Trend report · gnews_celebrity · 2026-05-31

Bombay High Court flags concerns over AI-generated celebrity simulations during Shilpa Shetty personality rights hearing - Storyboard18

When the Bombay High Court raised alarms about AI-generated simulations of Shilpa Shetty during a personality rights hearing last week, it underscored a legal frontier that platforms are still scrambling to map. The court's concerns weren't hypothetical—AI tools can now produce photorealistic likenesses of public figures, complete with voice timbre and gestural patterns, in minutes. What's emerging is a split between jurisdictions that haven't caught up to the technology and the detection infrastructure trying to slow it down. This piece examines how content moderation systems actually identify AI-generated content in 2026, what survives the current detection layers, and why stripping metadata plus injecting fresh device identity is the only durable workaround for creators who need their work to pass through.

What Platforms Scan For in 2026

Modern detection pipelines are more layered than most people assume. They don't just look for "fake images"—they inspect specific metadata fields and structural signatures that honest creators often overlook. Here's what actually gets checked:

C2PA Manifests and Provenance Chains

The Coalition for Content Provenance and Authenticity standard has matured significantly. C2PA embeds cryptographically signed manifests directly into image files using the application/x-c2pa content type, storing claims in a JSON structure under the c2pa.actions field. When a file passes through an AI generation pipeline—Stable Diffusion, Midjourney v6, Sora—tools that support C2PA write entries like ClaimGenerator and Edits that disclose the software origin. Platforms like Meta and TikTok now check for valid C2PA chains as a first-pass filter. Files without C2PA metadata or with broken chains get flagged for secondary inspection. The field c2pa.signature_info containing the issuer certificate is specifically audited—if the issuer isn't on an approved trust list, the content is treated as unverified.

AI Metadata Fields

Even without C2PA, AI tools leave breadcrumbs. EXIF fields like Software, Artist, or custom XMP namespaces often contain references like Stable Diffusion, DALL-E 3, or model version strings. Detection systems parse these fields directly. The gen_ai XMP flag—one of the standardized fields under the IPTC Photo Metadata standard—gets set by compliant AI generators and is checked by platforms that have adopted the 2025 guidelines. Anything with a populated AIGenerationTool or Prompt field under XMP gets rerouted to a secondary review queue. This is where most consumer-generated AI content gets caught, particularly images exported directly from Midjourney or Leonardo.ai without post-processing.

Encoder Fingerprints and Model Watermarks

Missing or Inconsistent GPS/EXIF Data

For video content especially, platforms cross-reference the GPSLatitude, GPSLongitude, and GPSAltitude EXIF fields against the content itself. A video claiming to be from Mumbai but lacking GPS data—or worse, showing GPS metadata from a Beijing-based server timestamp—gets flagged for geographic inconsistency. The Make and Model device fields are checked against known device lists; content generated by AI tools without any device attribution gets scored higher for synthetic origin. This is why stripped files often fail—they lose the legitimate device metadata that makes content look authentic in the first place.

What Actually Gets Flagged on Instagram and TikTok

Based on documented enforcement patterns and creator reports through 2025, here's what the platforms act on:

Content with active C2PA manifests naming AI generation tools—Instagram's AI label automation applies automatically and cannot be removed once attached
Videos with missing DeviceMake, DeviceModel, and DateTimeOriginal fields that exceed the platform's threshold for synthetic content probability
Images showing the characteristic checkerboard artifact patterns common to SDXL-generated outputs, detected through the platform's perceptual hashing pipeline
Audio tracks containing synthetic voice patterns flagged by TikTok's Content Authenticity Initiative compliance checks
Any content with conflicting metadata—C2PA claiming human authorship while EXIF shows AI software

The enforcement varies by region. Content in India, where the Bombay High Court ruling creates precedent, is now being escalated for human review more frequently than content from other markets, according to platform transparency reports. Creators posting AI-adjacent content from Indian IP addresses report higher rejection rates on content that passes without issue from US-based accounts.

Stripping and Injecting: The Durable Fix

The core problem is that stripping alone doesn't solve the detection challenge—it just removes legitimate provenance while leaving the statistical fingerprint. That's why stripping plus injecting fresh device identity is the approach that survives scrutiny.

Step-by-Step: How to Clean AI Content for Platform Upload

Strip all AI-era metadata — Remove EXIF, XMP, and ICC profile data entirely using a metadata stripper. Target the fields Software, Artist, Make, Model, GPSLatitude, and all XMP namespaces including xmpMM, stEvt, and gen_ai. This eliminates the AI tool fingerprint from the file header.
Remove C2PA manifests — Use a C2PA-aware stripper to clear the application/x-c2pa content and any associated manifests. If the manifest cannot be cleanly removed, re-encode the image through a trusted pipeline—PNG recompression or WebP transcoding—while preserving visual quality.
Inject authentic device identity — Rather than leaving the file metadata-empty, inject fresh metadata that matches a real device. Use fields like Make, Model, and DateTimeOriginal from known device profiles—iPhone 15 Pro, Sony A7IV, Samsung Galaxy S24 Ultra. The injected GPSLatitude and GPSLongitude should correspond to a plausible location consistent with the content.
Add consistent ICC profiles — Embed a standard ICC color profile (sRGB IEC61966-2.1 for web content) to prevent detection pipelines from flagging unusual color space inconsistencies that AI generators sometimes produce.
Verify before upload — Use a metadata inspector to confirm the file shows the injected fields, no AI tool references, and a valid geographic location. Check that C2PA claims have been cleared or that no C2PA data remains.

The injection step is critical. A file with no metadata is itself suspicious to modern detection systems. Content that shows the metadata profile of a legitimate photographed image—complete with device make, lens information, and GPS—passes the consistency checks that flag stripped AI content.

Why This Matters Now

The Bombay High Court's intervention in the Shilpa Shetty case signals that legal systems globally are moving toward active enforcement against AI-generated celebrity likenesses. Platforms face pressure to demonstrate they can detect and remove this content. That enforcement ripples to all AI-generated content, even content that's legitimate—art, satire, branded content created by human artists using AI tools. The detection infrastructure is blunt by design: it flags the category, not the intent. Creators who want their work to reach audiences need to understand the technical layers that determine whether content gets labeled, suppressed, or removed.

Understanding C2PA manifests, encoder fingerprints, and metadata injection isn't just technical trivia—it's the practical knowledge that determines whether your content survives platform moderation in 2026.

→ Try Calabi free at calabilabs.com — 3 cleans, no card.

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

Try free →