Trend report · hn_show · 2026-06-16
Show HN: ThoughtLeadin – AI-first LinkedIn where every post is corporate slop
If you are building tools that help people ship AI-generated content at scale—whether that is an AI-first LinkedIn post builder, a video renderer, or a content pipeline—platform detection is the problem that will eventually catch up with your users. Right now it is catching up fast.
What actually flags your file
When a piece of AI content gets rejected, shadowbanned, or suppressed on LinkedIn, Instagram, TikTok, or Reddit, it is rarely because a human moderator watched it. It is because an automated system ran a forensic scan on the file itself—and found signals that do not belong in a normal phone recording.
In 2026, those scans check three invisible layers. The first is C2PA / Content Credentials, stored as JUMBF atoms inside JPEG and video files. When you export from Midjourney, Sora, Runway, or any major AI generator, the file carries a cryptographic manifest that says exactly which model made it, when, and with what training data. ExifTool—the same tool newsrooms and platform trust-and-safety teams use—reads these JUMBF atoms and C2PA references. A raw AI export might carry 18 JUMBF atoms and 16 C2PA references. That is 34 separate cryptographic flags pointing directly at AI origin.
The second layer is XMP metadata. AI generators write fields like Iptc4xmpCore:DigitalSourceType = trainedAlgorithmicMedia, digiKam:TagsList = AI-generated, and tool-specific namespaces that identify the generator by name. A raw AI JPEG can carry over 144 metadata tags where a real phone photo carries fewer than 40. Platform scanners pattern-match on these tags. The moment they see DigitalSourceType: trainedAlgorithmicMedia, the file is flagged regardless of how good the content looks.
The third layer is encoder fingerprints. AI video exports carry codec signatures: Lavc (FFmpeg's libavcodec) in the bitstream, x264 SEI messages, and frame pattern anomalies that are statistically distinct from phone capture. H.264 and H.265 streams from FFmpeg contain encoder identification SEI NAL units that professional detection systems scan for. This is why re-encoding to "reset" the file often does not work—the Lavc fingerprint survives transcode if the codec chain is not fully broken.
LinkedIn, Instagram, and TikTok run these checks at upload, automatically, often within seconds. Reddit's AutoModerator applies similar rules to image posts. None of them need to "see" your content to act on it.
How Calabi handles it
Calabi is a one-pass web tool that strips these detection signals and injects authentic phone-capture identity in their place. It does not edit pixels, select regions, or change how the content looks. It works entirely on invisible metadata, manifests, and bitstream structures.
The pipeline runs in three stages:
- Strip. All JUMBF / C2PA atoms are removed. The 18 JUMBF atoms and 16 C2PA references go to zero. The DigitalSourceType: trainedAlgorithmicMedia flag is deleted along with every generator-specific XMP namespace and encoder fingerprint in the bitstream. The 144-tag AI export is reduced to roughly 94 neutral structural tags—the same baseline a phone produces.
- Inject. A real device profile is written into the metadata: Make, Model, Software version, GPS coordinates, capture timestamp, and a known-phone encoder name. Supported profiles include iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. This replaces the stripped AI identity with a plausible, consistent phone-capture fingerprint.
- Verify. A forensic proof card is generated showing the before/after ExifTool scan—the same scan platforms use. This is what you get to see what was stripped and what was injected before you download.
Why visible watermarks are a separate problem
If your AI tool's output has a visible logo or sparkle icon—Sora's watermark, Midjourney's credit stamp—that is a pixel in the corner of the image. Cropping removes it. Calabi does not erase logos pixel-by-pixel and never will. Calabi removes the invisible detection layer that survives cropping: the C2PA manifest, the XMP flags, the encoder fingerprint. Those signals are what actually trigger platform suppression, and they survive crop and re-encode because they are embedded in the file structure, not the visual content. Dropping a visible watermark and leaving the invisible layer intact is how you get flagged after you thought you were clean.
Step by step
- Go to calabilabs.com and upload your AI-generated image or video file.
- Calabi automatically strips all C2PA/JUMBF manifests, XMP AI flags, and encoder fingerprints in one pass.
- The pipeline injects a device profile (iPhone 16 Pro, Pixel 8 Pro, etc.) with GPS, timestamp, and phone encoder identity.
- Review the forensic proof card—see exactly what was removed and what was written in its place.
- Download the cleaned file and upload it to LinkedIn, Instagram, TikTok, or Reddit.
FAQ
Does re-encoding the video reset the AI detection signals?
Partially. Re-encoding disrupts some pixel-level patterns, but Lavc and x264 SEI fingerprints survive most transcode chains unless the codec path is fully broken. Calabi removes the encoder fingerprint at the bitstream level rather than relying on a downstream re-encode to mask it.
Can a platform still flag my content after cleaning?
No tool can guarantee zero detection—platforms evolve their models and results vary by source model and platform. Calabi removes the structural metadata signals (C2PA, XMP AI flags, encoder fingerprints) that are verifiably strippable. Perceptual hash-based detection may still trigger on some content depending on how the source model generates pixel patterns.
What device profiles are available?
Current profiles include iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. Each writes Make, Model, Software version, GPS coordinates, capture timestamp, and the corresponding encoder metadata that matches a real capture from that device.
If you are shipping AI content to platforms at volume, the metadata layer is not an edge case—it is the default detection path. Calabi handles it in one pass with a verifiable before/after report.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.