Rethinking Chat for Fun, Gameplay, and Civility

• A natural, fluid chat is vital for games where teams or communities play together. 

• Strings of ##### where inappropriate language was blocked can disrupt chat, be hard to follow, and interrupt gameplay.

• To help keep chat flowing and respectful, we’re enhancing our chat filters and rephrasing profanity to words that are within our guidelines.

• This is the first step on our long-term path to reducing ##### for a more natural chat experience. We’ll continue to learn, iterate, and evolve our tools to help support safe and civil chat, including additional controls for users.

Every day, millions of people connect and play on Roblox where real-time communication is important for user expression and coordinating gameplay. Using AI filters to block problematic text (visible to users as #####) has always been central to our approach to safety and increasing civility. These hashmarks are generated in real-time and help users understand when something they type is not allowed. Of course, when users repeatedly break the rules, they will face consequences beyond their words being ###.

In May 2025, we introduced proactive, real-time warnings to help users understand our policies. These are now active for both text and voice chat. These notifications tell users when they’ve broken a rule, prompt them to reconsider their language next time, and help them become better digital citizens. In experiments last year, we saw that issuing in-experience text chat notifications and time-outs resulted in a 5% reduction in filtered chat messages and a 6% reduction in consequences from abuse reports.

Today, we’re adding an additional feature to help keep in-experience text chat civil: We’re leveraging AI to automatically rephrase messages, starting with profanity. As these systems scale, they create a flywheel for civility, where real-time feedback helps users learn and adopt our Community Standards. All users, not just the sender of the message, see what language is allowed on Roblox (and the sender sees what is not allowed). This provides more feedback to users who may not know the rules and builds on our foundation of filtering and real-time warnings to help support  safe, civil, and fun conversation on Roblox. 

Real-Time Chat Rephrasing: Moving from Stop Signs to Guidance

For many users, a string of ##### can be a conversational dead end, with no guidance on how to rethink their language to follow our Community Standards. This can often cause confusion and make it harder for teams within a game to collaborate. Imagine someone trying to tell their friend that there’s a monster behind them and typing something like “Watch out! There’s a fkng monster on your ahhh!” to warn them. The friend only sees “###########.” That’s not very helpful. 

Starting today, filtered text will be translated into more respectful language that still feels natural to Roblox users. At launch, this real-time rephrasing is focused on profanity. For example, “Hurry tf up!” might become “Hurry up!”, which is much closer to the original intent than #####. To be clear, when we rephrase a message that violates our profanity policy—it is still a violation and the same rules still apply. A user who continues to try to curse in chat will still face the same consequences for repeatedly violating our policy. 

Think of this system as a steering wheel, rather than a stop sign: If a message violates our policy, everyone in the chat is notified that the text has been rephrased to keep things civil. By matching restricted words to phrases that users actually use in everyday chat, we help keep chat respectful without losing the authentic feel or flow. By replacing these stop signs with real-time guidance, we can reduce some of that friction—and still reinforce our rules against cursing.

We’re still at the beginning with rephrasing and providing more context to users when text is blocked. Over time, like any AI translation tool, it will become increasingly accurate and nuanced in how it preserves a user’s original tone or context. Our priority remains building a safe and civil way to communicate on Roblox that continues to prioritize informing, educating, and guiding our users as they learn our standards. 

While rephrasing reduces some of the disruption in chat, our multilayered safety system remains in effect for more serious behavior. Rephrasing is available exclusively for in-experience chat, which is available only between age-checked users in similar age groups and their Trusted Connections. Rephrasings are supported in all languages currently available through our automatic translation tools. Our goal is to achieve a level of accuracy that allows us to reduce the ### until we can retire it entirely, making digital communication as natural and clear as an in-person conversation.

As we developed this feature, we consulted with our Teen Council members to build a system that reflects how teens actually chat. Council member Sofia told 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 safety features continually evolve to stay true to how users chat on Roblox. Our text filters are trained to detect and block anything that goes against our Community Standards. That needs to happen on a global scale and in real time, as users are chatting.

Initially, this system had two components: 1) A rules-based system designed to detect and block content that breaks our rules within minutes when real world events occur. 2) A set of specialized models that were trained off of a larger model, based on a set of simulated and real samples that are continuously updated to include evolving types of language. 

These specialized models are very good at specific tasks, like finding a list of words we don’t allow, but they can’t reason very well, so they aren’t good at finding terms that are evolving. Extremely large models don’t work in real time, and hallucinations (or miscategorizations) often increase false positives. 

In this update, we’ve introduced a new, third component to our text filter system. Now, when the specialized models aren’t certain whether something is breaking the rules or not, they are sent to very large models that are better at reasoning and assessing more context. These large models can make decisions that are more complex. 

Our experiments show that this combined approach has significantly improved our filters. The filters can now better detect leet-speak, or letters replaced with numbers or symbols, and more sophisticated attempts to bypass our filters. No system is ever perfect, but we’re excited that this most recent innovation has allowed us to reduce the prevalence of false negatives for sharing or soliciting personal information, like social handles or phone numbers by 20x.

This is a bold new approach and we won’t always rephrase things perfectly. But we’ll continue learning and experimenting as we move forward and we will continue to reduce the number of ##### users are seeing in their chats, with the long-term goal of retiring them entirely. Our ultimate goal is to empower users to collaborate and connect in a way that is both fun and respectful.