Discord deploys machine learning to classify user ages

Discord is rolling out a machine-learning system that estimates whether users are teenagers or adults from account and behavioural signals, allowing more than 90% of its users to avoid submitting identification or completing a facial age check.

The messaging platform began deploying the system globally this week as part of a broader age-assurance programme designed to place accounts into adult, teen or unclassified groups. Adults are defined as users aged 18 or over, while the teen category covers those aged 13 to 17. Discord says its minimum permitted age remains 13.

The rollout is phased, so accounts may receive classifications gradually.

The model examines patterns already associated with an account, including how long it has existed, subscription history, email or phone verification status, aggregated activity, device information and the communities and games connected to its social graph. Discord says it does not analyse message or voice-call content, uploaded media, usernames or profile biographies to estimate age.

A central component is Discord Entity-Relationship Embeddings, or DERE, which converts relationships among users, servers and games into numerical representations. Discord says these social-graph embeddings are the model’s most predictive features because people of similar ages tend to form clusters through friendships, communities and gaming activity. Production embeddings contain 32 dimensions.

The company uses XGBoost, a machine-learning technique based on ensembles of decision trees, to combine those embeddings with account metadata, engagement measures, community context and system information. It has excluded race, gender, ethnicity and other protected demographic attributes from the model’s inputs.

Rather than assigning every account an age category regardless of uncertainty, the system produces a probability that an account belongs to an adult. Scores above an upper confidence threshold lead to an adult classification, while scores below a lower threshold produce a teen classification. Accounts falling between those thresholds remain unclassified and receive additional protections unless their users confirm their age.

Discord also aggregates scores over a longer period instead of acting on a single day’s behaviour, an approach intended to make classifications more stable and harder to manipulate. The company has not published the precise thresholds, but says the model meets or exceeds the effectiveness of age-assurance methods already used in jurisdictions with active requirements.

Users classified as adults retain their normal experience. Teen accounts can continue messaging friends, joining voice calls and participating in non-age-restricted servers, but face stronger safety settings, including restrictions on age-gated content and spaces. Users can see their assigned status through account settings.

Those whom the model cannot classify confidently, new accounts for which Discord lacks sufficient information, or users disputing an assignment can use other age-assurance methods. Depending on region and device, available options can include a video selfie, an identity-document scan, a credit-card check or an age range shared through Apple’s App Store.

Discord says facial age estimation is processed on the user’s device, so the facial scan does not reach Discord or its vendors. For document-based checks, outside providers process the material and return an age signal. Discord says underlying identity documents are deleted after confirmation and the company receives the age result rather than the person’s identity.

The model was trained and evaluated using age-group labels generated through age-assurance deployments during 2025 and 2026. Users in those deployments had confirmed their age through facial estimation or government identification. Discord says the submitted facial or identity data was not supplied to it or used for model training; only the resulting teen-or-adult label was used.

The approach follows a change in Discord’s rollout plans after it delayed wider expansion in February, saying it would add verification choices, increase transparency about vendors and publish technical documentation. Age-assurance requirements have meanwhile expanded across jurisdictions, and Discord has deployed solutions in places including Texas and Brazil.