Trend report · hn_ai · 2026-06-14
AI Slop flooding maths YouTube [video]
![AI Slop flooding maths YouTube [video]](/blog/assets/ai-slop-flooding-maths-youtube-video-1cb74e.svg)
If you make AI-generated math videos for YouTube, you've probably noticed the vibe shift. Platforms are getting faster at flagging AI content, and creators using Midjourney animations, Runway exports, or AI voiceovers are hitting upload warnings, reduced reach, or demonetization—even on purely educational content. The detection layer isn't looking at whether your math is right. It's scanning the file itself.
What actually flags your file
Platforms like YouTube, TikTok, and Instagram don't detect AI visuals by analyzing pixels. They scan the metadata and encoding signatures baked into your file during export. Here's what's actually happening in 2026:
C2PA / Content Credentials is the big one. When you export from Midjourney, Sora, Runway, or Leonardo.ai, the file gets tagged with JUMBF (JPEG Universal Metadata Box Format) atoms containing a cryptographic manifest. This manifest explicitly states the model that generated the content, the prompt, and the timestamp of generation. ExifTool reads these as C2PA, ContentCredentials, or dc:creator fields. A raw Sora export carries 18+ distinct JUMBF atoms—all of which read as "made by AI" to platform scanners.
XMP AI flags are the second layer. Adobe, Microsoft, and the C2PA coalition settled on xmp:DigitalSourceType = trainedAlgorithmicMedia as the standard way to flag AI-generated assets. Midjourney exports routinely write this. So do Runway Gen-3 and Pika. If your file has this field, Instagram's classifier treats it as a strong AI signal regardless of visual quality.
Encoder fingerprints are subtler. When you export from an AI video tool, the bitstream carries a specific encoder signature. FFmpeg-based encoders write Lavc (FFmpeg's codec library) into the video stream's SEI (Supplemental Enhancement Information) messages. x264 and x265 write their own vendor strings. These aren't metadata fields—they're embedded in the compressed bitstream itself. Platform scanners parse them as "non-phone capture" signals. A video encoded with Lavc58.134.100 and h264_mediacode values matching FFmpeg looks nothing like a video from an iPhone 16 Pro.
Missing authenticity signals complete the picture. Real phone captures include Make (Apple), Model (iPhone 16 Pro), SoftwareVersion, GPS coordinates, and DateTimeOriginal in the EXIF block. AI exports typically have none of these, or they have placeholder values. YouTube's content ID system and TikTok's AI detection pipeline both treat a file with zero GPS data and no device identity as statistically likely to be synthetic.
How Calabi handles it
Calabi runs a three-stage pipeline against every file you upload. No sliders, no region selection, no manual review.
Stage 1: Strip. The tool removes every AI-detection signal without touching your visual content. C2PA JUMBF manifests are zeroed. The DigitalSourceType: trainedAlgorithmicMedia XMP field is deleted. Encoder fingerprints—Lavc strings, x264 SEI data, HEVC vendor tags—are stripped from the bitstream. The result: a file that no longer carries "AI export" in any standard metadata field.
Stage 2: Inject. Calabi writes authentic phone-capture identity into the file. You choose a device profile—iPhone 16 Pro, Pixel 8 Pro, Galaxy S24 Ultra, or others. The tool writes Make, Model, Software version, GPS coordinates (configurable), and DateTimeOriginal matching the device's encoder profile. The bitstream is re-encoded with the actual encoder that device uses, replacing the FFmpeg fingerprints with a genuine phone encoder signature.
Stage 3: Verify. Before download, you receive a forensic proof card—an ExifTool readout showing exactly what was stripped and what was injected. You see C2PA dropping from 18 atoms to 0, xmp:DigitalSourceType removed, Make changed from "unknown" to "Apple", Model set to "iPhone 16 Pro". This is the same scan platform detectors run. You verify it worked before you touch the file.
Step-by-step
- Upload your AI-generated video to Calabi. Drag-and-drop or click—no account required for the free tier.
- Select a device profile. iPhone 16 Pro, Pixel 8 Pro, or Galaxy S24 Ultra. This determines what identity gets written into your file.
- Calabi runs the strip + inject pipeline automatically. No settings, no toggles. It takes 10–30 seconds depending on file size.
- Review the forensic proof card. ExifTool output showing before/after metadata. Check that C2PA is 0, encoder is a real phone encoder, and device fields are populated.
- Download the cleaned file. Upload it to YouTube, Instagram, or TikTok as normal.
FAQ
What if my AI video has a visible watermark in the corner?
Calabi doesn't erase pixels—it strips the invisible detection layer. For visible logos or watermark text (like Sora's sparkle mark or Runway's corner badge), cropping removes the visible element. Calabi then cleans the metadata so the cropped file doesn't still carry AI detection signals underneath.
Does this guarantee my video won't get flagged?
No tool can guarantee that. Platform classifiers use multiple signals including perceptual hashes, behavioral patterns, and community reports. Calabi removes the metadata and encoder signals that automated scanners check. Results vary by platform and by how the platform evolves its detection methods.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.