Add microvolunteering task: Review AI-generated images

[edwh]

Overview

Add a new microvolunteering task that allows users to review AI-generated images and rate whether they are a good or bad representation of the item name.

Background

We now have a substantial library of AI-generated images in the ai_images table. While the automated people-detection helps filter out obviously problematic images, we need human review to assess whether images accurately represent the items they're meant to illustrate.

Proposed Implementation

Phase 1: Data Collection

1. New microvolunteering task type: "AI Image Review"

   • Display the AI image alongside the item name
   • Simple Good/Bad buttons for user feedback
   • Store vote counts directly in the ai_images table

2. Database schema change - add columns to ai_images:

   ALTER TABLE ai_images 
     ADD COLUMN good_votes INT UNSIGNED NOT NULL DEFAULT 0,
     ADD COLUMN bad_votes INT UNSIGNED NOT NULL DEFAULT 0;

3. Prevent duplicate votes: Track which images a user has reviewed in existing microvolunteering tracking (or a simple ai_images_reviewers table with just ai_image_id, userid to prevent re-review)

4. Task selection criteria:

   • Prioritize images with fewer total votes
   • Prioritize images used on more messages (higher impact)
   • Don't show the same image twice to the same user

Phase 2: Analysis (after sufficient data collected)

1. Identify problematic images:

   • Images with majority "bad" votes (bad_votes > good_votes)
   • Images with low confidence (close to 50/50 split)

2. Query for flagged images:

   SELECT ai.*, 
          (bad_votes + good_votes) as total_votes,
          bad_votes * 100.0 / NULLIF(bad_votes + good_votes, 0) as bad_percent,
          COUNT(ma.id) as usage_count
   FROM ai_images ai
   LEFT JOIN messages_attachments ma ON ma.externaluid = ai.externaluid
   WHERE bad_votes + good_votes >= 5  -- minimum votes threshold
   GROUP BY ai.id
   HAVING bad_votes > good_votes
   ORDER BY usage_count DESC, bad_percent DESC;

Phase 3: Manual Improvement

1. Admin interface to:

   • View flagged images
   • Trigger regeneration with current prompt
   • Reset vote counts after regeneration

2. Consider: Allow power users to suggest better item name descriptions for regeneration

Acceptance Criteria

• ai_images table has good_votes and bad_votes columns
• Users can rate AI images as good/bad in microvolunteering
• Votes increment the appropriate column
• Users cannot vote on the same image twice
• Report can identify images needing attention based on vote ratios

Related

• AI images are stored in ai_images table
• Images are generated via Pollinations.ai
• People detection already filters images with human figures