Quick answer: The Roblox chat filter catches profanity, slurs, phone numbers, and external platform names like Discord. Since March 2026 it uses AI to rephrase blocked words instead of showing #### hashtags. It does not catch grooming language, social engineering, coded bypasses, or any conversation pattern that builds trust and tests boundaries using normal words. A 2026 academic study of roughly two million Roblox messages confirmed that unsafe content passes through the filter at scale.
If your child plays Roblox and you have looked at their chat, you have probably noticed messages full of #### hashtags. That is the filter at work: it catches a word it does not like and blanks it out. Since March 2026, those hashtags are being replaced with AI-generated rephrasing that strips the bad word while keeping the sentence readable.
Both versions of the filter do the same job: they catch profanity. That is genuinely useful. But grooming does not look like profanity. A stranger who builds trust, isolates your child, and pushes the conversation to a private channel will never trip the filter, because every word they use is conversational.
This guide covers exactly how the filter works, what the AI rephrasing changed, what still passes through, and what parents can do about the gap. For the full parental controls setup, see our Roblox parental controls guide. For what parents can and cannot see in chat, see whether parents can see Roblox chat.
How the Roblox chat filter works
The filter runs on every text message sent through Roblox, including in-experience chat, direct messages, and party chat. It processes each message before it reaches other players.
The mechanics
Roblox uses a hybrid system: a combination of word and phrase blocklists, pattern matching, and machine learning classification. The system is provided by CommunitySift (now part of Two Hat). Every game on Roblox must run player text through this filter. Developers cannot weaken it, only add restrictions on top of it.
When the filter flags a word or phrase, it either blocks the word (historically replacing it with hashtags), or since March 2026, rephrases the message using AI (more on this below).
Age-based filtering tiers
The filter is not one-size-fits-all. Roblox assigns age groups based on verified age (facial age estimation has been required globally since January 2026), and the filter strictness varies:
Age group | Text chat default | Filter level |
|---|---|---|
Under 9 | Off by default (parent can enable) | Strictest. Blocks anything the filter cannot confidently classify as safe. |
9 to 12 | On, with strict filtering | Blocks profanity, PII (phone numbers, addresses, school names, schedules, other-platform usernames), and unrecognised text. |
13 and older | On, with lighter filtering | Blocks profanity and slurs. Allows more conversational freedom. |
The under-13 tier also blocks indirect personal identifiers that the 13+ tier allows: school references, repeated mentions of real-world locations, and usernames from external platforms. This is a COPPA compliance measure. For children under 9, Roblox has launched a separate Roblox Kids tier (announced April 2026) with chat off by default and curated content.
The AI rephrasing update (March 2026)
On 5 March 2026, Roblox announced real-time chat rephrasing: instead of replacing banned words with #### hashtags, AI now rephrases the message to remove the violation while keeping the meaning.
The example Roblox gave: "Hurry TF up!" used to become "####". Now it becomes "Hurry up!"
What changed
The system uses a large language model, but Roblox describes it as "highly prescriptive and limited in scope."
It is currently scoped to profanity only.
It works in in-experience chat between age-verified users in similar age groups.
Rephrased messages still count as violations. Users who repeatedly try to send profanity face the same consequences as before.
Other players are notified that a message was rephrased.
Alongside the rephrasing, Roblox upgraded the underlying filter to better detect leetspeak (letter-to-number substitution), with early results showing improved detection rates.
What the rephrasing does not change
The AI rephrasing improves the user experience by making filtered conversations flow naturally instead of showing walls of hashtags. For safety, it solves the wrong problem. It makes profanity filtering smoother, but grooming, social engineering, and manipulation were never caught by the profanity filter in the first place. The rephrasing system does not address those gaps.
What the filter catches
Profanity and slurs: common swear words, racial slurs, ethnic slurs, and known variations.
Personal information (under 13): phone numbers, addresses, surnames, school names, schedules, and usernames from other platforms.
External platform names: "Discord" is explicitly blocked. Other external communication platforms and discord.gg invite links are filtered.
URLs and links: most external links are blocked, including scam and phishing URLs.
Explicit sexual language: known sexual terms and explicit content.
This is a real safety floor. Catching profanity and blocking personal information for young children prevents the most basic harms. The problem is what is left.
What the filter does not catch
This is the section that matters. Every category below has been confirmed in real-world data, including a May 2026 academic study (University of Arizona and Arizona State University) that analysed roughly two million Roblox chat messages and found unsafe content passing through the filter at scale.
Grooming language
Grooming does not use banned words. It uses normal conversation: compliments ("you're really mature for your age"), trust-building ("I won't tell anyone"), boundary-testing ("do you have a boyfriend/girlfriend"), and isolation ("come to my private server, it's just us"). The filter processes each message independently and has no memory of what the same user said previously, so patterns of escalation are invisible to it.
A stranger can build a complete grooming relationship through Roblox chat without triggering the filter once, because every individual message is conversational. For more on how grooming works in games, see our guide to grooming warning signs in gaming.
Luring to other platforms
The filter blocks the word "Discord" in some contexts, but "add me on [platform]" or "let's talk somewhere private" often passes through, especially on 13+ accounts. This is the single most predictive grooming signal: once a conversation moves off Roblox, it leaves all of Roblox's moderation behind. Our guide on what to do when someone asks to move to another app covers this pattern in detail.
Coded language and filter bypasses
Children and adults routinely bypass the filter using:
Letter substitution: replacing characters with numbers or symbols (e.g., "5h1t").
Spacing tricks: adding spaces or periods between letters.
Unicode characters: using visually similar characters from other alphabets.
Community code words: slang that evolves faster than the filter can adapt.
Roblox's March 2026 update improved leetspeak detection, but the filter remains in an ongoing arms race with bypass methods. Active tutorials for bypassing the filter are widely available online.
Context-dependent manipulation
The filter has no memory across messages. It cannot detect:
A conversation that escalates over 20 minutes from friendly to inappropriate.
A user who compliments a child in one message and tests a boundary in the next.
A pattern of gift-giving (Robux, items) followed by requests for personal information.
Each message is evaluated in isolation. The conversation-level pattern is where grooming lives, and the filter does not operate at that level.
False positives (the Scunthorpe problem)
The filter also blocks legitimate words that happen to contain banned substrings. Reported examples include "can't," common greetings, and non-English text. This makes the filter frustrating for normal conversation, which in turn trains children to view it as broken rather than protective.
Sentinel: the conversation-level layer
Roblox does have a system that tries to catch what the filter misses. Sentinel is a separate AI safety layer (active since late 2024) that analyses text chat patterns across conversations, not just individual messages. It captures one-minute snapshots of chat and looks for grooming and child endangerment signals that span multiple messages.
Flagged cases go to expert human analysts, not automated enforcement. In the first half of 2025, Sentinel contributed to roughly 1,200 NCMEC (National Center for Missing and Exploited Children) reports. Thirty-five percent of those cases were proactive, meaning they were caught before any player reported them. Roblox open-sourced Sentinel's methodology in August 2025 for use by other platforms.
Sentinel is a real advancement. But it is a backend investigation tool, not a real-time blocker. It flags conversations for review after they happen. It does not stop a grooming conversation in progress, and parents have no access to anything Sentinel detects.
What parents can do about the gap
The filter catches the surface. It does not catch the conversations that lead to harm. Here is what you can do.
Settings that help
Set text chat to Friends only (never Everyone). This limits who can message your child to their approved friends list.
Review the friends list regularly. Ask about anyone you do not recognise.
Lock the chat settings with a Parental PIN through the Roblox parental controls.
For children under 9, verify that text chat is off (the default on correctly-aged accounts).
Tools that help
Tool | What it does | Platform | Limitation |
|---|---|---|---|
Roblox chat filter | Blocks profanity, PII, and platform names. Rephrases blocked words via AI (March 2026). | All | Word-level only. No grooming detection. No parent notifications. No chat history. |
Roblox Sentinel | Conversation-level AI that flags grooming patterns for human review | All (backend) | Not real-time. Not parent-facing. No notifications to parents. |
Bark | Flags concerning text in Roblox chat on Android only | Android | No voice. Text is one-sided (you do not see what the other person wrote). No iOS coverage for Roblox text. |
Halo | Monitors the voice chat and the text chat in the game being played, and it never records a thing, alerting you only on genuine danger patterns (grooming, bullying, self-harm) | iPhone, iPad, Mac, Windows | No Android yet. No console. On iPhone and iPad covers voice and on-screen text in Roblox. Audio is processed on-device and never uploaded or stored; only the resulting alert is sent to you. 7-day free trial, then US$99.99/yr (about US$8/mo billed annually) or US$12.99/mo. |
The conversation that helps most
No tool replaces talking to your child about what to watch for. The pattern that matters most is simple and teachable: if someone your child met in a game wants to talk somewhere more private, whether that is a private Roblox server, Discord, or Snapchat, that is a red flag worth mentioning to you. The grooming warning signs guide covers the full pattern.
Frequently asked questions
What does the Roblox chat filter block?
The filter blocks profanity, slurs, phone numbers, addresses, and known external platform names like Discord. For under-13 accounts it also blocks additional personal identifiers including school names, schedules, and real-world locations. Since March 2026, blocked profanity is rephrased by AI instead of replaced with hashtags.
Does the Roblox chat filter stop grooming?
No. The filter processes individual messages and catches explicit words. Grooming uses normal conversational language: compliments, trust-building, boundary-testing, and isolation tactics like "come to my private server." None of this triggers a word-level filter. A 2026 academic study of roughly two million Roblox messages confirmed that unsafe content, including grooming patterns, passes through the filter.
What is the Roblox AI chat rephrasing?
Since March 2026, instead of replacing flagged words with #### hashtags, Roblox uses AI to rephrase the message while removing the violation. "Hurry TF up" becomes "Hurry up" instead of "####". It is currently scoped to profanity and works between age-verified users in similar age groups. It does not address grooming, manipulation, or social engineering.
Can children bypass the Roblox chat filter?
Yes. Common bypass methods include letter substitution (replacing characters with numbers or symbols), spacing tricks, Unicode characters from other alphabets, and community-specific code words. Roblox improved leetspeak detection in March 2026, but the filter remains in an ongoing arms race with bypass techniques.
Does the Roblox filter work differently for under 13?
Yes. Under-13 accounts have stricter filtering that blocks personal information (phone numbers, addresses, school names) and anything the filter cannot confidently classify as safe. Under-9 accounts have text chat turned off by default. Accounts aged 13 and older receive lighter filtering that allows more conversational freedom.
Can parents see what the Roblox filter blocked?
No. Parents cannot see what messages were filtered, what their child attempted to send, or what other players said. The parent dashboard shows playtime, friends, and spending, not chat content. For more on what parents can and cannot see, read our guide on whether parents can see Roblox chat.
What is Sentinel on Roblox?
Sentinel is a separate AI safety layer that analyses text chat patterns across conversations, not just individual messages. It looks for grooming and child endangerment signals that the word-level filter misses, and flags cases for human review. In the first half of 2025, Sentinel contributed to roughly 1,200 NCMEC reports. It is a backend investigation tool, not a real-time blocker or a parent-facing feature.
Sources
[Roblox] "Rethinking Chat for Fun Gameplay and Civility." AI rephrasing announcement, 5 March 2026. about.roblox.com
[TechCrunch] "Roblox launches real-time AI chat rephrasing to filter out banned language." March 2026. techcrunch.com
[Notebookcheck] "Roblox now rewrites profane chat in real-time instead of covering it with hashtags." March 2026. notebookcheck.net
[University of Arizona / Arizona State University] Academic study of approximately two million Roblox chat messages (arxiv 2605.04491). May 2026. Found that grooming, threats, and harassment pass through the chat filter at scale.
[Help Net Security] "Roblox chat moderation bypassed in study of ~2 million messages." May 2026. helpnetsecurity.com
[Roblox] "A New Era of Safety: Roblox Requires Age Checks to Access Chat." November 2025 / January 2026. about.roblox.com
[Roblox] "Child Safety at Roblox." Sentinel methodology and NCMEC reporting. about.roblox.com
[NCMEC] "2024 Reports by Electronic Service Providers." Roblox filed 24,522 CyberTipline reports. missingkids.org
[Roblox Support] "Parental Controls FAQ." en.help.roblox.com
[Roblox Developer Hub] "TextService:FilterStringAsync" documentation. create.roblox.com
This guide covers Roblox's chat filter as of August 2026. Roblox updates its moderation systems regularly; verify at roblox.com.





