Calabi Labs · Guide · 2026-06-15
Jorja smiths label requests share of royalties from ai cloned tiktok v
Here's the clean HTML page:
```html
Jorja Smith's Label Is Seeking Royalties From the AI-Cloned Song That Went Viral on TikTok
Jorja Smith's record label, FAMM, has filed a legal claim against British dance act Haven, demanding a share of royalties from the viral track "I Run" — which the label says was made using AI to clone the Grammy-nominated singer's voice without her consent. The case, reported in December 2025, is being closely watched across the music and AI industries as a test of how rights are handled when an artist's voice is reconstructed by a machine.
The song racked up millions of streams after spreading on TikTok in October 2025, with listeners quickly noting that the vocals sounded unmistakably like Jorja Smith. FAMM says no license was sought, no permission was given, and the label is now pursuing a share of all royalty earnings. Haven has not publicly responded to the claims.
What Actually Gets an AI Music File Flagged
When a track like "I Run" is uploaded to TikTok, Instagram, YouTube, or Spotify, the platform runs it through automated detection systems that don't just listen — they read the file's invisible infrastructure. Here's what those systems are actually scanning:
- C2PA / Content Credentials: This is a cryptographic manifest embedded in compatible media files. It stores a "made by AI" provenance chain — who trained the model, what tools generated the output, and when. If Haven used a text-to-music or voice-cloning tool that supports C2PA, that manifest travels with the file and flags it immediately.
- XMP AI metadata flags: Fields like
DigitalSourceType: trainedAlgorithmicMediaget written into a file's XMP metadata when certain AI generation tools export. This is a direct, machine-readable signal that the audio did not originate from a physical recording. - Encoder fingerprints: AI audio generators and some export pipelines use specific encoders — names like Lavc, FFmpeg, or particular x264/x265 SEI markers in video — that forensic tools can identify. A real iPhone or Pixel recording has a different encoder signature than a track rendered through an AI pipeline.
- Missing device metadata: A genuine phone recording carries a Make, Model, Software version, GPS coordinates, and a capture timestamp. AI-generated files typically lack all of this. That absence is itself a red flag.
- Perceptual hashes and acoustic fingerprints: Platforms can compare the audio fingerprint against databases of known AI-generated vocal models. If the vocal signature matches a cloned Jorja Smith model, that gets logged.
The "I Run" case reportedly attracted attention partly because listeners on TikTok spotted the vocal similarity — but the platform's own automated scanning would have picked up on the metadata layer independently, had the track been flagged through that route.
Why the Obvious Fixes Don't Actually Work
If you're an AI creator who has made a track and now worries about it being challenged — or if you're trying to distribute AI-generated audio without running into Content ID or platform takedowns — you might instinctively try some common workarounds. Here's why they fail:
- Re-recording or re-uploading: If you take an AI-generated audio file and play it through your speakers, record it with your phone, and re-upload it, you've changed the encoder fingerprint but the C2PA and XMP metadata may still be embedded underneath. Platforms have gotten better at looking through the surface layer.
- Cropping or trimming: Cropping works for visible watermarks — a corner logo — but it has zero effect on invisible metadata. The C2PA manifest and XMP tags survive most trimming operations unless the file is fully re-encoded from scratch.
- Metadata strippers: Basic EXIF/XMP strippers remove some visible tags, but they often miss C2PA atoms (stored as JUMBF boxes in MP4/MOV files), proprietary AI generator tags, and encoder fingerprint data. A forensic scan with ExifTool will still expose the full generation history.
- Converting to a different format: Converting MP4 to WAV or moving from one container format to another does not guarantee that AI-generation metadata is removed. Some flags persist through transcoding.
The underlying problem is that the detection layer is invisible to you as you look at the file — and most consumer-friendly "metadata cleaning" tools don't target the right fields. You think the file is clean because it looks normal, but the forensic scan a platform runs sees a different story.
How to Actually Clean an AI-Generated Audio or Video File
If you're creating AI audio or video content and want it to present as a normal capture — or if you want to understand what a proper cleaning tool actually does — here's the real process:
- Strip the AI signals: Remove C2PA / Content Credentials JUMBF manifests, XMP
DigitalSourceType: trainedAlgorithmicMediaflags, generator and tool tags, and encoder fingerprints (Lavc, x264 SEI markers) that identify AI export pipelines. The goal is to reduce an AI export's 100+ metadata tags down to a neutral structural set. - Inject authentic device identity: Write in real phone capture metadata — a real device Make and Model (e.g., iPhone 15 Pro, Pixel 8 Pro), a software version, GPS coordinates, and a capture timestamp. This replaces the absence of device data that would otherwise flag the file.
- Verify with a forensic proof card: Before distributing the file, run a scan using ExifTool — the same forensic tool newsrooms and platform trust teams use — to confirm exactly what was stripped and what was injected. You want to see 0 C2PA atoms, 0 references to trainedAlgorithmicMedia, and a clean device profile.
That's the pipeline Calabi runs: a one-pass process that handles all three stages automatically, so the file that comes out reads, at the file level, as a normal phone recording. No manual editing, no guessing which fields to remove.
FAQ
Does this mean the "I Run" track should have been caught and removed?
Potentially, yes — but platform enforcement varies. TikTok and YouTube have automated AI-content detection, but the speed and accuracy depend on the implementation. The FAMM case shows that even when platforms don't act automatically, rights holders are watching viral AI tracks and pursuing claims directly.
If I use AI to create music, can someone claim royalties from me?
This is genuinely unsettled legal territory. FAMM's claim is based on voice impersonation and unauthorized use of likeness — not just metadata. The outcome of cases like this will help define what rights artists have over reconstructed voices. For now, the safest move is to ensure your AI-generated files don't carry the technical signals that make them trivially identifiable as AI-made.
Is this the same as removing a visible watermark from an image?
No — and it's important to be clear about the difference. A visible logo or watermark is a visual artifact you can crop out or paint over. The problem Calabi addresses is the invisible detection layer: the metadata, provenance manifests, and encoder fingerprints that tell a platform "this was AI-generated." Removing those signals is a file-level operation, not a pixel-level one. Calabi doesn't edit what a file looks like — it edits what a file says about itself.
Try Calabi free at calabilabs.com — 10 cleans, no card.
```