AI moderation
Score every message against hate, harassment, violence and more, and act on what keyword filters cannot catch.
What it does
AI moderation sends message text to an external moderation model, which scores it against thirteen categories: harassment, hate, sexual content, violence, self-harm and illicit activity, each with a stricter sub-category. When a score crosses the threshold you set, the rule fires through exactly the same pipeline as any other automod rule. It catches the things a word list never will, because it reads meaning rather than matching strings.
Turn it on
Open AI Moderation in your server dashboard and switch Enable AI moderation on. Saving runs a live check against the scoring service first, so a misconfigured deployment fails loudly at that moment instead of looking enabled while silently doing nothing. It is off by default and no message leaves your server until you turn it on.
Pick a sensitivity
Each category has its own sensitivity: how sure the model must be before the rule fires. Very Low (95%) only acts on near-certain violations. Medium (75%) is the default and the setting most communities want. Oppressive (40%) flags anything remotely suspicious and will produce false positives, so pair it with an alert rather than a ban. Custom lets you set an exact value. Every preset and every custom value is free on every plan.
Choose what happens
The rule shares the automod action set: delete the message, alert staff, warn, mute, kick or ban. You can also override the action per category, so hate speech can ban while mild harassment only alerts. Start with alert-only for a week and read the log before you let it act on its own, exactly as you would with a new automod rule.
Scope and exemptions
Channel scoping and role exemptions are ordinary rule conditions, so you can run it only in public channels, or everywhere except a vent channel. Individual users can be exempted too. Staff roles are not exempt automatically; if you want them skipped, exempt the role explicitly.
What is actually sent
Not every message. Anything under twelve characters, bot commands, and messages that are only emoji or mentions are filtered out before a request leaves your server. Repeated identical text, which is most of what a raid looks like, is answered from a short-lived local cache rather than being scored again. Content is not used to train models, and Nighthawk stores only a per-server count of how many checks were made, never the content or who wrote it.
When the service is down
Every failure mode fails open. A timeout, an error or a rate limit means no scores, which means no match, which means the message is allowed. Repeated failures trip a circuit breaker that pauses checks for a minute rather than adding latency to every message. AI moderation going quiet never blocks your chat.
Allowances
Scoring is metered per server per calendar month, and the Usage panel shows where you stand. Running out pauses AI scoring for the rest of the month and nothing else: word and link filters, spam and raid detection, verification and logging all keep running. You get a direct message and a log entry when it happens. See the DeepIntel guide for the numbers.