Problem
Text displayed on a screen inside a video is classified as metadata-like content even though the creator is merely showing a real document for analysis.
Solution
Root Cause / Diagnostic:
Optical Character Recognition (OCR) systems extract text from captured computer screens, physical documents, and signage within the video frames. The moderation engine evaluates this extracted text against blacklists, mistaking analyzed documentary evidence for spam or policy-violating text.
Actionable Fix:
1. Highlight Documentary Document Analysis: Explain in the appeal that the on-screen text is an authentic primary source document subjected to critical journalistic evaluation.
2. Mask or Obscure Non-Essential Sensitive Phrases: In future edits, redact phone numbers, URLs, offensive phrases, or sensitive terms inside documents using digital black bars or blur.
3. Pair On-Screen Documents with Strong Voiceover Context: Ensure voiceover immediately deconstructs and refutes any questionable on-screen assertions to prevent classification as endorsement.
Pro Tip:
When showing documents or social media posts on screen, apply a 20% transparent overlay or slight gaussian blur over non-essential text so OCR models cannot extract raw strings.
Optical Character Recognition (OCR) systems extract text from captured computer screens, physical documents, and signage within the video frames. The moderation engine evaluates this extracted text against blacklists, mistaking analyzed documentary evidence for spam or policy-violating text.
Actionable Fix:
1. Highlight Documentary Document Analysis: Explain in the appeal that the on-screen text is an authentic primary source document subjected to critical journalistic evaluation.
2. Mask or Obscure Non-Essential Sensitive Phrases: In future edits, redact phone numbers, URLs, offensive phrases, or sensitive terms inside documents using digital black bars or blur.
3. Pair On-Screen Documents with Strong Voiceover Context: Ensure voiceover immediately deconstructs and refutes any questionable on-screen assertions to prevent classification as endorsement.
Pro Tip:
When showing documents or social media posts on screen, apply a 20% transparent overlay or slight gaussian blur over non-essential text so OCR models cannot extract raw strings.