Calabi Labs · Guide · 2026-06-19
The walt disney company and openai reach landmark agreement to bring b
Disney and OpenAI have reached a landmark agreement that will integrate OpenAI's technology across Disney's content pipelines, with a focus on authentication, verification, and transparency standards that are reshaping how major studios handle AI-generated and AI-assisted media.
What Actually Gets Flagged in AI-Generated Content
When Disney—or any studio—uploads content to platforms like YouTube, Instagram, or TikTok, automated systems scan for invisible signals that betray the file's AI origins. These aren't visible watermarks or logo overlays. They're embedded metadata that lives in the file itself, independent of what you see on screen.
The primary signal is C2PA (Coalition for Content Provenance and Authenticity)—a cryptographic manifest stored as JUMBF (JPEG Universal Metadata Box Format) that acts as a digital "made by AI" certificate. This contains the DigitalSourceType: trainedAlgorithmicMedia XMP tag, which explicitly declares the content was generated using a trained AI model. Beyond that, encoder fingerprints embedded during export—tools like Lavc (Libavcodec) or x264 SEI messages in video bitstreams—leave unmistakable markers that the file was processed through non-consumer software pipelines. Missing GPS coordinates, non-sequential capture timestamps, and generic device identifiers also trigger detection systems.
A single AI export can carry 100-150 metadata tags that platforms use to make their determination. Most creators have no idea these tags exist until their video gets flagged or suppressed.
Why the Obvious Fixes Fail
If you've tried removing a Midjourney or Sora watermark by cropping the image, taking a screenshot, or re-uploading the file through a different platform, you already know: the suppression doesn't stop. Here's why.
Cropping and screenshots remove the visible artifact—the corner logo or sparkle icon—but the metadata layer underneath survives intact. Platforms don't primarily scan what you see; they scan the file's invisible structure. A cropped AI image still carries its C2PA manifest, its XMP AI flags, and its encoder fingerprints.
Re-uploading through another platform sometimes strips display-level metadata (like Instagram compressing a video), but the forensic metadata survives deeper in the file structure. Some platforms even add their own AI detection flags on top, compounding the problem.
The fundamental issue is that these detection signals operate at the file level, not the visual level. Pixel editing and visual cropping address the wrong layer entirely.
How Calabi Actually Handles AI Content at the File Level
Calabi works on the invisible layer that cropping and screenshots never reach. The process runs automatically in one pass:
- Strip: Remove every detectable AI signal from the file—C2PA / Content Credentials JUMBF manifests, XMP tags like DigitalSourceType: trainedAlgorithmicMedia, generator/tool identification metadata, and encoder fingerprints like Lavc and x264 SEI entries.
- Inject: Replace the stripped identity with authentic phone-capture characteristics: real device profiles including iPhone 15/16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra, complete with Make, Model, Software version, GPS coordinates, and capture timestamps.
- Verify: Before download, Calabi generates a forensic proof card using ExifTool—the same forensic tool newsrooms and platform trust—which shows exactly what was stripped and what was injected. You see what platforms will see.
A raw AI export with 144 metadata tags typically comes out as approximately 94 neutral structural tags. The 18 JUMBF/C2PA atoms and 16 C2PA references reduce to 0. The trainedAlgorithmicMedia flag disappears entirely.
Frequently Asked Questions
Does this mean platforms like YouTube and TikTok will stop flagging AI content?
Calabi removes the metadata signals that automated detection systems look for. Results vary by platform and source model, but the invisible layer—C2PA manifests, XMP AI flags, and encoder fingerprints—that automated systems scan for is fully stripped by Calabi's pipeline.
What about visible watermarks like Midjourney's corner logo?
Calabi does not edit pixels or remove visible overlays. If a visible watermark needs to be addressed, cropping removes it—but Calabi handles the invisible metadata layer that survives cropping, which is what platforms actually scan for when they suppress AI content.
Is this only for video, or does it work on images too?
Calabi processes both images and video. The metadata signals—C2PA manifests, XMP tags, encoder fingerprints, and capture metadata—are present in both file types, and Calabi's pipeline strips them regardless of format.
Try Calabi free at calabilabs.com — 10 cleans, no card.