time, and hallucinations (or miscategorizations) typically improve false positives. 

In this replace, we’ve launched a brand new, third part to our textual content filter system. Now, when the specialised fashions aren’t sure whether or not one thing is breaking the principles or not, they’re despatched to very giant fashions which might be higher at reasoning and assessing extra context. These giant fashions could make choices which might be extra complicated. 

Our experiments present that this mixed method has considerably improved our filters. The filters can now higher detect leet-speak, or letters changed with numbers or symbols, and extra refined makes an attempt to bypass our filters. No system is ever excellent, however we’re excited that this most up-to-date innovation has allowed us to cut back the prevalence of false negatives for sharing or soliciting private info, like social handles or telephone numbers by 20x.

This is a daring new method and we received’t all the time rephrase issues completely. But we’ll proceed studying and experimenting as we transfer ahead and we are going to proceed to cut back the variety of ##### customers are seeing of their chats, with the long-term purpose of retiring them totally. Our final purpose is to empower customers to collaborate and join in a method that’s each enjoyable and respectful.