diff --git a/src/server/routers/assignment.ts b/src/server/routers/assignment.ts
index 94b66b5..d8bfd71 100644
--- a/src/server/routers/assignment.ts
+++ b/src/server/routers/assignment.ts
@@ -74,10 +74,22 @@ async function runAIAssignmentJob(jobId: string, roundId: string, userId: string
         description: true,
         tags: true,
         teamName: true,
+        projectTags: {
+          select: { tag: { select: { name: true } }, confidence: true },
+        },
         _count: { select: { assignments: { where: { roundId } } } },
       },
     })
 
+    // Enrich projects with tag confidence data for AI matching
+    const projectsWithConfidence = projects.map((p) => ({
+      ...p,
+      tagConfidences: p.projectTags.map((pt) => ({
+        name: pt.tag.name,
+        confidence: pt.confidence,
+      })),
+    }))
+
     const existingAssignments = await prisma.assignment.findMany({
       where: { roundId },
       select: { userId: true, projectId: true },
@@ -124,7 +136,7 @@ async function runAIAssignmentJob(jobId: string, roundId: string, userId: string
 
     const result = await generateAIAssignments(
       jurors,
-      projects,
+      projectsWithConfidence,
       constraints,
       userId,
       roundId,
diff --git a/src/server/services/ai-assignment.ts b/src/server/services/ai-assignment.ts
index 60bde8c..17e162b 100644
--- a/src/server/services/ai-assignment.ts
+++ b/src/server/services/ai-assignment.ts
@@ -38,7 +38,7 @@ const ASSIGNMENT_SYSTEM_PROMPT = `You are an expert jury assignment optimizer fo
 Match jurors to projects based on expertise alignment, workload balance, and coverage requirements.
 
 ## Matching Criteria (Weighted)
-- Expertise Match (50%): How well juror tags/expertise align with project topics
+- Expertise Match (50%): How well juror tags/expertise align with project topics. Project tags include a confidence score (0-1) — weight higher-confidence tags more heavily as they are more reliably assigned. A tag with confidence 0.9 is a strong signal; one with 0.5 is uncertain.
 - Workload Balance (30%): Distribute assignments evenly; prefer jurors below capacity
 - Minimum Target (20%): Prioritize jurors who haven't reached their minimum assignment count
 
@@ -99,6 +99,7 @@ interface ProjectForAssignment {
   title: string
   description?: string | null
   tags: string[]
+  tagConfidences?: Array<{ name: string; confidence: number }>
   teamName?: string | null
   _count?: {
     assignments: number
@@ -539,7 +540,7 @@ export function generateFallbackAssignments(
 
         return {
           juror,
-          score: calculateExpertiseScore(juror.expertiseTags, project.tags),
+          score: calculateExpertiseScore(juror.expertiseTags, project.tags, project.tagConfidences),
           loadScore: calculateLoadScore(currentLoad, maxLoad),
           underMinBonus: calculateUnderMinBonus(currentLoad, minTarget),
         }
@@ -586,24 +587,44 @@ export function generateFallbackAssignments(
 
 /**
  * Calculate expertise match score based on tag overlap
+ * When tagConfidences are available, weights matches by confidence
  */
 function calculateExpertiseScore(
   jurorTags: string[],
-  projectTags: string[]
+  projectTags: string[],
+  tagConfidences?: Array<{ name: string; confidence: number }>
 ): number {
   if (jurorTags.length === 0 || projectTags.length === 0) {
     return 0.5 // Neutral score if no tags
   }
 
   const jurorTagsLower = new Set(jurorTags.map((t) => t.toLowerCase()))
+
+  // If we have confidence data, use weighted scoring
+  if (tagConfidences && tagConfidences.length > 0) {
+    let weightedMatches = 0
+    let totalWeight = 0
+
+    for (const tc of tagConfidences) {
+      totalWeight += tc.confidence
+      if (jurorTagsLower.has(tc.name.toLowerCase())) {
+        weightedMatches += tc.confidence
+      }
+    }
+
+    if (totalWeight === 0) return 0.5
+
+    const weightedRatio = weightedMatches / totalWeight
+    const hasExpertise = weightedMatches > 0 ? 0.2 : 0
+    return Math.min(1, weightedRatio * 0.8 + hasExpertise)
+  }
+
+  // Fallback: unweighted matching using flat tags
   const matchingTags = projectTags.filter((t) =>
     jurorTagsLower.has(t.toLowerCase())
   )
 
-  // Score based on percentage of project tags matched
   const matchRatio = matchingTags.length / projectTags.length
-
-  // Boost for having expertise, even if not all match
   const hasExpertise = matchingTags.length > 0 ? 0.2 : 0
 
   return Math.min(1, matchRatio * 0.8 + hasExpertise)
diff --git a/src/server/services/anonymization.ts b/src/server/services/anonymization.ts
index 9b25656..0032cb0 100644
--- a/src/server/services/anonymization.ts
+++ b/src/server/services/anonymization.ts
@@ -52,7 +52,7 @@ export interface AnonymizedProject {
   anonymousId: string
   title: string
   description: string | null
-  tags: string[]
+  tags: Array<{ name: string; confidence: number }>
   teamName: string | null
 }
 
@@ -209,6 +209,7 @@ interface ProjectInput {
   title: string
   description?: string | null
   tags: string[]
+  tagConfidences?: Array<{ name: string; confidence: number }>
   teamName?: string | null
 }
 
@@ -253,7 +254,9 @@ export function anonymizeForAI(
         description: project.description
           ? truncateAndSanitize(project.description, DESCRIPTION_LIMITS.ASSIGNMENT)
           : null,
-        tags: project.tags,
+        tags: project.tagConfidences && project.tagConfidences.length > 0
+          ? project.tagConfidences
+          : project.tags.map((t) => ({ name: t, confidence: 1.0 })),
         teamName: project.teamName ? `Team ${index + 1}` : null,
       }
     }
@@ -524,7 +527,7 @@ export function validateAnonymization(data: AnonymizationResult): boolean {
     if (!checkText(project.title)) return false
     if (!checkText(project.description)) return false
     for (const tag of project.tags) {
-      if (!checkText(tag)) return false
+      if (!checkText(typeof tag === 'string' ? tag : tag.name)) return false
     }
   }