[Update] March 26, 2026
Hi Creators,
We are advancing our moderation strategy to moderate user behavior in games without taking down the experience. This upgrade overcomes the limits of traditional filters that can’t detect violations from complex combinations of approved assets - to continuous, real-time, multimodal review across the platform. This approach also enables us to detect and halt rule-breaking behavior as it’s created by users in real-time, such as when creating an object in a UGC experience. By stopping this behavior as users create it, we reduce the possibility of your experience suffering from reputational damage, becoming a place where unsafe situations can develop, or being taken down.
Announcing real time multimodal moderation
We recently introduced a new system that employs a family of AI models to enable real-time, multimodal moderation. This allows us to scan these combined elements simultaneously in the context of an entire scene, including avatars, text, and 3D objects in a specific moment and assesses whether the full scene breaks our rules. If this type of problematic behavior repeatedly occurs in a single game instance, the system will shut down just that server rather than the entire game.
When violative content and behavior is detected repeatedly in the same server, we will automatically shutdown the specific game servers. When a server is shut down for rule violations, users in that experience will see their gameplay blurred. A screen will appear explaining the server shutdown, giving them the choice to either reconnect to a new server or return to the Roblox dashboard.
We’re working to scale this multimodal system to capture and monitor content violating our sexual content and discrimination policies. In the future, we plan to expand to other high severity policies in our Community Standards.
Transparency into Server Shutdowns in Creator Analytics
To give creators better visibility into how many servers are being shut down due to automated moderation, we are adding a new chart to the Creator Safety Dashboard. This enables creators to quickly spot spikes and address potential issues, such as adjusting custom emotes, avatar editing, or in-game user creation features, before a broader community impact.
Note: The categories shown above are associated with the server instance shutdown.
Caption: New chart showing the number of servers shut down per day in a creator’s experience due to the detection of bad user behavior. Spikes could indicate an increase in problematic behavior.
An increase in this number suggests more server shutdowns are occurring within your experience, which could signal that bad actors might be disrupting your experience with behavior that violates our Community Standards. When you have an increase in server shutdowns, this should encourage investigation into potential underlying causes and solutions to address issues before they escalate. We are initially launching with models to shut down servers that detect scenes of “Discrimination” or “Romance or sex” and will continue to expand this set over time.
How to Reduce Server Shutdowns in your experience
Taken together with the recently announced abuse report metrics in Safety Analytics, having visibility into the level of user-submitted abuse reports and moderated activity helps equip Creators to have a better understanding of civility within their experience, and take actions to reduce violative user behavior that can disrupt fun and civil playtime, create unsafe situations, and ultimately put their Experience reputation at risk.
We are actively enhancing the safety toolkit available to developers within their experiences, which currently includes the BanAPI, KickAPI, and IsVerified API. Additionally, we are developing a safety callback API (anticipated for Q2 2026) to provide increased context to developers through warnings or notifications before any server shutdowns occur.
Early Impact and Future Investment
Since launching this multimodal system, automated intervention is rare, and, as of today, we are shutting down only 0.006% of daily server instances that violate our Community Standards. As we train and scale, we’re constantly improving our accuracy and working with the community to minimize false positives. We’re working to scale this multimodal system to capture and monitor 100% of playtime.
We are committed to evolving our moderation technology to address bad actors who try to circumvent our systems. Currently, some moderation actions result in server shutdowns, which unfortunately disrupts the game experience for well-intentioned players. To minimize this disruption, we will increasingly shift our focus from server shutdowns to user level consequences to target individual bad actors. We continuously refine our moderation model’s precision through a human evaluation loop. Reviewers assess shutdown rates at both the Experience and individual user levels. This process allows us to improve the model and identify potentially chronic bad actors for further consequences.
FAQs
How does Roblox prevent potential abuse that could lead to an intentional shutdown, such as ‘griefing’?
-
We actively monitor for “weaponization” daily with alerts for unusual shutdown frequency in single games, and conduct human review for flagged instances.
For violative behaviors that Roblox can attribute such as violative avatars, chat messages, Roblox will take direct user level actions. For violative behaviors that Roblox cannot attribute to the user, we are working on tools to provide to developers to attribute and take user level actions.
Can we view specific users and experiences related to server shutdowns?
- Not at the moment, however, we are exploring ways to share more detail in the future to enable more targeted action on your part.
What are the ‘granular tools’ you are building to help Developers actively moderate their experiences?
- We are working on a system that allows developers to take user-level actions (like individual kicks or bans) rather than relying on full server shutdowns
How do I trust that the AI system is making the right decisions?
- The performance of the AI system is assessed daily through evaluation by human reviewers to determine the system’s false positive rate. Our goal is that server shutdowns only happen when the system’s false positive rate is low.
Why can’t experiences just moderate themselves?
- We encourage everyone to self-moderate and help keep our community safe. Please keep in mind that server shutdowns are rare; this only occurs when we detect serious violations (sexual or discriminatory), and we want to protect other players in the server from violative content.
##### #### ##### ########## ###########