Trend report · gnews_detection · 2026-06-18
HYDAWAY DIGITAL COMPLETES INTEGRATION OF FULL MULITMODAL AI DETECTION SUITE ONTO HYDAWAY GPU INFRASTRUCTURE - PR Newswire Canada
AI Detection Is Now Infrastructure, Not an Afterthought
When Hydaway Digital announced it completed integration of a full multimodal AI detection suite onto its GPU infrastructure, the message was clear: automated AI content scanning is no longer a feature platforms might add later. It's a底层 layer running at scale, right now, on major platforms you post to every day.
Whether you're uploading a Sora clip, an AI-edited photo, or a generated still frame embedded in a video, platforms like Instagram, TikTok, YouTube, and Reddit are scanning every file within seconds of upload. The detection isn't guessing — it's reading specific, documented signals baked into your file's metadata and bitstream. If those signals say "AI-generated," your content can get flagged, shadowbanned, or rejected before a human ever sees it.
Understanding exactly what gets scanned — and why stripping that data and replacing it with authentic phone-capture identity is the only durable fix — is the difference between content that survives the upload and content that doesn't.
What Actually Flags Your File in 2026
Platform scanners don't look at your content the way a human does. They read invisible forensic signals embedded in the file structure. Here's what's actually being checked:
C2PA / Content Credentials (JUMBF manifests): This is the big one. C2PA embeds a cryptographically signed manifest — stored as JUMBF atoms — directly in compatible image and video files. It says, in machine-readable form, exactly which model generated the content, when it was created, and what tools were used. If your AI-generated file includes C2PA data, it's essentially walking through the platform's door with a label that says "made by AI." A single file can contain multiple JUMBF atoms and C2PA references. Hydaway's detection suite almost certainly scans for all of them.
XMP AI flags: Outside the C2PA layer, XMP metadata can carry a DigitalSourceType value of trainedAlgorithmicMedia — a direct AI-generation flag that predates C2PA adoption. Some exports also include Generator or Software tags identifying the AI tool by name. These fields survive re-saves and re-encodes in many cases.
Encoder fingerprints in video: Video files carry structural metadata about how they were encoded. AI-generated videos exported from tools using FFmpeg carry encoder strings like Lavc (FFmpeg's libavcodec) and x264 SEI messages that are structurally distinct from encoders found in real phone captures. These aren't metadata fields — they're embedded in the bitstream itself. Detection systems parse the codec signature the same way a forensic analyst would.
Missing capture metadata: A real phone recording in 2026 carries Make, Model, Software version, GPS coordinates, and a capture timestamp. An AI export carries none of these. The absence of these fields is itself a signal. Platforms weight this differently, but it's a consistent factor in multi-signal detection models like Hydaway's multimodal suite.
Perceptual hashes: Some platforms compare pHash or aHash values against known AI-generated content databases. These are less deterministic than metadata scanning — results vary by source model — but they factor into confidence scoring on platforms running multi-modal detection.
How Calabi Handles It: Strip, Inject, Verify
Calabi is a one-pass web tool that rebuilds your file's forensic identity from the ground up. It doesn't edit pixels, doesn't use inpainting, and doesn't touch the visible content. It operates entirely on the invisible signals and metadata that platforms actually scan.
Stage 1 — Strip: Calabi removes every detectable AI signal from your file. This includes all JUMBF / C2PA atoms and references (verified down to zero), XMP fields carrying trainedAlgorithmicMedia or generator tool tags, and encoder fingerprints in video bitstreams. A raw AI export with 144 metadata tags gets reduced to roughly 94 neutral structural tags — the kind any normal media file would carry.
Stage 2 — Inject: Calabi writes authentic phone-capture identity into the file. This includes a real device Make, Model, and Software version, GPS coordinates matching a plausible capture location, a 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 injected data is structurally consistent — every field you'd expect from a real phone recording is present and properly formatted.
Stage 3 — Verify: Before download, Calabi returns a forensic proof card showing exactly what was stripped and what was injected. This proof card is generated using ExifTool — the same forensic tool newsrooms, courts, and platform trust-and-safety teams use. You see the same scan the platform will run. If the proof card shows clean phone identity, your file is passing as a normal upload.
On visible watermarks: if your AI export has a visible logo or watermark in the frame — like Sora's sparkle mark — cropping removes the visible element. Calabi removes the invisible detection layer (the metadata and encoder signals) that survives cropping and re-encoding. That's the layer that actually gets you flagged on platforms that scan uploads automatically.
The Process: Upload to Clean File in Four Steps
- Upload your AI-generated video or image. Drag and drop or select the file. Calabi accepts most common formats. No account required to start.
- Automatic pipeline runs. Strip → Inject → Verify happens in one pass. No settings to adjust, no manual editing, no region selection.
- Review the forensic proof card. See exactly what was removed and what was injected. Compare the before/after metadata counts. This is the same ExifTool output platforms use to make detection decisions.
- Download the cleaned file. The file is rebuilt with clean phone identity and a neutral encoder signature. Post it directly to Instagram, TikTok, YouTube, or Reddit.
FAQ
Can Calabi remove a visible watermark like Sora's sparkle logo?
No — Calabi doesn't edit visible content pixel-by-pixel. A visible logo or watermark in the frame requires cropping to remove. Calabi handles the invisible detection layer — the C2PA manifest, XMP AI flags, and encoder fingerprints — that survives cropping and is what platforms actually scan for.
What if a platform uses perceptual hashing (pHash)?
Re-encoding disrupts some perceptual hash patterns, but results vary by source model and platform. Calabi fully removes the deterministic metadata and encoder signals, which are the primary detection layer on most platforms in 2026. Perceptual hash detection is less consistent and harder to defeat uniformly.
Does this work on videos exported from specific AI tools?
Yes. Calabi handles the encoder fingerprints (Lavc, x264 SEI) and metadata stripped from FFmpeg-based exports. The device profiles it injects are consistent with real captures from iPhone 15/16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra — the most common real-phone encoders platforms see at scale.
No tool can guarantee a platform won't flag any specific file. Results vary by platform, source model, and detection algorithm updates. Calabi removes the documented, deterministic signals that detection systems check — the C2PA layer, XMP AI flags, and encoder fingerprints — and replaces them with verified phone-capture identity.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.