Trend report · gnews_detection · 2026-06-09
SDG&E Launches AI Wildfire Detection Project - Newsradio 600 KOGO
In breaking news from San Diego, SDG&E has launched an ambitious AI wildfire detection project that uses computer vision to identify smoke patterns and heat signatures before fires spread. It's a reminder that artificial intelligence isn't just generating content—it's increasingly being deployed to detect it. That dual reality is reshaping how platforms like Instagram, TikTok, and YouTube handle authenticity in 2026.
The Detection Stack: What Platforms Actually Scan
Modern AI content detection operates on multiple layers, and understanding each one is essential for anyone creating or distributing media online.
C2PA (Coalition for Content Provenance and Authenticity) is the most significant structural change. C2PA embeds cryptographically signed metadata into images, video, and audio at the point of creation. Cameras, editing software, and AI generation tools that support C2PA add a manifest block containing fields like actions (describing how the content was created or modified), assertions (including stds.schema.org:CreativeWork with fields like authorName and generator), and signatureInfo (with issuer, time, and signer). Platforms including Adobe, Microsoft, Google, and Intel have adopted C2PA. When you upload media, Instagram's classifier checks for the presence of a valid C2PA signature block. If the manifest is missing on content that should have it—or if the signature fails verification—the upload enters a secondary review queue.
AI metadata flags go beyond C2PA. Platforms parse EXIF, XMP, and IPTC headers looking for tags like Software: Adobe Firefly, Generator: DALL-E 3, or AiGenerated: True. TikTok's moderation system specifically looks for the Generator field in XMP data. Instagram scans for embedded prompt strings that some AI tools insert into metadata. Even if you strip obvious markers, trailing artifacts in the metadata structure often reveal the generation tool.
Encoder signatures represent a subtler detection vector. Each video encoder—whether it's H.264, H.265, or AV1—leaves statistical fingerprints in the compressed bitstream. AI-generated video tends to exhibit specific compression anomalies: quantized DCT coefficient distributions that differ from natural footage, motion vector patterns that lack the micro-jitter of handheld capture, and GOP (Group of Pictures) structures that don't match typical camera behavior. YouTube's Content ID adjacent systems and TikTok's upload classifier both perform encoder fingerprint analysis as a first-pass filter.
Missing GPS and sensor data is a surprisingly effective red flag. Modern smartphones embed GPS coordinates, accelerometer readings, gyroscope data, and lens metadata in every photo. When content arrives without any GPSPosition or AccelerometerData fields, platforms infer a higher probability of AI generation or heavy post-processing. Instagram's Stories detection specifically flags images where expected sensor metadata is absent or inconsistent with the claimed capture device.
What Actually Gets Flagged
Concrete examples make this tangible. A photographer using Lightroom to edit an AI-generated base image will likely trigger flags if the original C2PA manifest from the AI tool remains intact, because the actions chain will show generation followed by editing without proper re-signing. The platform sees an unsigned discontinuity in provenance.
On Instagram, resharing an AI-generated image without modification often results in the "AI-generated content" label being automatically applied, even if you didn't create it. TikTok applies this label when the upload classifier detects statistical signatures in the first 30 frames of video consistent with video generation models. YouTube is more aggressive on the upload side: content missing expected C2PA manifests from known AI generation tools enters manual review.
The pattern is consistent: platforms are building interlocking checks rather than relying on any single signal. Absence of C2PA isn't itself a flag—but absence of C2PA combined with missing GPS, known encoder artifacts, and AI-associated metadata fields creates a high-confidence detection.
The Durable Fix: Strip and Re-Inject
Stripping all AI metadata and encoder signatures is necessary but not sufficient. Platforms have learned to flag content that arrives with suspiciously clean metadata—content that should have GPS but doesn't, or that lacks any embedded software signatures. The durable solution requires a two-step process: comprehensive stripping followed by injection of authentic sensor identity from a clean device.
Here is the specific step-by-step workflow:
- Strip all metadata comprehensively. Remove EXIF, XMP, IPTC, and any C2PA manifest blocks. This includes the
xmpMM:DocumentID,photoshop:Links4, anddc:formatfields that many tools preserve. Use a hex-level stripper that rewrites the file from decoded media rather than merely truncating header sections. - Remove encoder fingerprints. Re-encode the video through a different codec pipeline than the original. Transcode from H.265 to H.264, or from AV1 to VP9, allowing the new encoder to write fresh motion vectors and GOP structures. This breaks the statistical fingerprint chain.
- Inject clean sensor identity. Take metadata from a genuine capture—GPS coordinates, accelerometer calibration strings, lens profile data, and device make/model fields—from a device that wasn't used to generate AI content. Inject this as a complete, self-consistent metadata block. The
GPSLatitude,GPSLongitude,GPSAltitude,Make,Model, andLensModelfields must be internally consistent with each other and with the GPS coordinates provided. - Add C2PA manifest if appropriate. For content intended for platforms that require provenance, sign the content with a legitimate C2PA manifest from a camera or editing tool that supports it. This creates a verifiable chain of custody that satisfies platform requirements.
- Verify before upload. Run the file through a detection emulator that simulates platform classifiers. Check that C2PA verification passes, that metadata fields are internally consistent, and that encoder fingerprints match expected patterns for the claimed capture device.
Why Strip-and-Inject Is the Only Durable Approach
Platforms update their detection models continuously. A solution that works today by patching specific metadata fields will break when new detection layers are added. The strip-and-inject approach is durable because it addresses the fundamental detection logic: content must have verifiable, consistent provenance that matches expected sensor data for its claimed origin. By providing authentic sensor identity from a clean device, you create metadata that will pass future detection iterations—not just today's classifiers.
SDG&E's wildfire AI watches for anomalies in smoke and heat patterns. Platform AI watches for anomalies in content provenance. The defense in both cases is the same: create signals that are internally consistent, historically plausible, and structurally indistinguishable from authentic sources.
The tools and techniques continue to evolve, but the principle holds: genuine identity, properly embedded, is the only metadata that survives scrutiny across all detection layers.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.