config.yml

@@ -40,6 +40,7 @@ ai_moderation:
 # Block all images, GIFs, embeds, and URLs from these users
 # Add Discord user IDs here
 blocked_user_ids:
+    - 552511874269315073
     # Example: - 123456789012345678
 
 # NSFW Video Domain Blocklist (No AI cost)

src/guardden/services/ai/openai_provider.py

@@ -40,25 +40,36 @@ class OpenAIProvider(AIProvider):
         sensitivity: int = 50,
     ) -> ImageAnalysisResult:
         """Analyze an image for NSFW or inappropriate content."""
-        system = """Analyze this image for content moderation. Classify NSFW content with precision.
+        system = """You are a content moderation system for a Discord server. Your job is to analyze images and classify them for NSFW content. This is a legitimate moderation task to protect users.
 
-NSFW CATEGORIES (in order of severity):
-- none: Safe content, no sexual elements
-- suggestive: Revealing clothing, provocative poses, cleavage, but no nudity
-- partial_nudity: Underwear visible, partial exposure, censored nudity
-- nudity: Full nudity visible (breasts, genitals, buttocks) but no sexual acts
+IMPORTANT: You MUST analyze ALL artwork types equally, including:
+- Human characters
+- Anime/manga characters
+- Furry/anthropomorphic characters (animals with human traits)
+- Fantasy creatures
+- Any drawn or digital artwork
+
+Apply the SAME standards to drawn/artwork content as you would to photographs. Artistic style does NOT exempt content from classification.
+
+NSFW CATEGORIES (classify based on what is depicted):
+- none: Fully clothed, no sexual elements, safe for work
+- suggestive: Revealing clothing, provocative poses, emphasis on body parts, tight/skimpy outfits, seductive positioning
+- partial_nudity: Underwear/lingerie visible, bare chest, partially exposed, strategically covered nudity
+- nudity: Exposed breasts, buttocks, or genitals visible (even if stylized/drawn)
 - explicit: Sexual acts, pornographic content, genitalia in sexual context
 
-DETECTION CRITERIA:
-- Focus on human bodies; anime/artwork should be judged by same standards
-- Swimwear in beach/pool context = suggestive at most
-- Artistic nudity (classical art, sculptures) = nudity but note in description
+KEY DETECTION RULES:
+- Leather/latex bodysuits, harnesses, BDSM gear = suggestive or higher
+- Exposed chest (any gender, any species) = partial_nudity or higher
+- Sexualized poses with minimal clothing = suggestive
+- Characters in underwear/lingerie = partial_nudity
+- "Bara", "yiff", or similar adult artwork styles = likely nudity or explicit
 
 ALSO CHECK FOR:
 - Violence or gore (blood, injuries, weapons used violently)
 - Disturbing content (shock imagery, extreme content)
 
-Respond in JSON format:
+Respond ONLY with valid JSON:
 {
     "is_nsfw": true/false,
     "nsfw_category": "none|suggestive|partial_nudity|nudity|explicit",
@@ -66,11 +77,13 @@ Respond in JSON format:
     "is_violent": true/false,
     "is_disturbing": true/false,
     "confidence": 0.0-1.0,
-    "description": "Brief description including context",
+    "description": "Brief factual description of what you see",
     "categories": ["category1"]
 }
 
-NSFW SEVERITY GUIDELINES: none=0, suggestive=20-35, partial_nudity=40-55, nudity=60-75, explicit=80-100"""
+SEVERITY SCALE: none=0, suggestive=25-40, partial_nudity=45-60, nudity=65-80, explicit=85-100
+
+If unsure, err on the side of caution and classify higher rather than lower."""
 
         if sensitivity < 30:
             sensitivity_note = " SENSITIVITY: LENIENT - Allow suggestive content, only flag partial_nudity and above, set is_nsfw=false for suggestive."