AI Transparency
Where Brilliant uses a learned model, what it is given and what it may decide. The models run on our own machines.
What is and is not AI
| What | Learned model? | Can it punish anyone? |
|---|---|---|
| AutoMod rules | No | Yes, automatically. |
| The threat model | Yes | Only with Aegis. It proposes, and Aegis must confirm. |
| Aegis rules | No | Yes. They check the model, and act alone when it is unsure. |
| System Zero | Yes | No. It is in testing in our own server and only recommends. |
| Modmail | No | No. Only staff answer members. |
The threat model
For raids, staff bursts, webhooks and new bots, a small neural network reads sixty-nine numbers about the event, with no text, usernames or IDs, and leads the decision. Aegis carries out a punishment it proposes only if its own rules also see a threat, and never more harshly. If the model is sure there is no threat, Aegis stands down unless the threat is clear-cut. If the model is unsure, Aegis decides alone. The model leads only while at least 99.8% of the resulting moderations are right on test situations it never trained on.
System Zero
System Zero is in testing. It runs only in Brilliant’s own support server, and there it only recommends. It carries out nothing on its own, and it does not run in any other server.
For security incidents, such as a malicious bot, a webhook backdoor, an app phishing for access, a role quietly given Administrator or security settings turned down, it reads the last fifteen minutes of the audit log as structured facts: what kind of account, bot, app, webhook, role or channel was involved, how old it is, what permissions changed and when.
It also looks for abuse aimed at members and at the server: pornography raids, harassment, threats, doxxing, spam networks of selfbots and farmed accounts, abuse of Brilliant’s commands, and grooming or sextortion. Trash talk between friends is left alone. Slurs and heavy abuse are not trash talk, and count as harassment even once, even when both people are doing it. So does one member going after another again and again without being provoked, and many accounts turning on one member. For these it reads each message once and turns it into a few numbers, such as whether it links to an adult or gore site, whether it looks like someone’s address or phone number, whether it threatens someone or tries to move them off Discord, and how many other accounts posted the same thing. The text is dropped straight away and never stored, and it never looks at images. It also counts joins, reports filed with /report and commands refused for missing permissions. Messages it flags are left out of Brilliant’s message logs, so removing them never copies them into a log channel.
It was trained on simulated servers and on audit log activity from Brilliant’s own support server, reduced to numbers with no names, IDs or message text, and never on any other server. Every incident it finds is posted to the Aegis alert channel with what it recommends, and it does not punish anyone.
Limits
AutoMod and Aegis’s rules are fixed rules that act automatically. Everything described here can be wrong, and by using Brilliant you accept that risk. A server’s staff are responsible for reviewing it. Any claim about an automated decision is subject to the release and the dispute rules in our Terms of Service.