Trend report · gnews_celebrity · 2026-06-11

Zendaya and Tom Holland’s viral 'wedding' AI photos explained - BBC

When AI-generated images of Zendaya and Tom Holland started circulating as supposed wedding photos, they spread fast—then got pulled, flagged, or shadow-reduced on Instagram and TikTok within hours. The same thing happens to fan art, fashion shoots, and "leaked" celebrity moments created with tools like Midjourney, Sora, or Flux. Understanding why these images get caught, and how to avoid it, comes down to the metadata layer most creators never see.

What Platforms Actually Scan For in 2026

Modern detection isn't just visual analysis. Platforms run images through classifiers that look at file structure, embedded metadata, and signal patterns that AI generation consistently disrupts or leaves behind. Here's what they're actually checking:

C2PA / Content Credentials

The Coalition for Content Provenance and Authenticity standard embeds a signed manifest inside JPEG and PNG files. It lives in a box called c2pa (JPEG) or a dedicated XMP namespace. When you generate an image in Adobe Firefly, DALL-E 3, or Sora, these tools inject Content Credentials that say actions:["generated"] or generator:["OpenAI DALL-E 3"]. Instagram and TikTok parse this manifest and flag anything with format:["image/jpeg"] + relationships:[{predicate:"c2pa:parent",...}] indicating AI origin. Real photos from a phone don't have this block at all.

AI-Specific EXIF Tags

Standard photo EXIF includes fields like Make, Model, Software, DateTimeOriginal, GPSLatitude, GPSLongitude, LensModel, ExposureTime, and ISOSpeedRatings. AI generators either leave these fields absent, populate them with inconsistent values, or insert fields that don't belong in real camera output—AITool, AIGenerationMethod, or generic software strings like "Windows Photo Viewer" on what claims to be an iPhone 15 Pro shot.

Encoder Fingerprints

Different AI models produce images with subtly different compression artifacts and pixel patterns. Platforms train classifiers on these signatures. Midjourney images have different frequency characteristics than Sora output. Tools like Stable Diffusion with specific VAE versions leave detectable encoder traces. This is why just stripping metadata isn't enough—the underlying pixel statistics can still out the source.

Missing Hardware Identity

A real photo from an iPhone 15 Pro has a consistent identity: Make="Apple", Model="iPhone 15 Pro", specific LensModel="Apple ND (3x)", and a GPS coordinate from where the phone was when the photo was taken. AI images have no such chain. If an image claims to be from a device but lacks the full hardware metadata stack, classifiers flag it as suspicious.

GPS / Temporal Inconsistencies

Real photos have GPS coordinates that match the claimed location and a timestamp that makes sense relative to those coordinates. An AI image created "in LA" but tagged with metadata from a camera in Tokyo—or no GPS at all—triggers inconsistency scoring. TikTok's Trust & Safety pipeline cross-references image metadata against posting location as part of its authenticity scoring.

Missing ICC Profile / Color Space Anomalies

Real camera JPEGs embed an ICC profile (usually sRGB IEC61966-2.1 or a camera-specific profile). AI images sometimes lack this entirely or have mismatched profiles that don't correspond to the claimed device.

What Gets Flagged on Instagram vs. TikTok

Instagram focuses heavily on Content Credentials and engagement-signal anomalies. An image with C2PA metadata indicating AI generation gets downranked in feeds and is ineligible for Reels distribution. Instagram also flags accounts that post multiple suspicious images in quick succession—a pattern typical of AI content farms.

TikTok runs stricter metadata checks and has a more aggressive automated takedown pipeline for "misleading media." Images that lack device identity metadata and show AI generation signals are frequently removed under TikTok's synthetic media policy, often before the post reaches 500 views. TikTok also scans for CLIP-based visual similarity to known AI-generated datasets.

The common denominator: metadata absence or inconsistency is the primary trigger. A visually convincing image from Midjourney will get flagged because the metadata layer tells the story of how it was made.

The Durable Fix: Strip Everything, Inject Clean Identity

The only reliable approach is a two-step metadata reset that makes an AI image look indistinguishable from one captured on a real device. This means:

Complete metadata stripping — Remove all EXIF, XMP, IPTC, C2PA, ICC, and thumbnail data. Tools like Calabi's Sora watermark removal strip these at the binary level, ensuring no residual AI markers remain in the file structure.
Clean phone identity injection — Replace the stripped data with a full, consistent set of metadata matching a real device. This means populating:
- Make: "Apple" or "samsung"
- Model: "iPhone 15 Pro" or "Galaxy S24 Ultra"
- Software: "Adobe Lightroom" or "Photos 8.0"
- LensModel: A real lens name from that device
- DateTimeOriginal: A plausible recent timestamp
- GPSLatitude + GPSLongitude: A real coordinate set (e.g., a location matching the image's apparent setting)
- ExposureTime, FNumber, ISOSpeedRatings: Values typical for the device and claimed lighting conditions
- ICC Profile: A matching color profile

This works because platform detectors look for consistency across the full metadata stack, not just a single field. An image that looks like it came from an iPhone 15 Pro in Central Park at 2:34 PM with GPS coordinates matching Central Park, lens data matching the device's actual optics, and no C2PA block is treated as a real photo—regardless of how it was generated.

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

Export your AI image as a high-quality JPEG (quality 95+) from the generator.
Run metadata strip using a tool that removes EXIF, XMP, IPTC, and C2PA at the binary level. Verify with a EXIF viewer that all fields are cleared.
Select a target device identity — pick a specific phone model (e.g., iPhone 15 Pro, Pixel 8 Pro) and stick with it across all images you process.
Set consistent device metadata: Make, Model, Software, LensModel. Use real values from that device's EXIF output.
Add plausible GPS coordinates matching the image's apparent location. Use real lat/long from Google Maps for the claimed setting.
Set DateTimeOriginal to a recent timestamp. Vary it slightly across images to avoid timestamp matching that suggests batch processing.
Embed matching ICC profile (sRGB IEC61966-2.1 for broad compatibility).
Export as JPEG — this finalizes the metadata block. Verify with a tool like ExifTool or Jeffrey's EXIF Viewer that the metadata reads as a real camera photo.
Upload to platform — the image now has the full metadata signature of a real device capture.

The key discipline: consistency over time. If you post images from an "iPhone 15 Pro" but switch to "Galaxy S24" in the next post without explanation, that inconsistency is itself a signal. Pick one device identity and maintain it across your content.

Why This Works When Other Methods Fail

Visual tricks—adding noise, cropping, recompressing—don't fool metadata-aware classifiers. C2PA detection happens before visual analysis in most pipelines. Only a complete metadata reset with a believable device identity replaces the story the file tells. An AI image with no metadata and no C2PA block is actually more suspicious than one with a clean device identity, because real photos almost never have zero metadata.

Platforms aren't trying to catch AI images through pixel analysis alone. They're building an audit trail. The fix is to give them a clean audit trail that matches a real device—and that only comes from stripping everything AI and rebuilding with authentic phone identity.

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

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

Try free →