Calabi Labs · Guide · 2026-06-15
Jorja smiths label famm seeks share of royalties from viral track i ru
```html
Jorja Smith's Label FAMM Is Suing Over an AI-Cloned Track — Here's What That Means for Creators
Jorja Smith's independent label FAMM has filed a legal claim against the dance act Haven, demanding a share of royalties from the viral track "I Run." FAMM alleges the song used AI to clone Smith's voice without authorization — training a model on her existing recordings and generating new vocal tracks that closely resemble her. The track went viral on TikTok in late 2025 and was later re-released with new vocals, though FAMM says both versions infringe on Smith's rights. The case is one of the highest-profile examples yet of a recording artist fighting back against AI voice cloning, and it's raising hard questions about how platforms detect AI-generated music — and how creators who use AI tools can end up exposed.
What Actually Gets a Track Flagged as AI-Generated
Most people assume a track gets flagged because it sounds like someone. That's not wrong, but it's not the whole picture. Platforms like TikTok, YouTube, and Instagram scan the invisible metadata layer of uploaded files — not just the audio itself. That layer carries signals that experienced forensic tools can pick up, and they're increasingly automated.
One major signal is C2PA / Content Credentials — a cryptographic manifest embedded in some files that records how the media was made, including whether AI generation tools were used. If a track was produced using an AI vocal synthesizer or a model trained on别人的声音, that process can leave a trace in the file's metadata. The XMP tag DigitalSourceType: trainedAlgorithmicMedia is a specific flag that explicitly marks a file as having been produced using machine learning. That's a red flag the moment a platform scans for it.
Encoder fingerprints are another layer. When audio is exported from AI tools, it often passes through specific software libraries — names like Lavc (FFmpeg's encoder) or video files carrying x264 SEI messages — that leave a recognizable fingerprint in the bitstream. A phone recording doesn't carry those fingerprints. An AI export does. And on the metadata side, a typical AI-generated audio file can carry over 144 metadata tags, many of which identify the generation tool, the model, or the training data source. Platforms scanning uploads don't need to listen to your track — they can pull that information straight from the file header.
Why the Obvious Fixes Don't Actually Work
If a creator releases an AI-assisted track and worries about detection, the instinct is to make it look different. Common approaches include converting the file format, screenshotting and re-recording audio, or cropping out visible markers. None of these remove the invisible signals that platforms actually scan.
Converting formats — MP3 to WAV, for instance — strips some metadata, but rarely all of it, and it doesn't remove encoder fingerprints embedded in the bitstream. Screenshotting audio as video and re-recording it does disrupt certain metadata tags, but it doesn't remove C2PA manifests or XMP flags that may already be embedded at the frame level. And cropping a video file to remove a visible watermark doesn't touch the metadata layer underneath, which is exactly what automated scanning tools read. The forensic signals that actually trigger platform flags are in the file structure, not on the screen.
The legal case between FAMM and Haven illustrates why that matters. Even after Haven re-released "I Run" with what appear to be new vocals, FAMM contends both versions still infringe on Smith's rights — suggesting the AI origin signal survived the re-release. That's because removing visible content is not the same as removing forensic metadata.
How to Actually Manage AI Signals in Your Files
For creators working with AI-assisted production — whether that's AI vocals, AI-mixed stems, or AI-generated backing tracks — the challenge isn't about hiding what you made. It's about controlling what the file says about how it was made. That's where a tool like Calabi operates: it works at the file level, stripping the invisible signals that platforms scan for and injecting the metadata profile of a standard mobile device recording.
Here's what that process looks like:
- Upload your track. Drop the AI-generated or AI-assisted file into Calabi. No manual editing, no selecting regions.
- Automatic stripping. Calabi removes C2PA / Content Credentials manifests, XMP AI flags like DigitalSourceType: trainedAlgorithmicMedia, encoder fingerprints (Lavc, x264 SEI), and excess metadata tags. A raw AI export carrying 144 metadata tags gets reduced to roughly 94 neutral structural tags — the kind a phone recording naturally carries.
- Injection of authentic device identity. Calabi injects Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder name — profiles drawn from actual devices like iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra.
- Review the forensic proof card. Before downloading, you see exactly what was stripped and what was injected — the same ExifTool scan that forensic investigators and platform tools use. You know what the file looks like under the hood before it goes anywhere.
- Download the cleaned file. The output carries the metadata profile of a genuine phone recording, with the AI-generation signals removed.
This is not photo editing. There's no inpainting, no pixel-level reconstruction, no selecting or filling regions. It's purely a metadata operation — but in 2026, that's exactly what platform scanning tools are built to read.
Frequently Asked Questions
Does this make the track no longer AI-generated?
Calabi removes the forensic signals that platforms scan for — the C2PA manifests, XMP flags, encoder fingerprints, and excess metadata that identify AI origin. It doesn't change the actual audio content. Whether that matters legally or ethically depends on your jurisdiction and the rights of any artists whose voices were used to train the model — FAMM's lawsuit against Haven is pursuing exactly that question.
Can Calabi remove a visible watermark or logo from a video?
No. Calabi does not erase, crop, or edit visible content. If an AI video export carries a visible watermark — a corner logo, a sparkle icon, a text overlay — cropping the frame removes the visible mark. But that crop doesn't touch the metadata layer, which is what platform scanners read. Calabi handles the metadata side; visual cropping handles the visible side. They address different layers of the same file.
Can platforms still detect AI audio even without metadata signals?
Some platforms also use perceptual hashing — analyzing the actual audio waveform for patterns associated with specific AI models or training data. Results vary by platform and by source model. Calabi fully removes the metadata and encoder signals that automated scanners flag at the file-header level. It disrupts some perceptual patterns, but no tool can guarantee a platform won't flag content on other grounds. Results vary, and the legal landscape around AI voice cloning — as the FAMM lawsuit shows — is developing fast.
Try Calabi free at calabilabs.com — 10 cleans, no card.
```