Thirty-two numbers derived from an account's place in Discord's social graph will now help decide whether that user is treated as an adult. The vector encodes relationships among users, servers and games, and Discord says it is the most predictive input in its new age-estimation model. According to the company's technical explanation, the system should classify more than 90% of users without asking them for an ID, credit card or video selfie. The convenience comes from making everyday participation part of the age check.
Discord began rolling out the system this week across most of the world, seven months after pausing an earlier expansion that drew privacy objections. The Verge reported that Australia and the UK are excluded from the new model for now. Accounts elsewhere will receive one of three statuses: adult, teen or unconfirmed. A wrong or uncertain result leaves the account open. It can still block age-restricted spaces and settings until the user supplies another age signal.
That makes the model more consequential than a recommendation score. It sits in front of access controls, and the data it reads includes which communities an account joins. Discord's technical post names the model family and its feature groups. It leaves out the figures needed to judge how often teens pass as adults or adults are sent to verification.
Your place in Discord becomes the credential
Discord uses XGBoost, a tree-based machine-learning system, to combine ordinary account fields with graph embeddings. The company's model description lists account tenure, subscription history, email or phone verification, active days, server ownership and message counts among its inputs. Device details such as operating system, browser type and hardware platform also feed the classifier. Message counts are used, while message text, uploaded media, usernames and profile biographies are excluded.
The unusual input is Discord Entity-Relationship Embeddings, or DERE. Discord maps connections among accounts, servers and gaming titles into a 32-dimensional vector for each user. Accounts with similar connection patterns sit near one another in that numerical space. The DERE explanation says age clusters emerge because people of similar ages tend to interact, join related communities and play similar games. The embedding itself is trained on graph structure rather than age labels.
Age enters later. Discord says verified teen-or-adult results collected during regional age-assurance deployments in 2025 and 2026 supplied labels for training and evaluation. The company says it never received the underlying face scans or identity documents, and those materials were not model features. The resulting age group signal was used as the label attached to an account's behavioral and graph data, according to the training-data section.
This design reduces the number of people who must hand over fresh identity evidence. It also gives existing metadata a new purpose. A server's name and description help form the model's community-context features, while social relationships and game connections shape the account embedding. Discord says no single server or friend determines the outcome. Even so, the pattern of communities around an account now affects whether the service grants adult access without another check.
Two thresholds create three outcomes
The model produces a probability between zero and one that an account belongs to an adult. A score above an upper threshold becomes adult, and a score below a lower threshold becomes teen. Scores between those lines remain unconfirmed. Discord says it aggregates scores over a longer window instead of acting on one day's activity, which should prevent a short burst of behavior from changing a classification. The company has not disclosed either threshold or the length of that window in its technical post.
The uncertain middle limits false certainty. Discord expects many unconfirmed accounts to belong to adults because adults make up most of its users, while added protections remain in place until those people confirm their age. New accounts also remain unconfirmed until the system has enough history. The threshold design reduces the chance that sparse data will unlock adult spaces and transfers the cost of uncertainty to the user who wants access.
Discord says the classifier matches or exceeds methods already used under age-assurance laws when measured on a held-out set. It also says the system will be checked against newly confirmed age data and retrained or recalibrated when performance drifts. The performance section does not give the underlying accuracy, false-adult rate, false-teen rate, sample composition or regional breakdown. Readers therefore cannot compare the model with facial estimation or see whether sparse, atypical or recently created accounts fail more often.
A wrong result changes who can reach you
Adult accounts keep their existing experience. Teen accounts send messages from non-friends into a request inbox, warn about some friend requests, blur or block sensitive media, and lose access to age-restricted servers, channels and bot commands. Discord also limits full profile and activity visibility to friends and people in smaller shared servers. Its rollout announcement says unconfirmed users receive most of the same protections, though profile visibility works somewhat differently.
An adult placed in the teen or unconfirmed group can go to User Settings, then Account Status, and choose an available confirmation method. Discord's confirmation guide lists Apple App Store or Google Play age-range sharing, a credit card check, Google Wallet, an on-device video selfie, an ID scan and AgeKey. Availability depends on the user's region and device. Discord app version 343.0 or later is required to see the status on mobile.
Each route leaves a different data trail. Facial estimation runs on the user's device, and Discord says the image never reaches it or vendor k-ID. An ID and matching selfie go to k-ID and are deleted after the result is produced. The credit-card option may place a temporary charge of less than $0.50, while payment processors can retain a truncated financial record for as long as ten years to meet legal obligations. Discord's retention policy says the platform keeps the resulting age signal until the account is deleted or that information is updated.
Discord's design avoids a universal document or face scan. It still makes an age decision for nearly every account. Users who accept the automatic result may never see the inference process. Those caught on the wrong side of a threshold must pick among verification methods with different privacy and access costs. The model decides who encounters that extra step.
The redesign follows a failed first attempt
Discord paused its planned global expansion in February after backlash over face and ID checks. The pause also followed a breach at a former third-party customer-service provider that exposed some user information and scanned photo IDs, The Verge noted. Discord now says it has retired that manual-review process. Video-selfie processing must stay on the device, and vendors receive documents only when a user chooses a method that needs them.
The revised system exchanges a visible demand for proof for an automatic inference drawn from data Discord already holds. That reduces fresh collection for the majority of users. The accuracy question remains open, and the company acknowledges that mistakes will happen. Privacy and fairness now depend on feature selection, thresholds, training labels and the ease of correction as much as on document handling.
The next useful evidence will be a breakdown of adult, teen and unconfirmed classifications, followed by error and appeal rates across regions, account ages and verification methods. Discord has promised to expand age-assurance reporting in future transparency reports. Until those numbers arrive, the 32-dimensional vector explains how the system can avoid most face scans. It cannot tell a user how often the social graph puts an account on the wrong side of the door.