New
·
Local AI · no OpenAI
The AI Moderator is an intelligent moderation system that automatically analyses messages in your server for harmful behaviour. Messages shorter than three words are skipped, unless they contain a threat, hate against a group, an insult aimed at someone or self-harm. When something is detected, the bot acts immediately according to your settings — issuing a warning, deleting the message, or simply logging it for your moderators.
Unlike most "AI" Discord bots, our detection runs entirely on our own servers. Messages are not forwarded to Google, OpenAI or any other external AI service. Only if a server admin switches on the second opinion when in doubt themselves do the doubtful cases, without a name, go to Anthropic (Claude); that second opinion can only make a penalty milder, never harsher.
Type a test message and watch our local AI score it live — the exact same model that protects your server. One difference: this demo also scores one- and two-word messages, which the moderator skips in Discord.
By default messages are analysed on our own infrastructure: no data to third parties, no tracking, no external APIs. Only if a server admin switches on the second opinion when in doubt themselves do doubtful cases, without a name, go to Anthropic.
Six languages for tone detection: English, Dutch, French, German, Spanish and Italian. The profanity and slur lists exist only for English and Dutch; the threat patterns hold English and Dutch terms and run on every scanned message. The bot's own alerts and commands come in English and Dutch.
Under 100 milliseconds per message. Messages go through a queue one at a time, so during a spike they wait their turn briefly; once 50 messages are queued, a new one is skipped. If the second opinion is on, a doubtful case waits for Claude's answer.
No API costs; the AI Moderator is included in the free bot by default. It is not unlimited: if 50 messages are already queued during a spike, a new one is skipped. The second opinion when in doubt has its own ceiling of 60 questions per server per day.
You set the sensitivity and which actions the bot takes. From mild (serious violations only) to strict (catches light insults too).
Discord's AutoMod works on fixed word lists. The AI Moderator also scores the tone of the whole message with a sentiment model, and has separate patterns for threats.
A hybrid detection system that combines multiple signals into one reliable score.
A BERT classifier that understands the tone of a message in 6 languages (EN, NL, FR, DE, ES, IT). Runs on-device, on our own server.
Hand-built English and Dutch lists of profanity, split into categories (mild, explicit, slurs).
Detects threats like "you should kill yourself", "i'm gonna kill you" or the Dutch "ik ga je vermoorden" — a fixed list of English and Dutch patterns applied to every scanned message. The abbreviation "kys" or a bare "go die" is judged too: a threat counts even in a message shorter than three words. A few spellings ("k.y.s", "kysss") are recognised, most creative spelling variants are not.
Hate speech and identity attacks (racism, homophobia, etc.) are automatically given the highest priority.
All signals are combined into one score between 0 and 1. Action is only taken once it crosses your configured threshold.
| Setting | What it does |
|---|---|
| On / Off | Enables or disables the AI Moderator entirely, per server. |
| Sensitivity: Mild | Serious violations only (threshold 0.85). |
| Sensitivity: Normal RECOMMENDED | Balanced — recommended for most servers (threshold 0.70). |
| Sensitivity: Strict | Catches lighter insults too (threshold 0.55). |
| Alert channel | Optional channel where moderators get notified on every AI action. |
| Auto-delete | Removes messages immediately on high severity. |
| Auto-warn | Gives the user a warning + a chat message (disappears after 10 sec). |
| Second opinion when in doubt | Off by default. If you turn it on, then for messages the moderator is in doubt about, only the text, without a username, goes to Claude from Anthropic for a second opinion. That opinion can only make a penalty milder, never harsher. How that works. |
Go to the SlakBot Dashboard → Moderation → AI Moderator
Flip the switch to ON
Choose your sensitivity (Mild / Normal / Strict)
Optionally set an alert channel for your moderators
Click Save — done! The AI Moderator now protects your server automatically.
Not by default. Moderation runs locally on our own servers; nothing is then shared with Google, OpenAI, Anthropic or any other party. Only if a server admin switches on the second opinion when in doubt themselves do doubtful cases, without a name, go to Anthropic (Claude). That is their own choice, not default behaviour. A mode in which messages went to Google Gemini no longer exists.
Yes. Tone detection works in six languages: English, Dutch, French, German, Spanish and Italian. Profanity and slur lists exist only for English and Dutch, and the threat patterns hold English and Dutch terms — they run on every scanned message, whatever language is detected. The bot itself — its alerts and commands — speaks English and Dutch.
No, the AI Moderator is included in the free bot by default. No API costs. It is not unlimited: messages pass through a queue of at most 50. If it is full, the bot skips a message instead of waiting. The second opinion when in doubt has its own ceiling of 60 questions per server per day; once that is used up, the bot decides locally, just as it would without a second opinion.
Yes. You decide whether messages are deleted, whether the user gets a warning, and whether moderators receive an alert.
On the "Normal" setting the system is balanced. If you do see incorrect detections, you can switch the sensitivity to "Mild".
SlakBot uses AI in four places: the AI moderator runs locally, and in three places we use Claude from Anthropic. Here, in plain words, is what that AI does, how to recognise it and how to keep it off. The exact arrangements about your data are in §2.10 of the privacy policy.
A suggestion from the writing help or from 'Quick setup' always carries the label 'Suggestion made by AI'. There the AI never saves or posts anything by itself: a person reads the suggestion, changes it where needed and presses Save themselves. The AI moderator works differently: it does remove and warn on its own, if your server is set up that way. Such a warning is issued in the name of the AI Moderator, not a person, and the report channel shows whether there was a second opinion.
The writing help does nothing until you press the button: not using it means it's off. The second opinion is off by default; you turn it on and off again in the AI moderator in the new panel. The AI moderator itself is switched on or off there, per server. If you don't want us to use Claude for the answer to your support question, just say so; we then write it without.
Discord allows people from 13, so servers with SlakBot also have young members and sometimes young admins. That is why every AI suggestion says it comes from AI, and the writing help and 'Quick setup' save nothing until a person presses Save. The AI moderator does step in on its own; you can then see it was the AI Moderator, and the second opinion can only make such a penalty milder. SlakBot itself is intended for people aged 16 and over (privacy policy §9).
A suggestion that is wrong or inappropriate, or a decision by the AI moderator that isn't right? Report it via support@slakbot.nl, the support Discord or slakbot.nl/en/contact.
SlakBot follows Anthropic's usage policy and the transparency obligation of the European AI Act: anyone dealing with AI gets to see that it is AI. Anthropic does not train on what SlakBot sends, and SlakBot does not keep your text or the suggestions.
Add SlakBot and turn on the AI Moderator in your dashboard. No credit card, no external AI by default, no hassle.