The AI: joke or serious?
A word filter can't tell “you're so dumb 😂” between friends from a real attack. Automod's AI can: it judges each message the way a moderator would, then applies your server's settings.
What the AI looks at
For each message in a module set to AI mode, the AI answers with three pieces of information:
- the category: insult or hate, sexual content, personal data, scam;
- the severity: 1 mild (banter, swearing), 2 clear (direct insult, belittling mockery: looks, love life, family…), 3 severe (racist or homophobic slur, threat, encouraging suicide, harassment, obvious scam);
- the tone: 😂 joke (teasing between friends) or 😠 serious.
It takes context into account: the author's previous messages and the message they reply to. To avoid easy excuses, a joke only counts if there is a real sign of it — “lol”, “lmao”, 😂, 💀… in the message or in the one it replies to. Mean mockery “in a chill tone” stays serious.
Invitations to continue on another network (“insta account in dm”, “add me on snap”, WhatsApp, TikTok…) count as personal data: on a teen server, it's often how a scam or harassment starts. Discord DMs are not concerned.
Sensitivity decides
The AI gives its opinion; the module's sensitivity (0 to 100 %) decides whether to sanction. It can be adjusted on the Pro and Max plans; otherwise the standard sensitivity (50 %) applies. A joke counts one severity level lower:
| Sensitivity | Serious message | Joking message |
|---|---|---|
| 0 % | nothing (everything is a joke) | nothing |
| 1 to 34 % | severe | never |
| 35 to 64 % | clear and severe | severe |
| 65 to 99 % | everything | clear and severe |
| 100 % | everything | everything (jokes no longer exist) |
Changing the sensitivity costs nothing: the instructions sent to the AI are the same whatever the setting.
Who is analysed
- Only members: never the owner, the buyer, administrators or bot admins.
- Not members, roles or channels that are bypassed.
- An identical message is only analysed once every 10 minutes.
- If the AI doesn't answer (outage, limit reached), the classic rules take over: the server is never left unprotected.
Test a message
/iatest (or =iatest) followed by some text shows, without deleting anything, what the rules
and the AI think of it: the tone, the severity and “sanctioned” or “ignored” with the sensitivity needed. Note: a real
message you post yourself as an admin is never analysed — that's on purpose.
Turn the AI on
In /controls → Automatic moderation (or on the dashboard → Modules), set a content module to
AI mode. Two ways to get access:
- the Max plan, which uses Automod's AI;
- your own AI key, with no subscription.
Plug in your own AI key
On the web dashboard → Developer tab, pick your provider (Mistral, OpenAI, Claude, Gemini, Groq, OpenRouter, DeepSeek, xAI or any OpenAI-compatible API), the model, and paste your key. AI modules then work without the Pro or Max plan, without Automod's daily quota, and your provider bills you. The key is encrypted and never shown again.
The Estimated cost card in the same tab works out what it will cost you based on the model and your number of messages per day. Ballpark figures (prices noted in November 2026, with the 4 AI modules):
| Model | Per 1,000 messages |
|---|---|
| Mistral Small | ≈ $0.15 |
| Claude Haiku 5.5 | ≈ $0.23 |
| Claude Opus 5.5 | ≈ $8 (it thinks before answering: overkill for moderation) |
What about my data?
A message's text is only sent to an AI if the server set a module to AI mode. Nothing is sold or used for anything other than moderation. Everything is explained in the privacy policy.