Trend report · hn_ai · 2026-06-16
In Age of AI, Leading Deepfake Expert No Longer Trusts His Own Eyes
When the world's leading deepfake researcher says he can no longer trust his own eyes, you know we've hit an inflection point. Hany Farid — the Dartmouth professor whose face-detection work helped define the field — told the New York Times he's now skeptical of video evidence he hasn't verified through forensic tools. In 2026, that's not paranoia. That's realism.
What actually flags your AI content on platforms
Creators assume platforms detect AI content visually. They don't. Detection happens at the file level, before a single frame displays, by scanning invisible metadata and structural signatures that survive re-encoding and even cropping.
The three signal families that get you flagged:
- C2PA / Content Credentials (JUMBF atoms) — A cryptographic manifest embedded in the file that says "this was generated by AI" with cryptographic signatures. A raw Sora export carries 18 of these atoms. A Midjourney v7 export carries 12. Platforms like Instagram and Reddit scan for these at upload, not by looking at pixels.
- XMP metadata flags — Fields like
DigitalSourceType: trainedAlgorithmicMedia,GeneratorSoftware, andAIleDatatags sit in the file header. ExifTool reads them in under a second. So does TikTok's automated upload scanner. - Encoder fingerprints — FFmpeg's Lavc encoder writes a specific SEI NAL unit signature into H.264/H.265 bitstreams. x264 and x265 each have distinct artifact patterns. A file exported from Runway or Pika carries this fingerprint whether it looks like a phone recording or not.
Beyond metadata, platforms run perceptual hash comparisons against known AI-generation models. A file missing a GPS coordinate, capture timestamp, or Make/Model field — all standard in phone captures — raises an automatic flag even if the metadata is otherwise clean. That's why cropping or re-encoding a Sora export doesn't fix the problem: the detection signals are in the file structure, not the visible pixels.
How Calabi handles it — strip, inject, verify
Calabi runs a single automatic pipeline that addresses all three signal families at once. No manual editing. No selecting regions. Upload, wait, download.
Stage 1 — Strip. The pipeline removes every detectable AI signal from the file. C2PA manifests are zeroed out. XMP fields like DigitalSourceType and GeneratorSoftware are deleted. Lavc and x264 SEI fingerprint units are stripped from video bitstreams. In a single Calabi clean, a raw AI export drops from 144 metadata tags to roughly 94 neutral structural tags — the same count as a screen recording.
Stage 2 — Inject. Calabi writes authentic phone-capture identity into the file. This includes Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder profile. Available device profiles include iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. The injected data is structurally identical to what a native phone camera app writes.
Stage 3 — Verify. Before you download, Calabi generates a forensic proof card — the same ExifTool scan that platforms use. It shows exactly what was stripped and what was injected. You see the before and after: 18 JUMBF atoms reduced to 0, 16 C2PA references to 0, the trainedAlgorithmicMedia flag removed. This is the document you hold if anyone asks where the file came from.
The step-by-step
- Upload your AI-generated video or image. Drag and drop or select from your library. File size up to 500MB.
- Calabi runs the strip + inject pipeline automatically. No settings, no options to configure. The system picks the device profile and generates fresh GPS/timestamp data.
- Review the forensic proof card. See exactly what was removed and what was written. ExifTool output, side by side.
- Download the cleaned file. The proof card is included in your download package as a text file — handy if you're working with clients or need an audit trail.
FAQ
What about visible watermarks — like Sora's sparkle icon or a corner logo?
Cropping removes the visible mark. Calabi removes the invisible detection layer that survives cropping — the C2PA manifest, XMP flags, and encoder fingerprints that get you flagged even after you've cut the logo out. Use both: crop first, then run Calabi.
Can I use a specific phone model for the injection?
Yes. Device profiles include iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. The injected metadata uses that device's exact Make/Model/Software string.
Does re-encoding through HandBrake do the same thing?
Re-encoding strips some encoder fingerprints but leaves C2PA manifests and XMP flags intact. It also degrades quality. Calabi removes all three signal families in one pass and keeps your original file quality intact.
Farid's point isn't that AI content is bad. It's that the infrastructure we built for a world of photographic truth — where "I was there, I saw it" was evidence — is broken. The fix isn't trusting your eyes less. It's making sure the file that carries your content looks, at the forensic level, like it came from a phone.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.