Trend report · gnews_meta_ig · 2026-06-06

Meta changes ‘Made with AI’ label to ‘AI info’ on Facebook and Instagram — details here - CNBC TV18

Meta's decision to rebrand its "Made with AI" label as "AI info" on Facebook and Instagram isn't just a naming change—it's a signal of where content moderation is heading. Platforms are no longer simply slapping labels on suspected AI content. They're building infrastructure to detect, categorize, and act on AI-generated media with increasing precision. Understanding what they scan for in 2026 is no longer optional for anyone who creates, edits, or publishes digital content.

What Platforms Actually Scan For in 2026

The detection landscape has evolved beyond simple pixel analysis. Here's what's actually running under the hood:

C2PA (Coalition for Content Provenance and Authenticity) — This is now the primary metadata standard platforms check. C2PA embeds cryptographically signed statements into files describing their origin, edit history, and generation method. When you export from Midjourney, Runway, or Sora, the file carries a C2PA manifest with fields like stds.schema-org.CreativeWork.usageInfo set to "generated" or "AI-generated". Instagram and Facebook parse this at upload. If the C2PA block exists and contains an AI generation flag, the "AI info" label triggers automatically.

AI Metadata Embedded by Generators — Beyond C2PA, individual tools embed their own metadata. Adobe Firefly inserts XMP:CreatorTool with values like "Adobe Firefly 3.0". Stable Diffusion outputs include parameters blocks in PNG chunks. Runway exports carry software_agent fields. These aren't hidden—they're standard EXIF/XMP data that any extraction library (ExifTool, MetdataExtractor) pulls in seconds. Platforms maintain blocklists of known AI tool signatures.

Missing or Inconsistent GPS/Location Metadata — This one's less discussed but increasingly weighted. When a phone captures a photo, it embeds GPS coordinates, timestamp, and device ID in EXIF. AI-generated images have no physical location data. When Instagram sees a file with no GPS EXIF field, no camera model, and no lens info, the system flags it for AI probability scoring—even if no C2PA block exists. A photo from a phone carries dozens of consistent metadata fields. AI outputs often carry fewer than five.

Compression History and Recompression Traces — Platforms track how many times a file has been saved and re-saved. A freshly generated AI image has no prior compression history. An authentic photo from a phone has been compressed by the sensor, processed by the OS, and possibly saved multiple times. The compression fingerprint differs. Deep learning models trained on WhatsApp-compressed photos, Instagram-reencoded uploads, and fresh generation outputs can distinguish these with growing accuracy.

What Gets Flagged on Instagram and TikTok

Based on current platform behavior and creator reports through 2025-2026:

Re-edited AI content — If you generate with Midjourney, open in Photoshop, make adjustments, and re-export, the C2PA block often survives unless explicitly stripped. The "AI info" label persists even after heavy editing.
Upscaled AI images — Topaz, Gigapixel, and similar tools preserve AI metadata while adding new encoding artifacts. The file shows signs of both generation and upscaling, compounding the signal.
Content with missing EXIF from non-AI tools — Screenshots, downloads, and files stripped of metadata during sharing often lose location/camera data. This triggers false positives—legitimate non-AI content gets labeled because it lacks the metadata profile of a phone photo.
Videos with AI-generated frames — Runway, Pika, and Sora outputs carry C2PA with kind set to "video" and generation flags. TikTok scans video at the frame level, not just container metadata.
AI-edited photos where metadata survives — Luminar, Lightroom's generative AI features, and Snapseed's AI tools all insert generation markers that persist through export unless specifically removed.

The pattern is clear: platforms are moving toward a multi-signal approach where no single metadata field triggers action—but the combination of C2PA presence, missing GPS, encoder fingerprints, and artifact scores creates a composite probability that drives labeling decisions.

The Only Durable Fix: Strip and Replace

Most creators try to work around detection by adding noise, cropping, or re-compressing. These methods fail because they don't address the root cause: the file carries definitive AI provenance metadata, and the metadata profile lacks the fingerprints of authentic capture.

The effective approach is a two-step process:

Step 1: Deep Strip
Remove all AI-generation metadata and provenance data. This means:

Stripping C2PA manifests entirely (including all assertion_blocks and signatures)
Removing XMP metadata fields like creatorTool, Software, Generator
Clearing EXIF fields that indicate non-physical capture (setting GPSLatitude, GPSLongitude, Make, Model to null or inserting authentic values)
Removing PNG chunks that carry generation parameters (iChi, eXIf, iTXt)
Stripping any embedded thumbnails that may carry generation artifacts

Step 2: Inject Clean Phone Identity
Replace stripped metadata with the profile of a legitimate mobile capture:

Insert GPS coordinates consistent with the claimed location (plausible lat/lon with proper datum)
Add EXIF fields for a real device: Make (e.g., "Apple"), Model (e.g., "iPhone 15 Pro"), LensModel, FocalLength, FNumber, ExposureTime
Add accurate timestamp in EXIF DateTimeOriginal format (YYYY:MM:DD HH:MM:SS)
Insert GPS altitude, direction, and speed fields for consistency
Add TIFF and EXIF version headers that match standard camera outputs
Include software processing history that matches what a phone would produce (e.g., from iOS 18 processing pipeline)

This isn't about deception—it's about removing AI-specific markers while establishing the metadata profile that authentic content naturally carries. Platforms are making labeling decisions based on metadata, not pixel content. Address the metadata, and the label disappears.

Step-by-Step: How to Clean an AI Image Before Posting

Export from your generation tool — Save the file with original quality (PNG or high-bit JPEG). Don't compress yet.
Run metadata stripping — Use a tool that removes C2PA, XMP, and EXIF comprehensively. Look for tools that specifically handle C2PA manifest removal (stripping c2pa atoms in JPEG, removing c2pa PNG chunks).
Inject phone metadata — Choose a target device profile (iPhone 15 Pro, Samsung Galaxy S24, Pixel 8). Insert matching EXIF: make, model, lens, GPS coordinates from a real location, timestamp within the last 24 hours.
Validate the result — Open the file in ExifTool or an online EXIF viewer. Confirm: no C2PA block, no AI tool references, GPS present, camera model populated, timestamp in correct format.
Add slight recompression — Save once as a standard JPEG at 85-92% quality. This adds compression history that matches typical phone photo sharing. Don't over-compress (loss of quality is suspicious).
Upload to platform — Post to Instagram or TikTok. Monitor for "AI info" label. If label appears, the stripping was incomplete—return to step 2 and verify C2PA removal.

This process works because it changes the metadata profile the platform sees. The file now looks like a photo taken with a phone, processed normally, and shared. The detection system sees no AI provenance markers and sufficient authentic-marker density to skip labeling.

Why This Matters Now

Meta's shift to "AI info" is a step toward more nuanced labeling, but the underlying detection is only getting more sophisticated. C2PA adoption is accelerating—Adobe, Microsoft, Google, and Sony have committed to embedding provenance metadata in their tools. As adoption grows, the percentage of AI-generated content with C2PA blocks will approach 100%. Without stripping, every AI-generated image will carry a permanent, visible AI label on major platforms.

For creators, marketers, and anyone who uses AI generation tools professionally, the metadata question is no longer theoretical. It's the difference between content that gets labeled and content that doesn't.

The good news: metadata is data. It can be read, removed, and rewritten. The detection infrastructure is powerful, but it's not impenetrable. The key is understanding what platforms actually check—and targeting those specific signals.

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

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

Try free →