Trend report · gnews_detection · 2026-06-18
The World’s Leading Deepfake Expert No Longer Trusts His Own Eyes - The New York Times
The Expert Who Built Deepfake Detection Now Doubts His Own Eyes
When the world's leading deepfake detection expert starts squinting at his own monitor, you know the problem has crossed a line. A recent New York Times profile documented how even the researchers who built AI content scanners can no longer trust visual evidence the way they once did. That's not hyperbole — it's the new normal for anyone working at the frontier of synthetic media in 2026.
But here's what the headlines don't tell you: the detection arms race has shifted. It's no longer about whether an image looks real. Platforms don't rely on eyeballs — they scan files for invisible forensic signals, and those signals are what get creators flagged, shadowbanned, or demonetized. Understanding what platforms actually scan is the only way to stay ahead of the filter.
What Actually Flags Your File in 2026
Platforms like Instagram, TikTok, YouTube, and Reddit run automated content scans within seconds of upload. They're not analyzing pixels — they're reading metadata, checking cryptographic manifests, and fingerprinting encoder signatures. Here's what's actually triggering the algorithms:
- C2PA / Content Credentials (JUMBF manifests): This is the big one. The Coalition for Content Provenance and Authenticity embeds cryptographically signed manifests inside files using JUMBF (JPEG Universal Metadata Box Format). These manifests explicitly state "generated by AI" and include the model name, generation tool, and timestamp. A Sora export, Runway clip, or Midjourney render carries this manifest by default. ExifTool — the same forensic tool newsrooms use — reads it in under a second.
- XMP AI metadata flags: Even without C2PA, XMP metadata fields like
DigitalSourceType: trainedAlgorithmicMediaact as explicit AI declarations. Adobe, OpenAI, and most major generators write these tags automatically. - Encoder fingerprints: Video files carry SEI (Supplemental Enhancement Information) NAL units from encoders like Lavc (libavcodec) or x264. These fingerprints identify the software that encoded the file — and no phone records a video through x264.
- Missing device identity: A file with no GPS coordinates, no camera Make/Model tag, no capture timestamp, and no real-phone encoder name stands out immediately against the billions of genuine phone captures uploaded daily.
- Perceptual hashes (pHash / aHash): Platforms compare your file's perceptual hash against databases of known AI outputs. These hashes survive recompression and are updated continuously as new models ship.
The brutal reality: cropping a Sora video or removing a visible sparkle watermark does nothing to these invisible signals. The C2PA manifest, the XMP flags, and the encoder fingerprint survive. That's what gets you flagged — not the pixels you can see.
How Calabi Fixes the Invisible Problem
Calabi is a one-pass web tool that rebuilds your file's forensic identity from the ground up. It doesn't edit pixels, doesn't erase logos, and doesn't touch what the image actually looks like. Instead, it operates on the metadata and structural signals that platforms actually scan.
The pipeline runs three stages:
- Strip: Calabi removes every AI-detection signal in one pass. JUMBF / C2PA manifests are zeroed out — 18 JUMBF atoms and 16 C2PA references reduced to 0. The
trainedAlgorithmicMediaXMP flag is deleted. Encoder SEI fingerprints from Lavc, x264, and similar software-only encoders are stripped. A raw AI export's 144 metadata tags compress down to roughly 94 neutral structural tags. - Inject: Calabi writes authentic phone-capture identity into the file. This includes a real device profile — iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra — with matching Make, Model, Software version, GPS coordinates, and capture timestamp. The encoder name switches to match the target device's actual codec.
- Verify: Before download, Calabi generates a forensic proof card — the same ExifTool scan platforms use — showing exactly what was stripped and what was injected. You see what Instagram's algorithm will see.
What Gets Creators Flagged on Instagram vs. TikTok
Both platforms scan for the same core signals, but their tolerances differ:
| Signal | TikTok | |
|---|---|---|
| C2PA manifest present | Immediate flag | Immediate flag |
| XMP AI flag | High risk | High risk |
| Software-only encoder | Moderate risk | High risk |
| Missing GPS/timestamp | Moderate risk | Low risk |
| Clean phone identity injected | Cleared | Cleared |
TikTok has been more aggressive on encoder fingerprints since 2025. Instagram tends to weight C2PA manifests and XMP flags more heavily. Either way, stripping the AI metadata layer and injecting a real device profile clears both.
Step-by-Step: How to Clean a File with Calabi
- Upload your AI-generated video or image. Drag-and-drop or click — no account required for the free trial.
- Calabi's pipeline runs automatically. No settings, no configuration screens. Strip → Inject → Verify happens in one pass.
- Review the forensic proof card. See exactly what was removed (C2PA manifest, XMP flags, encoder fingerprint) and what was injected (device profile, GPS, timestamp).
- Download the cleaned file. The file is now structurally identical to a phone recording from the device profile you selected.
- Upload to your platform of choice. The invisible signals that triggered previous scans are gone.
FAQ
Does removing the visible watermark from a Sora export fix the detection problem?
No. Visible watermarks — the corner logo, the sparkle icon — are cosmetic. The detection signals that get you flagged are invisible: the C2PA manifest, the XMP metadata, and the encoder fingerprint. These survive cropping and are what platforms actually scan.
What device profiles does Calabi support?
Current profiles include iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. Each profile injects the corresponding Make, Model, software version, encoder name, and GPS coordinates.
The expert in that Times profile still can't trust his own eyes — and neither can the platforms. That's precisely why they've shifted to scanning the file structure instead. The fix isn't about making your content look more real to human viewers. It's about making your file look like a real phone recording to the algorithm.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.