Trend report · gnews_detection · 2026-06-18
Emirati researcher develops AI platform to detect deepfakes and fake content - Gulf News
Another Deepfake Detector Hit the News — Here's What Actually Gets Your Content Flagged
When Gulf News reported that an Emirati researcher built an AI platform to spot deepfakes and fake content, it landed in the same feed as a dozen similar headlines this year. Researchers are racing to build better detection models. Meanwhile, platforms like Instagram, TikTok, YouTube, and Reddit have already deployed automated scanning that flags AI-generated content before a human ever sees it — and those systems aren't looking at pixels. They're reading metadata.
If you're posting AI-generated or AI-edited videos and images, the question isn't whether deepfake detection exists. It's what invisible signals in your file are already triggering those filters right now.
What Actually Flags Your File
Platforms in 2026 don't primarily rely on AI-vs-real classification. They scan for specific technical signals embedded in the file itself — signals that survive recompression and often survive cropping.
C2PA / Content Credentials
The most significant flag is C2PA — the cryptographic manifest standard that stores who made a piece of content and with what tools. Adobe, Microsoft, Google, and most major camera manufacturers back it. When you export from Midjourney, Sora, Runway, or Leonardo AI, the file gets tagged with a JUMBF box containing a C2PA manifest. That manifest references a digitalSourceType of trainedAlgorithmicMedia — a direct statement: this came from an AI model.
A single AI export can carry 18 or more JUMBF atoms and 16 distinct C2PA references. Platforms parse these with ExifTool or similar parsers and reject or label the content automatically.
XMP AI Metadata Flags
Outside the C2PA block, XMP packets carry DigitalSourceType, CreatorTool, and Generator fields. These appear in plain text. A Runway export might show Generator: Runway Gen-3 Alpha in the XMP. A Sora export writes Software: OpenAI Sora. These aren't hidden. They're read by platform scanners as explicit AI disclosure.
Encoder Fingerprints
AI video models output with specific encoder signatures baked into the bitstream. x264 and x265 encodeurs write SEI (Supplemental Enhancement Information) messages. Libavcodec (Lavc) writes its own identification tags. Platform scanners maintain blocklists of encoder combinations that only appear in AI-generated content. A file with Lavc 61.x or x264 core 164 SEI headers from an AI pipeline gets a hit.
Missing Context Signals
Real phone captures carry GPS coordinates, precise capture timestamps, Make/Model tags, and software version strings. AI exports have none of these — or they inherit a generator string instead. The absence of phone-identity metadata is itself a signal. Platforms weight this alongside the positive AI flags.
Perceptual Hashes
Some platforms compute pHashes or aHashes on the visual content itself. These are less reliable than metadata scanning and results vary, but for high-profile accounts or viral content, perceptual hashing is part of the multi-signal check that triggers review queues.
How Calabi Handles It
Calabi is a one-pass web tool that doesn't edit pixels or guess what's real. It works on the metadata and bitstream layer — stripping the signals that trigger automated flags and injecting the identity profile of a real phone capture.
Stage 1: Strip
The pipeline removes every detectable AI signal. C2PA manifests get zeroed — all 18 JUMBF atoms and 16 references reduced to 0. XMP fields like DigitalSourceType: trainedAlgorithmicMedia, CreatorTool, and Generator are stripped. Encoder SEI messages from Lavc and x264 that identify AI pipelines are cleared. A raw AI export that carried 144 metadata tags comes out with roughly 94 neutral structural tags and zero AI disclosures.
Stage 2: Inject
Calabi then writes authentic phone-capture identity into the file. You select a device profile — iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra — and the tool injects Make, Model, Software version, a plausible GPS coordinate, and a capture timestamp in the correct EXIF/XMP field format. It also replaces the encoder fingerprint with a real-phone codec signature (not a generator string).
The result looks, to automated scanners, like any other phone video posted without editing.
Stage 3: Verify
Before download, Calabi generates a forensic proof card — an ExifTool readout showing exactly what was stripped and what was injected. This is the same tool newsrooms and platform trust-and-safety teams use. You see the same scan they'll see. No guessing whether it worked.
Step-by-Step: What the Workflow Looks Like
- Upload your file — video or image, up to the platform's size limit. No manual settings required.
- Select a device profile — iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra. The tool handles the full metadata profile for that device.
- Automatic pipeline runs — C2PA strip, XMP cleanup, encoder signature rewrite, phone identity injection. You wait. No decisions to make.
- Review the forensic proof card — See the before/after ExifTool output. Confirm the AI signals are gone and phone identity is present.
- Download the cleaned file — Post it to Instagram, TikTok, YouTube, or Reddit. The metadata layer matches a real phone capture.
FAQ
What about visible watermarks, like Sora's sparkle or a corner logo?
Calabi removes the invisible detection layer — the metadata and encoder signals that survive after you crop out a visible mark. If you crop the visible watermark, Calabi sanitizes everything that would survive that crop. No tool rewrites pixels to erase a visible logo.
Can platforms still detect my content as AI-generated?
No tool can guarantee a platform won't flag you — detection methods evolve. Calabi eliminates the metadata and encoder signals that power automated scanning. Results vary by platform and source model, but the metadata layer that automated filters read is gone.
Does this work on screenshots of AI images?
Screenshots carry the screen capture device's metadata — Make, Model, Software — which helps. But they often still contain embedded AI tool metadata from the original render if the screenshot preserved EXIF data. Calabi strips残留 signals regardless of how the file reached its current state.
The Detection Gap Is a Metadata Gap
Researchers building AI classifiers are solving a hard problem — and they're solving it slowly, with accuracy rates that fluctuate by source model and compression level. Platform scanners aren't waiting for better classifiers. They're using metadata signals that are already there, already parseable, and already automated. The detection gap isn't about spotting pixels. It's about reading what a file says about itself.
Calabi closes that gap by making your file say something different.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.