Roblox has started rewriting its users’ chat messages in real time using AI, altering what people actually typed into something the platform considers more appropriate.
The feature, rolling out now, goes further than the existing filter that replaces flagged words with “#” symbols. Under the new system, banned language gets silently reworded into what Roblox calls “more respectful language that remains closer to the user’s original intent.”
The platform’s example: type “Hurry TF up!” and the message your recipient sees reads “Hurry up!” Roblox says everyone in the chat is notified when this happens, though the person who typed the original message has no way to stop the substitution before it goes out.
The definition of “banned language” extends beyond profanity. It covers “misspellings, special characters, or other methods to evade detection of profanity,” meaning the AI is also tasked with catching deliberate workarounds and rewriting those too.
Roblox is simultaneously expanding its text filtering system to “detect more variations of language that break its Community Standards,” so the net is getting wider at the same time, and the consequences of being caught in it are changing.
What Roblox has built is a system that goes beyond blocking speech. It replaces it. The message that leaves your keyboard is not the message that arrives. The recipient reads words you didn’t choose, attributed to you, with a notification that your original phrasing was deemed unacceptable. The platform decides what you said.
For now, the feature applies to “in-experience” chats between age-verified users in similar age brackets, and to conversations with “Trusted Connections,” a feature for users 13 and older who’ve completed an age check and connected with people they know. Roblox started requiring age verification for chat features last month. Once verified, users can talk with players in adjacent age groups: the 9–12 bracket can chat with the 13–15 bracket, and so on.
Minecraft filters profanity too, but its approach is more honest about what it’s doing. Flagged words get replaced with symbols, or the message gets blocked. The words you typed don’t reappear as different words under your name. Roblox’s system does something categorically different: it puts words in your mouth.
A system that rewrites what users say in real time, without consent, trained on definitions of “acceptable language” written by a private company, and expanding to catch more variations of speech the platform dislikes, is a significant piece of infrastructure. It starts with profanity. The architecture works the same way for anything else the platform decides to flag.
-----
Article
Archive
