AI Moderation

AutoMod matches patterns. AI Moderation reads meaning. It looks at messages in context and catches what no keyword list can: veiled threats, harassment spread across sentences, hate speech with zero banned words in it.

They run together. AutoMod handles the mechanical stuff instantly, AI Moderation handles the judgment calls.

Under the hood it works in two steps. A fast classifier checks every message. When a message is clearly fine or clearly a violation, that is the end of it: clear-cut violations are acted on instantly. Borderline messages get a second look from an AI reviewer that reads the recent conversation before deciding, so sarcasm and banter between friends stop tripping false flags, and harassment dressed up in polite words still gets caught.

One dial controls how aggressive the whole system is. Drag the sensitivity bar and every category's threshold moves together. Punishments are tiered: a low-severity match might just delete the message, while a high-severity, high-confidence match can escalate to a timeout or worse. You decide what each tier does.

The AI Server rules trigger lets you write a rule the way you would explain it to a new moderator: plain English, no keywords, no regex. "No self promotion outside #promo." "Keep politics out of general." "Don't ask people to DM you for giveaways." AI reads each message against your rules and applies the action you chose when one is broken.

It is a normal AutoMod rule in every other way: it lives on the AutoMod page, respects your exemptions, and its hits show up in the same logs and stats as every other rule.

AI Moderation acts immediately. A harmful message is gone in seconds, not after a human wakes up. The trade is a one-click Revert on every single action, in the log and on the case page. Undo it, restore what can be restored, and the case records that it was reverted.

Fast when it's right. Cheap to fix when it's wrong.

The AI log is the full history of everything the AI has done, with the category, confidence, and outcome for each entry. Approve or dismiss entries inline as you review. Entries held for review show the action and duration the rule proposed; the flagged message itself is already removed, so approving only applies the member punishment. Approved punishments can be reverted just like automatic ones.

You have three separate monthly balances: Sentiment + Subjects, Moderation, and Chat. Standard includes 30,000, 20,000, and 15,000 tokens respectively. Pro includes 150,000, 30,000, and 75,000. Free includes 500 Chat tokens.

Sentiment and Subjects share one analysis and one charge. Moderation tokens cover AI rules and deeper reviews. Chat tokens cover Zero answers and daily insight generation. Basic screening and rule-based AutoMod do not consume tokens.

Tokens measure paid AI work, so message length, conversation context and the work required change the cost. They are Metryx billing units, separate from raw model tokens. You do not have a fixed message allowance.

Open Settings, AI usage for all three balances, reserved work, recent usage and email alerts. Optional usage billing costs $1 per 1,000 additional tokens, with one shared monthly spend cap across all pools.