Trend report · gnews_detection · 2026-06-15
How hard can it be? Testing the dependability of AI detection tools - Times Higher Education
Why AI Detection Tools Keep Flagging Your Content — And What Actually Fixes It
A recent Times Higher Education investigation put AI detection tools through their paces, asking a deceptively simple question: how reliable are they really? The answer, unsurprisingly to anyone who's uploaded an AI-generated video to Instagram only to watch it get throttled, is "not very — but that's only half the problem." Detection tools vary wildly in accuracy, but the real issue is that platforms aren't primarily relying on detection tools. They're scanning the metadata, encoding signatures, and cryptographic manifests embedded in your file. That's what gets you flagged, and that's what you need to address.
What Actually Flags Your File on Instagram, TikTok, and YouTube
In 2026, major platforms run automated scans on every upload — often within seconds. They're not guessing whether your video "looks AI." They're reading the invisible identity layer baked into every AI-generated export. Three things trigger flags most often:
- C2PA / Content Credentials — The cryptographic manifest (stored as JUMBF atoms) that declares "this file was generated by an AI model." A Sora export, for instance, carries multiple JUMBF atoms referencing the model that created it. Platforms like Instagram and Reddit parse these atoms directly. If your file has 18 JUMBF atoms, that's 18 direct signals saying "AI-generated."
- XMP AI metadata flags — Fields like
DigitalSourceType: trainedAlgorithmicMedia,GeneratorSoftware, andAIGeneratedLabel. These survive re-encoding if not explicitly stripped. A raw AI export can carry 144 metadata tags; after a basic re-save you're still looking at 90+ tags with AI fingerprints baked in. - Encoder fingerprints — Video files carry encoder signatures in the bitstream. Lavc (FFmpeg's encoder) SEI messages, x264编码参数, and similar signatures are dead giveaways. Platforms maintain lists of known AI-generation encoder patterns. Your file looks like it came from a render farm, not a phone.
Missing GPS, timestamp, and device identity compounds the problem. A phone-recorded video carries Make (Apple), Model (iPhone 15 Pro), Software version, GPS coordinates, and a capture timestamp. An AI export has none of that. The absence itself is a signal.
The Visible Watermark Problem — And Its Honest Solution
If your AI video has a visible watermark — a corner logo, Sora's sparkle indicator — cropping removes it. That's the honest fix for the visible layer. But here's what the marketing blurbs don't tell you: the invisible detection layer survives cropping. The JUMBF atoms, the XMP flags, the encoder fingerprints — those stay embedded in every pixel and every metadata block of the cropped file. Cropping fixes what you see. It does nothing to the signals platforms actually scan.
That's the gap Calabi fills — not by touching a single pixel of image data, but by rewriting the file's invisible identity.
How Calabi Handles It: Strip, Inject, Verify
Calabi runs a three-stage pipeline on every upload. No manual editing, no sliders, no region selection.
Stage 1 — Strip: Remove every AI-detection signal from the file. C2PA / JUMBF atoms reduced to 0. C2PA references zeroed out. XMP fields like DigitalSourceType and GeneratorSoftware deleted. Encoder fingerprints (Lavc SEI, x264 signatures) stripped from the video bitstream. A raw export's 144 metadata tags compressed to roughly 94 neutral structural tags — no AI language anywhere.
Stage 2 — Inject: Write authentic phone-capture identity into the file. Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder name. Calabi draws from real device profiles: iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra. The file now reads as a normal phone recording at the metadata level.
Stage 3 — Verify: Return a forensic proof card — the same ExifTool scan platforms use — showing exactly what was stripped and what was injected. You see before-and-after metadata. You know what's been removed. You know what identity was written in its place.
What Happens When You Upload After Calabi Processing
- Upload your AI-generated video or image to Calabi. Drag-and-drop, no account required for the free trial.
- The pipeline runs automatically. Strip → Inject → Verify. No settings to adjust, no regions to select.
- Review the forensic proof card. See exactly which JUMBF atoms were removed, which XMP fields were cleared, which encoder fingerprints were stripped. Confirm the injected device profile matches a real phone.
- Download the cleaned file. The platform sees a phone recording. Instagram, TikTok, YouTube, Reddit — all run the same automated scans. The cleaned file passes the same checks a genuine phone capture would pass.
FAQ
Can't I just re-export my video to remove the metadata?
Partially. A re-export via standard software strips some metadata, but encoder fingerprints embedded in the bitstream (Lavc, x264 SEI messages) persist. C2PA/JUMBF atoms are also notoriously sticky — they survive many re-encoding passes. Calabi targets every signal category systematically, not just the easy metadata fields.
Does this work for images, or only video?
Both. Calabi handles images and video through the same strip-and-inject pipeline. A PNG or JPEG from an AI generator carries the same C2PA/JUMBF and XMP issues as a video file.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.