Roblox tests AI rephrasing to kill off the dreaded '#####' chat

Anyone who has ever played Roblox knows the frustration of a teammate screaming a warning in chat, only for it to show up as a wall of useless "###########".

Roblox is finally fixing its most infamous communication bottleneck. The gaming giant just rolled out an AI-powered real-time chat rephrasing feature that swaps out banned profanity with polite, readable alternatives instead of slapping hashmarks over entire sentences.

Swapping stop signs for steering wheels

Authored by VP Rajiv Bhatia and Chief Safety Officer Matt Kaufman, the update focuses initially on profanity during in-experience text chat. For instance, an urgent "Hurry tf up!" gets automatically cleaned up and delivered as "Hurry up!".

Everyone in the chat will be notified when a message has been rephrased. This keeps tactical information flowing between players without breaking game immersion, while subtly teaching younger audiences what kind of language crosses the line.

To be clear, getting your text rephrased does not give you a free pass. Roblox confirmed that triggered replacements still count as policy violations, and persistent offenders will still face temporary bans or account suspensions.

For now, the feature is rolling out strictly to age-verified users playing with peers in similar age brackets and trusted connections. It supports all languages currently handled by Roblox's automatic translation system.

Reasoning models enter the moderation stack

Under the hood, Roblox also overhauled its underlying text moderation architecture to catch players trying to game the system.

The old setup relied on two layers: a rigid rule-based filter for instant blocks, and specialized lightweight models to catch known banned terms. The problem was that smaller models cannot reason well over complex context, while giant models are typically too slow for real-time multiplayer lobbies.

Roblox introduced a third layer powered by large reasoning models. When smaller models are uncertain whether a message violates community guidelines, the snippet is escalated to the larger model to evaluate context and decipher leet-speak—where letters are swapped for numbers or sneaky punctuation.

According to Roblox, this new hybrid approach reduced false negatives for sharing or soliciting personal information, like phone numbers and social handles, by a massive 20x. The company's stated end goal is to retire the ancient hashmark system altogether.

Why it matters

Replacing aggressive chat redaction with generative rephrasing is one of the smartest product applications of AI in multiplayer gaming we have seen so far.

For years, safety systems in live-service games have treated moderation like a blunt axe. Blocking entire sentences destroys team coordination and frustrates players, often leading to worse toxicity. Guiding users toward acceptable phrasing keeps the game playable while maintaining safety boundaries.

There are real operational hurdles, though. Real-time rephrasing relies heavily on nuance, and if an AI misinterprets sarcasm or tactical slang, it could spark entirely new arguments between teammates. That is almost certainly why Roblox is walling this off behind age verification first.

If Roblox pulls this off at scale, it will put tremendous pressure on competitive gaming ecosystems like Discord, Valve, and Riot to stop relying on blunt keyword blacklists and start treating safety as a conversational layer.