Trend report · gnews_detection · 2026-06-11

Deezer gives free access to AI-music detection tool on third-party playlists - Telecompaper

In March 2025, Deezer made headlines by offering free access to its AI-detection tool for third-party playlists—a watershed moment that signaled how seriously the music industry is taking synthetic audio. But Deezer's move is just one front in a much larger war. Platforms across the internet are now running increasingly sophisticated scans on every piece of content you upload. If you're a creator, a brand, or anyone who publishes media online, understanding what these scanners look for—and how to handle them—has become essential.

What Platforms Scan For in 2026

The detection landscape has evolved rapidly. In 2024, most platforms relied on basic heuristics. By 2026, the toolchain is mature, standardized, and running at scale. Here's what the scanners are actually checking:

1. C2PA (Coalition for Content Provenance and Authenticity)
   

   C2PA is now the de facto standard for content authentication. Developed by a consortium including Adobe, Microsoft, and Google, C2PA embeds cryptographically signed metadata into files at the moment of creation. The manifest includes fields like stds.schema-org.C2PAAssertions.digital_source_type and stds.schema-org.C2PAAssertions软 (though the technical manifest uses JSON-LD format). When you upload a JPEG, MP4, or WAV file, platforms extract the C2PA manifest and check for three things: whether the manifest exists, whether it's validly signed, and whether it claims the content is "algorithmicMedia" or "synthetic." Missing or invalid C2PA data is itself a red flag on high-trust platforms like YouTube's Content Authenticity Initiative integration.

2. AI Generation Metadata (EXIF and XMP)
   

   Beyond C2PA, platforms parse traditional EXIF and XMP metadata for AI fingerprints. Common flags include: AITool, Software, and Generator fields that list names like "Stable Diffusion 3.0," "DALL-E 3," or "Sora." For images, look for xmlns:xmpMM namespaces containing software-specific serial numbers. TikTok's classifier, for example, checks for the absence of typical camera EXIF profiles—modern phones add fields like Make, Model, GPSLatitude, and LensModel. An image generated by Midjourney often has Software="Midjourney" in the EXIF ImageDescription tag, which is a near-instant flag.

3. Encoder Signatures (Audio and Video)
   

   AI-generated audio leaves distinctive encoder fingerprints. When an AI model synthesizes speech (via ElevenLabs, OpenAI's Voice Engine, or similar), the output often passes through specific codecs. Platforms maintain hash databases of known AI audio artifact patterns. For video, the mdia.minf.stbl.stsd.mhlt atom structure in MP4s can reveal whether encoding was done by a human using Premiere Pro (which sets specific handler_name values and adds clap atoms with precise dimensions) versus an AI upscaler or generator that may leave the structure malformed or use non-standard brand identifiers.

4. Missing GPS and Sensor Metadata
   

   This is one of the most reliable signals. Real photos and videos taken on phones almost always contain GPS coordinates (GPSLatitude, GPSLongitude, GPSAltitude), accelerometer data, and gyroscope readings in the meta atom of video files. AI-generated images typically have no GPS tags whatsoever. Instagram's classifier flags any image where GPSAltitude is missing and the DateTimeOriginal field exists without corresponding location data. TikTok goes further: it cross-references the upload location (IP-derived) against the claimed GPS in metadata—if the image says "San Francisco, CA" but the IP suggests Virginia, that's a soft flag that triggers additional scrutiny.

What Actually Gets Flagged on Instagram and TikTok

Based on documented enforcement actions and creator reports through 2025-2026, here's what platforms are actively flagging:

• AI-generated profile pictures — TikTok has been particularly aggressive, flagging accounts using AI avatars for livestreams. The system detects stds.schema-org.C2PAAssertions manifests marked as digital_source_type: "algorithmicMedia".
• AI-upscaled or AI-interpolated videos — Instagram Reels flagged videos where the MotionPhoto metadata indicated frame interpolation software (common in apps like Remini or Topaz Video AI). The detection triggers on MakerNote tags that include "Topaz" or "DAIN" strings.
• Synthetic music clips — Following Deezer's lead, both platforms now run audio fingerprinting on clips, flagging AI-generated vocals detected through spectral analysis of formant frequencies and absence of natural breath patterns.
• AI-edited product photos — Commerce accounts have faced flags when product images show perfect lighting, zero dust particles, and perfectly consistent shadows—signals that the image passed through AI retouching pipelines.

The Durable Fix: Strip, Then Inject Clean Identity

Simply removing metadata isn't enough—platforms have moved beyond metadata-only checks. The durable solution is a two-step process:

1. Strip all embedded metadata
   • Remove C2PA manifests entirely (not just mark as invalid—remove the c2pa box from JUMBF files or the C2PA atom from MP4s)
   • Clear EXIF/XMP fields: use tools that zero out GPSLatitudeRef, GPSLongitudeRef, DateTimeOriginal, Software, Make, Model, LensModel, and all xmlns:xmpMM namespaces
   • Re-encode media through a clean pipeline that doesn't insert software markers—for images, convert through a library that produces minimal output (like libpng with --no-title flags)
   • For AI-generated video, removing embedded Sora or similar watermarks is part of this stripping step—these are visible in the perceptual hash layer, not just metadata
2. Inject clean device identity signals

   This is the step most guides skip. Metadata scanners check for absence of phone identity as much as presence of AI markers. You need to inject authentic device metadata that matches a real camera:

   • Add GPS coordinates that correspond to a plausible location (use real coordinates from a matching IP region)
   • Set Make and Model to match real phone models: e.g., Apple + iPhone 15 Pro, or samsung + SM-S928B
   • Add authentic timestamps: DateTimeOriginal, CreateDate, and ModifyDate in the correct EXIF date format (YYYY:MM:DD HH:MM:SS)
   • Include typical lens metadata: FocalLength, FNumber, ExposureTime that match the claimed camera model
   • For video, ensure the mdia.minf.stbl.stsd box uses standard avc1 or hvc1 brand identifiers with realistic Width and Height values

The key insight: platforms in 2026 don't just check "does this look AI?"—they check "does this look like it came from a real device?" A file with no metadata at all is itself suspicious. You need to provide the full fingerprint of a legitimate capture.

This is precisely what Calabi was built to handle. Our pipeline strips every trace of AI generation metadata, C2PA manifests, and encoder signatures—then re-injects the complete device identity profile of a real phone, with authentic GPS, timestamp, and camera metadata that passes platform scrutiny.

The Deezer news proves the direction: AI detection is moving from experimental to mandatory across every major platform. Get ahead of it.

→ Try Calabi free at calabilabs.com — 10 cleans, no card.

Related reading

• Remove the Seedance watermark
• What platforms actually scan for when detecting AI content
• Spoof iPhone 16 Pro metadata