Trend report · gnews_detection · 2026-06-14

AI or Not Achieves 100% Detection of Deepfake X-Rays in Independent Test, With 95% Overall Accuracy, Exceeding Results Reported for Radiologists and Leading Multimodal LLMs - PR Newswire

AI Detection Is Getting Harder to Beat — Here's What Actually Flags Your Content

A new independent test from AI or Not just hit 100% detection accuracy on deepfake X-rays and 95% overall — outperforming radiologists and leading multimodal LLMs. That news lands in the middle of a platform crackdown that's been building for two years. If you're posting AI-generated video or images on Instagram, TikTok, YouTube, or Reddit in 2026, automated scanners are already checking your files for exactly the same signals: C2PA manifests, XMP metadata, encoder fingerprints, and missing device context. One pass through Calabi strips those flags and injects authentic phone-capture identity — because that's the only fix that holds up under forensic review.

What Actually Flags Your File in 2026

Platform scanners don't look at your content the way a human does. They read the invisible metadata layer underneath — the stuff that lives in file headers, manifests, and bitstream structures. Here's what they're actually checking:

C2PA / Content Credentials (JUMBF manifests): This is the cryptographic "made by AI" manifest that Adobe, Microsoft, and the C2PA coalition built into AI generation tools. It stores structured claims about who created the file, what model generated it, and when. A Sora export can carry 18+ JUMBF atoms. Platforms read these with ExifTool or equivalent parsers and flag any file with active AI provenance data.
XMP AI flags: XML Metadata Platform tags like DigitalSourceType: trainedAlgorithmicMedia sit directly in the file metadata. This is a machine-readable declaration that the content came from a trained model. It's separate from C2PA and gets flagged independently.
Encoder fingerprints: AI export pipelines leave distinct encoder signatures. Lavc (FFmpeg's libavcodec), x264 SEI user data, and specific FFmpeg builds have identifiable bitstream patterns. A file encoded by an AI pipeline looks different under a hex-level scan than a file encoded by a phone camera app.
Missing device context: Real phone captures include Make, Model, Software version, GPS coordinates, and capture timestamp. A freshly generated AI image has none of this. The absence itself is a signal — platforms weight missing EXIF context as a moderate-risk indicator.
Perceptual hashes: Platforms maintain hash databases of known AI-generated content. A re-encode disrupts some hash matches, but this layer is the least reliable and most evasion-prone. It matters less than metadata.

How Calabi Handles It — Three Stages

Calabi doesn't edit your image or touch the visible pixels. It works on the invisible layer that platforms actually scan. The pipeline runs in one pass:

Strip: Remove all C2PA / JUMBF manifests, XMP AI flags (including DigitalSourceType), generator/tool tags, and encoder fingerprints from the bitstream. For a raw AI export, this can mean reducing 144 metadata tags down to about 94 neutral structural tags. The "made by AI" signals are gone.
Inject: Write authentic phone-capture identity into the metadata: a real device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), plausible software version, GPS coordinates, and a capture timestamp. The encoder identity shifts from Lavc/FFmpeg to a phone-native encoder profile.
Verify: Return a forensic proof card — the same ExifTool scan platforms use — showing exactly what was stripped and what was injected. You see the before/after before downloading.

What Gets Flagged on Instagram, TikTok, YouTube, and Reddit

Each platform runs its own scanning stack but they're looking at the same underlying signals. A video uploaded to Instagram without device metadata, carrying C2PA manifests and x264 SEI fingerprints from an AI pipeline, gets flagged within seconds on upload. TikTok runs ContentSafe on ingestion and checks both C2PA manifests and encoder signatures. YouTube's classifier flags videos with trainedAlgorithmicMedia XMP tags before they finish processing. Reddit's AutoModerator and spam filters check for missing GPS and timestamp EXIF fields as a soft signal.

The critical point: cropping your image or video does not remove these signals. C2PA manifests and XMP metadata survive most crops because they're stored at the file level, not embedded in pixel regions. A visible watermark — a corner logo, a sparkle icon from Sora — is a separate problem that cropping can address. Calabi removes the invisible detection layer that cropping leaves behind.

Why This Is the Only Durable Fix

Detection models improve. Hash databases grow. If you strip only the flags a specific scanner checks today, you get caught when that scanner updates tomorrow. Calabi's approach is to replace the AI file identity entirely with a legitimate phone-capture identity — not to patch holes, but to rewrite the forensic record from the ground up. A file that looks, under ExifTool, like a phone recording from an iPhone 15 Pro in a specific location at a specific time is structurally indistinguishable from one that actually came from that device. That's what survives platform updates.

FAQ

What about visible watermarks like Sora's sparkle icon? Calabi removes the invisible detection and metadata layer — not visible pixels. Cropping removes a visible watermark. Calabi removes the invisible signals that survive cropping and get you flagged even after you've cropped.
Can this guarantee my content won't be flagged? No tool can guarantee that. Platform classifiers evolve and add new signals. Calabi removes the metadata and encoder layer that current scanners check; results vary by platform and source model.
Does re-encoding disrupt detection? A re-encode disrupts some perceptual hash matches, but it doesn't remove C2PA manifests, XMP flags, or encoder fingerprints — those survive re-encoding. Calabi removes all three layers.

The Forensics Don't Lie — Until You Rewrite Them

The AI or Not test shows how precise detection has become: 100% on deepfake X-rays, 95% overall. That precision is built on reading metadata, encoder patterns, and provenance manifests — the same signals Calabi strips and replaces in a single pass. If you're shipping AI content to platforms, those signals are your liability. Remove them and write a clean phone identity instead.

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

10 free cleans. See the forensic proof before you download.

Try free →