Knowledge Base Article

AI Moderation Admin Settings

AI Settings: General

Settings > AI Settings

Turn the AI Settings toggle ON, then select a Primary LLM model (for content analysis) and a Fallback LLM model (backup if primary is unavailable).

Demo default models (recommended for testing)
For the testing phase, your instance is already configured with recommended defaults:
• Primary LLM: OpenAI GPT 5 Nano
• Fallback LLM: OpenAI GPT 4o Mini
You can change models if needed, but you do not need to change them to run accurate tests.

Moderation: Spam Management & AI Moderation

Settings > Moderation > Spam Management

Enable Spam Management first, it is a prerequisite for AI Moderation. Then enable AI Moderation to activate automated content analysis. If prerequisites are missing, you will see a message explaining what to configure first.

Confidence Threshold

Determines how certain the AI must be before taking automatic action. Higher threshold = more human oversight; lower threshold = more AI automation.

Threshold

Confidence

Best for

Aggressive AI Automation

Below 70%

High-trust communities; willing to accept some false positives

Balanced (Recommended)

~75%

Most communities starting out

Conservative (High Oversight)

Above 80%

Quality-critical, zero-tolerance, or regulated content

If AI confidence exceeds your threshold, it takes automatic action. If at or below, content is flagged for human review. Example at 75%: 80% confidence → auto-rejected; 60% confidence → flagged for human review.

You can override the community-wide threshold for individual boards via Board Settings > Moderation.

Bypass Spam Rejection: Configure which roles skip automatic spam rejection. Administrator, Moderator with Tags, Moderator, and LithiumInternalApiUser are exempt by default.

Community Guidelines: Upload or enter your community's rules. The AI uses these to understand what is and is not acceptable in your specific community. Write detailed and specific guidelines, instead of "no spam," write "no promotional posts, affiliate links, or unsolicited commercial content."

Updated 1 month ago
Version 2.0
No CommentsBe the first to comment