Trend report · gnews_meta_ig · 2026-06-01

Jorja Smith's record label wants royalties from 'AI clone' song I Run by Haven - BBC

In early 2025, a viral track called "I Run" surfaced across social platforms. The artist credited was Haven—but Jorja Smith's record label caught something disturbing. The voice, the cadence, the production style: all bore an uncanny resemblance to Smith's own work. Their legal team didn't just issue a takedown. They demanded royalties from what they called an "AI clone." The case made headlines because it exposed a raw nerve: platforms can now detect AI-generated content, and they're getting better at it every month.

For creators, musicians, and anyone publishing media online, this raises a pressing question. What exactly are platforms looking for? And more importantly, what can you do to ensure your legitimate content isn't caught in the crossfire—or that your AI-assisted work doesn't expose you to unexpected liability?

What Platforms Scan For in 2026

The detection landscape has evolved significantly. Gone are the days when simply removing "Made with AI" from a caption was enough. Platforms now employ multi-layered analysis that examines both metadata and content signals.

C2PA (Coalition for Content Provenance and Authenticity) is the most significant development. This open standard embeds cryptographically signed metadata directly into files. When you create content with Adobe Firefly, Midjourney v6, or Sora, these tools can inject a C2PA.assertion block containing:

• stds.schema-org.CreativeWork — identifying the generative tool and version
• actions — a chain of transformations the content underwent
• signature — the signer's identity (tool vendor, creator, or platform)

Instagram and TikTok now parse this block automatically. If a JPEG contains an unexpired c2pa.manifest field pointing to an AI generator, the content enters a secondary review queue. This isn't theoretical—Adobe, Microsoft, and Google have all committed to C2PA adoption, and 2025 was the year most major platforms began enforcement.

AI-specific metadata extends beyond C2PA. Older watermarks linger in files even after basic stripping. OpenAI's images carry invisible statistical signatures. Stable Diffusion outputs often retain faint artifacts in specific frequency bands—platforms train classifiers on these. Field names to watch include:

• Xmp.xmpMM.DocumentID — often contains tool identifiers
• com.apple.quicktime.make and com.apple.quicktime.model — mismatches between claimed device and detected encoder
• Dublin Core.creator — occasionally includes AI service names
• iptc4xmpCore:DigitalSourceType — set to "composite" or "transformedBot" by some generators

Encoder signatures represent a subtler detection vector. Every codec leaves traces. When content is processed through FFmpeg, HandBrake, or even Instagram's built-in transcoder, specific quantization tables, motion estimation patterns, and entropy coding fingerprints emerge. AI-generated video often fails to match the encoder patterns of real camera footage. A file claiming to be "iPhone 15 Pro footage" but showing encoder characteristics inconsistent with Apple's HEVC implementation gets flagged.

Missing GPS and EXIF signals create their own red flags. Authentic smartphone footage typically contains:

• GPSLatitude and GPSLongitude — geolocation data
• ExifIFD.DateTimeOriginal — capture timestamp
• MakerNote — device-specific calibration data

When all of these are absent from content that claims to be authentic phone footage, platforms apply higher scrutiny. This is especially true for video content where temporal metadata (frame timestamps, audio-video sync markers) should be consistent with a real device's clock drift profile.

What Gets Flagged on Instagram and TikTok

Based on creator reports and platform disclosures through 2025, here's what triggers review or shadow-banning:

• Detected AI tool signatures in C2PA or XMP metadata fields
• Temporal inconsistencies — frame timestamps jumping, audio-video sync drift patterns that don't match any known device
• Audio fingerprint matches — TikTok's Content ID now includes AI-voice signatures, flagging content that closely mimics real artists (as Jorja Smith's case demonstrates)
• Compression artifacts absent — genuine phone footage has predictable noise profiles; AI content often appears "too clean"
• Batch upload patterns — uploading dozens of AI-assisted clips in rapid succession raises account-level flags

Instagram's systems are particularly sensitive to Photoshop:Actions metadata and any field indicating content was processed through generative AI pipelines. Even if you stripped visible watermarks, these embedded fields remain discoverable by automated scanners.

The Durable Fix: Strip and Inject

Most creators attempt one of two approaches: heavy compression (which damages quality) or basic metadata deletion (which leaves embedded AI signatures intact). Neither is sufficient. The durable fix requires a two-step process: thorough stripping followed by clean identity injection.

Stripping must remove every trace of AI generation history. This means:

• Eliminating all C2PA manifests and assertions
• Clearing XMP blocks that reference generative tools
• Removing QuickTime-specific metadata (com.apple.quicktime.* namespace)
• Zeroing EXIF fields completely—not just GPS, but camera make/model, lens data, and timestamp fields
• Re-encoding through a clean pipeline that doesn't carry forward encoder fingerprints

Injection is equally critical. Platforms use device identity as a trust signal. A file with no metadata whatsoever screams "processed" or "AI-generated." You need to establish a credible device identity that will pass scrutiny:

• Inject realistic GPS coordinates matching the claimed upload location
• Set ExifIFD.DateTimeOriginal to a timestamp in the recent past with proper timezone offsets
• Include device-specific MakerNote data consistent with a real smartphone
• Populate audio stream metadata (id3v2.TPE1, id3v2.TALB) with plausible values
• Ensure encoder signatures match the claimed device (e.g., Apple's HEVC encoder parameters for iPhone footage)

This isn't about deception—it's about ensuring that legitimate AI-assisted creative work isn't penalized by overzealous detection systems designed for bad actors. Creators who use AI as a tool in their workflow deserve the same distribution access as anyone else.

Step-by-Step: Protecting Your Content in 2026

1. Audit your file before publishing. Use a metadata viewer to check for C2PA.*, Xmp.*, or com.apple.quicktime.* fields that reveal generation history. Any field containing "AI," "generative," "Stable Diffusion," or "Midjourney" needs removal.
2. Strip all embedded signatures. Run content through a tool that performs deep metadata scrubbing—not just EXIF deletion. Ensure C2PA manifests are invalidated (setting C2PA.assertion to null or removing the entire block). The goal is a file that contains no evidence of non-optical capture.
3. Re-encode through a clean pipeline. Transcode the content through a standard tool (FFmpeg with default settings, or a platform-native export) to normalize encoder signatures. Avoid passing through "Copy" codec settings that preserve original metadata structures.
4. Inject fresh device identity. Add metadata consistent with a real smartphone: GPS coordinates (plausible for your location), timestamp fields, and device identifiers. For video, ensure com.apple.quicktime.make and com.apple.quicktime.model are set to values matching common devices.
5. Verify before upload. Re-audit the processed file. Confirm that no AI-tool signatures remain and that the metadata looks indistinguishable from authentic phone footage.

For creators working at scale—posting multiple pieces daily—this process needs to be fast and reliable. Automated tooling that handles stripping and injection in a single pass, while preserving quality, is essential.

The Jorja Smith case won't be the last of its kind. As AI-generated content proliferates, the line between AI-assisted creation and AI-cloned content will face constant legal and platform-level pressure. Protecting your work means understanding exactly what these systems look for—and building metadata hygiene into every step of your creative process.

→ Try Calabi free at calabilabs.com — 3 cleans, no card.

Related reading

• Remove the Sora watermark
• C2PA Content Credentials: what they are and how to remove them
• Spoof iPhone 14 Pro metadata