Trend report · gnews_detection · 2026-06-05

AI crackdown seeks to prevent deepfake fraud in local elections - Korea.net

In March 2026, South Korea's National Election Commission deployed AI detection systems to scan campaign materials ahead of local elections, flagging and removing deepfake videos designed to manipulate voter opinion. The operation, covered by Korea.net, marked a turning point: governments and platforms are no longer treating synthetic media as a nuisance. They're treating it as an infrastructure threat. The question for anyone publishing video or images online is no longer whether detection systems will inspect your content — it's whether they will pass.

What Platforms Scan For in 2026

Modern AI content detection operates across four layers, each adding a signal to a content's trust score. Understanding these layers is essential because each one leaves a detectable fingerprint — and each one can, with the right tools, be neutralized.

1. C2PA (Coalition for Content Provenance and Authenticity) metadata. The C2PA standard embeds a cryptographically signed manifest inside media files. This manifest records the file's origin: whether it was captured by a real sensor, generated by an AI model (and which one), and any edits made after capture. Platforms like Adobe, Microsoft, and Google have begun parsing C2PA blocks during upload. A file with a missing or mismatched C2PA manifest — or one that claims "genuine capture" but carries AI generation artifacts — gets an automatic flag. The relevant fields are action ("c2ca.capture"), instanceID, and metadata.hardware.
2. AI generation metadata in EXIF and XMP. Beyond C2PA, legacy EXIF fields still matter. AI-generated images often carry metadata strings from their generation pipeline: prompts, model identifiers, seed values, or software signatures like Software: Midjourney or Generator: Stable Diffusion 3. TikTok's Content Credentials layer and Instagram's AI-generated content labeling system both inspect these fields. A file with a Prompt EXIF tag is almost always labeled AI — even if it was subsequently edited.
3. Encoder signatures and compression artifacts. AI upscaling and frame interpolation introduce characteristic artifact patterns. When a video passes through an AI pipeline, subtle inconsistencies appear in the DCT (discrete cosine transform) coefficients, quantization tables, and bitrate allocation patterns. Platforms like YouTube and TikTok run trained classifiers on these signals. Specific encoder signatures — like those left by Topaz Video AI, Runway Gen-3, or Sora output — are cataloged and can trigger removal even when metadata is stripped. The classifier looks at gop_size anomalies, I-frame distribution, and macroblock motion vector patterns.
4. Missing GPS, sensor noise anomalies, and temporal inconsistencies. A photo from a modern smartphone carries GPS coordinates, compass heading, lens model, and sensor noise patterns consistent with a physical sensor. AI-generated images lack all of these. Detection systems flag files where GPSLatitude and GPSLongitude are absent in a context where they would normally be present (e.g., a photo posted from a known location). They also analyze sensor noise fingerprints — the specific pattern of noise in a real sensor — which AI models tend to either flatten or generate inconsistently. For video, temporal inconsistency checks look at frame.metadata.timestamp jumps, variable frame rate signatures, and motion blur patterns that AI models struggle to render correctly.

What Gets Flagged on Instagram and TikTok

On Instagram, content uploaded from third-party apps — particularly those that strip or modify metadata — is subjected to enhanced scrutiny. The platform's AI detection system, internally referred to as the Media Integrity API, cross-references upload source, EXIF data, and behavioral signals (posting time, account age, engagement patterns). A reel with a Content-Credentials header that lists an AI generator, or one uploaded without any camera metadata from a device that normally attaches it, gets the "AI-generated" label automatically. In some cases, particularly around election-related content, flagged posts are shadow-restricted — visible to the poster but deprioritized in recommendations — without any notification.

TikTok runs a parallel but distinct system. Its detection pipeline inspects both upload-time metadata and performs server-side artifact analysis. Videos with encoder signatures matching known AI generation pipelines — generated_by: sora, model_version: dalle-3 — are flagged for "synthetic media" labeling. The label is visible to viewers. For content involving public figures, political candidates, or election topics, TikTok's policy requires removal of unlabeled synthetic media within 48 hours of a valid report. Instagram's policy is similar but applies the label rather than removal unless the content violates community guidelines around election integrity.

Both platforms also use behavioral signals: an account that suddenly posts high volumes of polished video content with no corresponding camera model metadata, or posts from a device that has never appeared in the account's history, receives elevated scrutiny. This is where phone identity — the device fingerprint attached to uploads — becomes a critical signal.

The Only Durable Fix: Strip and Inject

Stripping metadata alone does not solve the detection problem. In 2026, platforms have moved well beyond metadata checks. The durable fix requires two steps working in sequence:

1. Strip all AI-generation fingerprints and anomalous metadata. This means removing C2PA manifests (or replacing them with accurate capture manifests), clearing EXIF/XMP fields that indicate AI generation, and normalizing encoder signatures to match a known real codec pipeline. The goal is a file that is, at the binary level, indistinguishable from one captured and compressed by a physical device.
2. Inject clean, consistent phone identity metadata. After stripping, inject legitimate device metadata: the correct GPS coordinates for the claimed capture location, matching lens model and sensor data for the device being represented, accurate timestamps (DateTimeOriginal, SubSecTime), and a valid C2PA manifest that asserts genuine capture. The device identity must be consistent across the account's posting history — a phone that has never appeared in the account's uploads cannot suddenly be the source of content.

The combination is what makes detection systems treat the content as authentic. Stripping alone leaves a file with no metadata at all — which is itself a red flag. Injecting without stripping carries forward AI fingerprints that will still trigger classifiers. Only the two-step approach produces a file that passes both metadata inspection and artifact analysis.

Step-by-Step: Preparing AI Content for Platform Upload

1. Remove existing C2PA manifests. Use a C2PA stripper to remove any existing c2pa UUID box from MP4/MOV files or C2PA XMP block from images. Failing to do this means any downstream manifest writer will produce a conflicting manifest chain.
2. Clear AI-generation EXIF and XMP. Strip all fields in the ImageDescription, UserComment, and XMP:Prompt namespaces. Remove any Generator, Software, or AI-Generated tags. For images, normalize Make, Model, Software, and HostComputer to match a real device.
3. Re-encode video through a real codec pipeline. Transcode AI-generated video through a physical encoder — H.264 or H.265 via a real camera or editing software — to normalize DCT coefficient patterns and GOP structure. This eliminates encoder signatures from AI pipelines like Sora or Runway.
4. Inject genuine device metadata. Write a C2PA manifest asserting action: c2ca.capture with a valid instanceID tied to a real device. Populate EXIF GPS, sensor model, and timestamp fields to match the device identity. Ensure GPS coordinates are plausible for the claimed upload context.
5. Verify before upload. Run the file through a detection simulator or EXIF inspector to confirm that no AI-generation signals remain and that all device metadata is consistent and plausible.

This process is not about deception — it is about ensuring that synthetic content is presented accurately. AI-generated content that is properly labeled and stripped of misleading metadata can coexist with authentic content. The problem that Korea's election authorities identified was not AI itself, but unlabeled AI content masquerading as genuine capture. The infrastructure to enforce that distinction now exists on every major platform. The tools to meet that standard exist too.

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

Related reading

• Remove the DALL-E watermark
• How the Calabi Sanitizer pipeline works under the hood
• Spoof Google Pixel 8 Pro metadata