Calabi Labs · Guide · 2026-06-14
Ai detector
```html
What "AI Detector" Actually Means in 2026
When someone searches "AI detector," they're usually asking one of two things: how to spot AI-generated content, or how to get their own AI content past detection. If you're the second type—you're an AI creator who wants to post your work without an algorithm shadowbanning or suppressing it—here's what you're actually up against.
Platforms like Instagram, TikTok, YouTube, and Reddit don't use human eyes to scan your uploads. They run automated forensic checks on the file itself, reading invisible metadata and embedded signals that say "this was made by AI." That's the layer Calabi targets. The free trial is 10 cleans with no card required at calabilabs.com.
What Actually Gets Your File Flagged
The detection layer isn't one thing—it's a stack of signals baked into every AI export. Understanding each one matters because the obvious fixes only address the surface.
C2PA and Content Credentials are the biggest culprit. AI tools like Midjourney, Sora, and DALL-E embed a cryptographic manifest called JUMBF that stores the C2PA standard—essentially a tamper-proof "made by AI" label. This isn't in the visible metadata; it's a structured data blob that survives cropping, recompression, and most re-exports. ExifTool (the same forensic tool newsrooms use) reads it directly. A single Midjourney export can contain 18 separate JUMBF atoms declaring its AI origin.
XMP AI flags are the next layer. The DigitalSourceType tag—specifically set to trainedAlgorithmicMedia—tells any scanner "this came from an AI model." Generator software also writes tool names, model versions, and prompts into XMP fields that most users never see.
Encoder fingerprints are subtler. AI video exports from tools like Runway or Pika use specific encoders—Lavc (FFmpeg's encoder library), x264, or VP9—and embed SEI (Supplemental Enhancement Information) messages in the bitstream. These aren't visible to players, but a forensic parse reveals them instantly. A standard phone recording has none of these signals.
Missing capture metadata completes the profile. Real phone photos and videos carry Make, Model, Software version, GPS coordinates, and a capture timestamp. AI exports have none of this—or they have placeholder values. The absence itself is a signal.
Why the Obvious Fixes Don't Work
You've probably tried these already:
- Cropping removes the visible watermark—the corner logo or sparkle icon—but the C2PA manifest, XMP tags, and encoder fingerprints survive the crop. Platforms aren't flagging the pixels you can see; they're reading the ones you can't.
- Screenshotting loses some metadata, but you're still left with a screen recording that carries its own software fingerprint. Plus you lose resolution and add visible compression artifacts that scream "re-capture."
- Re-exporting through HandBrake or QuickTime strips some metadata, but C2PA manifests are designed to persist through re-encoding. You might shrink the file, but the "made by AI" manifest often survives intact.
- Adding your own metadata manually is time-consuming, error-prone, and you won't know which fields you missed. A single missed JUMBF atom can still flag you.
How Calabi Actually Cleans Your File
Calabi runs a three-stage pipeline in a single pass—no manual editing, no guessing.
- Strip: Calabi removes every detection signal in one operation—C2PA manifests (all 18 JUMBF atoms), XMP flags like
DigitalSourceType: trainedAlgorithmicMedia, generator/tool tags, and Lavc/x264 SEI encoder fingerprints. A raw AI export drops from roughly 144 metadata tags to about 94 neutral structural tags. - Inject: Calabi writes authentic phone-capture identity into the file—Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder profile. You can choose from device profiles including iPhone 15 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. The file now looks like it came from that device.
- Verify: Before you download, Calabi shows you a forensic proof card—the same ExifTool scan platforms run—displaying exactly what was stripped and what was injected. You see what the algorithm will see.
The result is a file that reads as a normal phone recording at the forensic level. No AI tool names, no C2PA manifests, no encoder fingerprints—just standard phone metadata. The visible content is unchanged.
Frequently Asked Questions
Can Calabi guarantee my post won't get flagged?
No tool can guarantee that—platforms update their detection methods constantly and perceptual hashes (invisible patterns in the actual pixels) are harder to disrupt than metadata. Calabi removes the metadata and encoder signals completely, which is what most platforms scan for automatically. Results vary by platform and source model.
Does Calabi change how my image or video looks?
No. Calabi doesn't touch the pixels. It works only on the invisible metadata layer, stripping detection signals and injecting phone-capture identity. If you have a visible watermark or logo, you'd need to crop it out separately—Calabi removes the invisible layer that survives cropping.
Try Calabi free at calabilabs.com — 10 cleans, no card.
```