Trend report · gnews_detection · 2026-06-14

‘Inconsistent’ AI detection ‘should prompt assessment rethink’ - Times Higher Education

When AI Detection Fails Everyone—Including the Humans

If you've been watching the AI detection debate in higher education, you've probably noticed a recurring headline: the detectors don't work. A Times Higher Education analysis just confirmed what many educators already suspected—AI detection tools flag human students, miss AI outputs, and contradict each other across platforms. But here's the part the academic debate keeps missing: the same detection infrastructure that's failing classrooms is actively scanning every video and image you upload to Instagram, TikTok, YouTube, and Reddit. And unlike a worried professor, those platforms act on it automatically.

The inconsistency in detection isn't a bug—it's a signal that the underlying technology is still being figured out. But that hasn't stopped platforms from deploying it at scale. If you're creating AI-generated content and posting it anywhere, you need to understand what these systems are actually checking, because they're not looking at the same things detection companies are selling to universities.

What Actually Gets Your File Flagged

In 2026, platforms don't "detect AI" the way a human would. They scan for embedded signals that indicate an AI model generated or modified the file. The three main categories:

C2PA / Content Credentials (JUMBF manifests): This is the big one. C2PA embeds a cryptographic manifest inside the file—stored as JUMBF atoms—that declares the file was made by AI. When you export from Midjourney, Runway, Sora, or Kling, this manifest gets baked in automatically. ExifTool, the same forensic tool newsrooms use, reads these manifests. A raw AI export often contains 18+ JUMBF atoms and 16+ C2PA references. Instagram and TikTok scan for these during upload, often within seconds.
XMP metadata flags: The field DigitalSourceType gets set to trainedAlgorithmicMedia in AI exports. Generator tags, software version strings, and tool identifiers appear in the XMP block. A typical AI export carries 144 metadata tags; a real phone photo carries maybe 30-40.
Encoder fingerprints: Video files from AI generators carry recognizable encoder signatures. Lavc (FFmpeg's libavcodec), x264 SEI messages, and similar markers are dead giveaways. A real iPhone 16 Pro records with Apple or VideoToolbox encoders—not Lavc.

Then there's the metadata layer: missing GPS coordinates, capture timestamps that don't match upload time, device Make/Model fields that reference software instead of hardware. A file without a plausible phone identity is a red flag, regardless of how it was generated.

The Platform Stack in 2026

Here's how this plays out on the big four platforms:

Instagram: Runs automated AI detection on upload. C2PA manifests trigger review. Visible watermarks can be cropped out, but the invisible metadata layer survives cropping—which is what gets scanned.
TikTok: Similar automated pipeline. The DigitalSourceType flag alone can flag your content before a human ever sees it.
YouTube: Scans uploaded content for AI-generated signals. C2PA Content Credentials are checked, and content with declared AI origins gets labeled or restricted.
Reddit: Automatically scans uploads for AI signals. Posts flagged as AI-generated get labeled or removed from certain feeds.

The gap the Times Higher Education article identifies in academic detection doesn't exist here—platforms aren't trying to "detect creativity." They're matching specific technical fingerprints. And unlike a professor second-guessing whether a paragraph sounds robotic, these systems execute deterministic checks on deterministic data.

How Calabi Handles It

Calabi works on the same principle: strip the signals that declare AI origin, then inject authentic phone-capture identity so the file looks like what it claims to be. Three stages:

Strip: Remove all C2PA / JUMBF manifests, XMP AI flags (DigitalSourceType: trainedAlgorithmicMedia), generator tags, encoder fingerprints, and extraneous metadata. A file that started with 144 tags ends up with roughly 94 neutral structural tags. The "made by AI" layer is gone.
Inject: Add real phone identity: device Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder name. Device profiles include iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. The file now carries the metadata footprint of an actual phone recording.
Verify: Return a forensic proof card—generated by ExifTool—showing exactly what was stripped and what was injected. This is the same scan platforms use, so you see what they'll see.

The Honest Edges

Calabi removes the invisible detection layer that survives cropping. If your AI export has a visible watermark—a corner logo, Sora's sparkle—cropping removes the visible mark. Calabi removes the invisible signals that would flag the file even after cropping.

Invisible pixel watermarks (embedded in the image data itself) are a different case. A re-encode disrupts some patterns, but results vary by platform and source model. Calabi fully removes the C2PA/metadata/encoder signals. No tool can guarantee a platform won't flag you for every possible detection method.

Step by Step

Upload your AI-generated video or image to Calabi.
The automatic pipeline strips C2PA manifests, XMP AI flags, and encoder fingerprints.
Clean phone identity is injected—choose a device profile or let Calabi select automatically.
Review the forensic proof card showing the before/after metadata state.
Download the cleaned file with phone-capture identity intact.

FAQ

Can I just delete metadata manually? You can strip some fields, but C2PA manifests are embedded at the bitstream level—they're not simple metadata fields you can clear with "Delete All." Calabi removes them structurally.

Will platforms still detect my content as AI? Platforms scan for the specific signals Calabi removes. Results vary by platform and source model, but those signals—C2PA, XMP flags, encoder fingerprints—are what automated systems check. Removing them eliminates the primary detection vector.

What device profiles are available? iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, Galaxy S24 Ultra—real phones with documented encoder and metadata profiles.

The academic world is right to call for an assessment rethink. But if you're posting AI content on social platforms, the question isn't whether detection is inconsistent—it's what signals your specific file is carrying, and whether you're okay with platforms matching them automatically.

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

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

Try free →