Guide · creator playbook
What platforms actually scan for when detecting AI content
AI detection on social platforms isn't magic. It's a checklist of specific signals. Here's the complete list of what Instagram, TikTok, YouTube, and Twitter check before they decide whether to show your post.
C2PA / Content Credentials
A cryptographically signed metadata block that proves AI origin. Sora, DALL-E, Adobe Firefly, Midjourney's newer versions all attach this. Platforms verify the signature in milliseconds.
Generator tool identifiers in XMP/EXIF
Tool tags like 'CreatorTool: Sora v2' or 'DigitalSourceType: trainedAlgorithmicMedia' baked into the metadata.
Invisible pixel watermarks
Some AI image tools embed invisible pixel-domain patterns that survive JPEG compression, cropping, and screenshots. These are the hardest signal to remove — far harder than metadata, and not always removable.
Encoder fingerprints in the bitstream
FFmpeg's 'Lavc libx264' compressor name and the embedded 'x264 - core 164' SEI NAL — both betray that the file wasn't encoded by a phone.
Missing or inconsistent GPS / timestamps
Real phones always write GPS coordinates and capture dates. AI-generated files often have NO GPS and a creation timestamp matching the moment you exported.
Perceptual hash matching
Platforms compute a 'fingerprint of the visual content' and compare it against known AI tool training data and previously-flagged uploads.
Frequency-domain analysis
Trained models look at the DCT coefficients and frequency distribution — AI-generated content has detectably different patterns than real camera captures.
File container quirks
MP4 atoms like 'udta' and 'meta' in AI tool exports often differ from phone-recorded files in structure. Subtle but checked.
Stop guessing. Calabi automates every step above — 3 free cleans, forensic proof on each.
Try free →
Why this matters now
AI detection on social platforms in 2026 is no longer a coin flip. Instagram, TikTok, YouTube, Twitter/X, and Reddit all run automated scanners on every upload before a human moderator sees it. The flag-rate on raw AI output is approaching 100% for the major generators (Sora, Veo, Kling, Midjourney). The only path forward is full sanitization — strip every signal, inject authentic phone identity, verify before posting.