Roblox is sharing updated artificial intelligence safety tools designed to detect grooming signals, requests for personal information, and unsafe voice chat involving children. The company announced on August 19 that it is contributing updated versions of three safety models to the Robust Open Online Safety Tools (ROOST) Model Community. Roblox is also releasing a new evaluation dataset to allow other platforms to test their own safety systems.
Roblox Co-founder and CEO Dave Baszucki stated that the company has "no tolerance" for bad actors and builds safety by default. The new AI safety tools focus on identifying attempts to obtain personal information, signs of potential child endangerment, and inappropriate behavior in voice chat.
These AI systems are designed to catch behaviors that could lead to dangerous situations, such as attempts to move conversations to other platforms or solicit personal details. The PII (Personally Identifiable Information) Classifier, for example, looks for attempts to direct players to other apps. Unlike older filters that relied on specific keywords, Version 2.0 of the PII classifier analyzes conversations holistically, evaluating messages together to understand context and intent, even if language is misspelled or coded.
Roblox reports that Version 2.0 expanded language support from 17 to 189 languages and improved its F1 score, a measure of accuracy, from 63.41 to 90.52. The system aims to understand the direction of a conversation rather than judging each message in isolation. This is particularly relevant in cases where initial contact occurs on one platform and then moves to another, potentially with fewer safety protections.
Another AI system, Roblox Sentinel, analyzes conversation patterns to detect early signals of potential child endangerment before interactions become explicitly inappropriate. This allows human reviewers to prioritize and address suspicious interactions. Roblox stated that nearly 70% of the child-endangerment cases it detected in the 12 months ending August 7, 2026, were identified through Sentinel's early detection.
Version 2 of Sentinel offers developers more ways to score suspicious behavior and understand the reasoning behind the system's scores, aiding in platform adaptation and tuning. Additionally, Roblox's voice safety classifier analyzes speech in real time for policy violations, issuing warnings and, for repeated or serious offenses, temporary voice-chat suspensions.
Roblox first made its voice safety classifier open source in 2024. Version 3 now supports 30 languages and eight violation categories, reporting 61% recall across languages at a 1% false-positive rate. However, the company acknowledges that AI moderation is not foolproof and can miss certain conversations.
Independent researchers have previously identified examples of unsafe messages, including those involving grooming, sexualization of minors, bullying, violence, and sensitive information sharing, that bypassed existing moderation systems on Roblox. This highlights the ongoing need for parental involvement.
Roblox previously implemented other AI safety measures, such as a system that analyzes text, avatars, and 3D environments in real time to shut down violating game servers. The company has also expanded age-based protections, including Roblox Kids accounts for ages 5-8 and Roblox Select accounts for ages 9-15, with varying limits on chat and features.
To enhance child safety, Roblox advises parents to treat requests to move conversations to other apps as a warning sign, set clear rules about sharing personal information, review Roblox's parental controls, pay attention to voice chat, and show children how to report or block users. Maintaining open communication with children about their online experiences is also emphasized, encouraging them to speak up if an interaction feels uncomfortable.