AutoMod with Artificial Intelligence
Advanced automatic moderation system powered by AI to detect and prevent unwanted behaviors.
ULTRA - All AI features in AutoMod require Server Premium ULTRA
Description​
AI AutoMod uses advanced language models to understand the real context of messages, reducing false positives and detecting sophisticated threats that traditional systems can't detect.
AI Model: OpenAI GPT-4o-mini with Chain of Thought reasoning
AI Features​
1. AI Context Analysis - Word Filter​
What does it do? Analyzes context before penalizing for forbidden words, understanding the difference between malicious and legitimate use.
Features:
- Detects sarcasm and irony
- Understands quotes and educational context
- Distinguishes friendly jokes from real insults
- Reduces false positives by 70-80%
Examples:
| Message | Block? | Reason |
|---|---|---|
| "let's kill this game" | NO | Legitimate gaming context |
| "I'm going to kill you" | YES | Direct threat |
| "killing time playing" | NO | Idiomatic expression |
| "I hate mondays lol" | NO | Harmless humor |
| "I hate [person]" | YES | Hate speech |
How does it work? The AI analyzes every message containing forbidden words and determines if the usage is offensive or harmless. It only blocks when it has 75%+ confidence that it's offensive.
Configuration:
Panel → AutoMod → Word Filter → AI Context Analysis (toggle)
2. AI Duplicate Detection + Paraphrasing​
What does it do? Detects duplicate messages even when they're paraphrased. Prevents sophisticated spam that changes words but maintains the same meaning.
Features:
- Semantic similarity analysis
- Synonym and variation detection
- Configurable threshold (60-95%)
- Message history per user
Detection Examples:
| User sends | Spam detected as |
|---|---|
| "Hello friends" → "Greetings everyone" | Paraphrase detected |
| "Buy here" → "Purchase here" | Minimal variation |
| "Promo link" (x5 with different text) | Promotional spam |
| Natural conversation with others | NOT detected |
Threshold:
- 60%: Very sensitive, may have false positives
- 75%: Recommended for most servers
- 90%: Very strict, only nearly identical duplicates
Configuration:
Panel → AutoMod → Anti-Spam → AI Duplicate Detection
3. AI Toxicity Detection​
What does it do? Detects advanced toxicity: insults, harassment, threats, hate speech, bullying, inappropriate content.
Toxicity Types:
- Insults and offensive language
- Threats and violence
- Harassment and bullying
- Hate speech
- Inappropriate sexual content
- Excessive aggressiveness
Features:
- Multilingual (Spanish and English)
- Confidence score (0-100%)
- Configurable threshold
- Actions: delete, warn, timeout, kick, ban
Recommended Threshold:
- 70%: Balance between precision and sensitivity
- 80%: Stricter, fewer false positives
- 60%: More sensitive, may over-moderate
Configuration:
Panel → AutoMod → AI Moderation → Toxicity Detection
4. AI Scam/Phishing Detection​
What does it do? Detects Nitro scams, crypto scams, phishing, fake giveaways, malicious links.
Detected Scam Types:
- Fake Nitro scams
- Crypto scams and pump & dump
- Fake giveaways
- Phishing links
- Pyramid schemes
- Malicious promotional spam
Features:
- Suspicious URL analysis
- Scam pattern detection
- Confidence score
- Automatic prevention
Recommended Threshold:
- 60%: Recommended (detects most scams)
- 70%: Stricter
- 50%: Very sensitive (may have false positives)
Configuration:
Panel → AutoMod → AI Moderation → Scam Detection
5. AI Raid Detection​
What does it do? Analyzes new member join patterns to detect automated raids.
Raid Indicators:
- Multiple simultaneous joins
- Similar or generated names
- Very new accounts
- Default avatars
- Coordinated behavior
Features:
- AI pattern analysis
- Action recommendations
- Automatic response
- Confidence score
Automatic Actions:
- Kick: Expel suspicious users
- Ban: Ban if high raid confidence
- Alert: Only alert staff
Configuration:
Panel → AutoMod → AI Moderation → Raid Detection
General Configuration​
Enable AI Moderation​
- Go to Panel:
panel.yumechanbot.com→ Your server → AutoMod - Enable AI Moderation: Main AI toggle to ON
- Verify ULTRA tier: Required for all AI features
- Configure each module: Enable only the ones you need
Adjust Thresholds​
What is a threshold? It's the minimum confidence level the AI needs to take action.
- Low (50-60%): Very sensitive, more detections but more false positives
- Medium (70-80%): Recommended, balance between precision and sensitivity
- High (85-95%): Very strict, only very clear cases
Select Actions​
For each detection type you can configure:
- Delete: Delete message only
- Warn: Warn the user
- Timeout: Temporarily mute (1min - 1 week)
- Kick: Expel from server
- Ban: Permanently ban
Best Practices​
Do​
- Start with high thresholds (75-80%) and adjust as needed
- Test in test channels before enabling server-wide
- Review logs regularly to see what the AI detects
- Adjust gradually if you see too many false positives
- Combine with manual moderation for best results
Avoid​
- Don't use thresholds that are too low (causes over-moderation)
- Don't enable everything at once (test module by module)
- Don't trust AI 100% (always review doubtful cases)
- Don't ignore feedback from users about false positives
Recommended Use Cases​
Gaming Server​
- AI Context Word Filter (allows gaming jargon)
- AI Duplicate Detection (prevents promotional spam)
- Toxicity 70% (allows friendly banter)
Educational Server​
- AI Context Word Filter (allows educational quotes)
- Scam Detection (protects students)
- Toxicity 80% (stricter environment)
Community Server​
- All modules enabled
- Balanced thresholds (70-75%)
- Raid Detection active (full protection)
How the AI Works​
Chain of Thought (CoT) Reasoning​
The AI analyzes messages following a step-by-step reasoning process:
- Content Analysis: Reads and understands the message
- Context Extraction: Identifies emotions, intent, tone
- Signal Evaluation: Looks for violation indicators
- Reasoning: Considers alternatives and context
- Final Decision: Determines if there's a violation and confidence score
JSON Schema Validation​
All AI responses use enforced JSON schema for:
- Guaranteed consistent format
- No invalid responses
- Easy automatic processing
Temperature 0​
We use temperature 0 for:
- Maximum consistency in decisions
- Deterministic responses
- Less random variability
Costs​
Each AI analysis costs approximately $0.0001-0.0003 (0.01-0.03 cents). With 10,000 daily messages, the cost would be ~$1-3/month.
Privacy​
- The AI only analyzes the specific reported message
- It does NOT store or save message content
- It does NOT have access to historical messages
- It only returns:
{violation: boolean, confidence: number, type: string}
Support​
If you have issues with AI AutoMod:
- Review logs in the panel
- Verify thresholds and configured actions
- Test in a test channel first
- Contact support on the official Yume-chan server