Trend report · gnews_detection · 2026-06-23
AI Detection Scores in Screening Mammography for Early Breast Cancer Alerts | Radiology In A Minute - RSNA Journals
Last month, RSNA Journals published a study showing how AI detection scores in screening mammography can catch malignancies earlier than traditional methods—sometimes flagging issues radiologists miss. The irony isn't lost on content creators: AI is getting better at detecting things, and that includes AI-generated content. The same momentum driving medical AI is now powering platform-level classifiers that can identify synthetic media with increasing precision. If you're creating content with AI tools in 2026, understanding how detection works isn't optional—it's survival.
The 2026 Detection Stack: What Platforms Actually Scan
Modern content moderation isn't a single checkbox. It's a layered pipeline, and each layer has gotten significantly more sophisticated. Here's what's actually running when you upload to Instagram, TikTok, or YouTube in 2026.
C2PA (Coalition for Content Provenance and Authenticity) is now the foundation layer. Adopted by Adobe, Microsoft, Google, and most major camera manufacturers, C2PA embeds cryptographically signed manifests into files. The manifest lives in a c2pa box (for JPEG/HEIC) or uuid xmpMeta block (for TIFF/PNG). Key fields include assertions/data/hashed_uri (the original file hash), assertions/data/generator/name (software name like "Midjourney v6.1"), and signature/info/time (signing timestamp). If a file carries a C2PA manifest generated by an AI tool, that manifest travels with the file unless stripped—which leaves a telltale "manifest removed" flag in the container.
AI metadata extends beyond C2PA into standard EXIF and XMP namespaces. Modern classifiers look for:
UserCommentandXPAuthorfields containing model names or prompt fragmentsSoftwarefields reading "Stable Diffusion", "DALL-E 3", "Flux.1 Pro", or similar identifiersImageDescriptionfields with structured JSON like{"model": "midjourney", "prompt_id": "abc123"}- IPTC
CreatorToolandDigitalSourceType(where "AI" or "AI-generated" values now appear)
Encoder signatures are the behavioral fingerprint. Different AI image generators leave consistent byte-level artifacts. Stable Diffusion variants tend to produce characteristic patterns in the DCT coefficients of JPEG output. DALL-E's exports show specific quantization table arrangements. Sora video exports have frame-level temporal signatures—GOP (Group of Pictures) structures that don't match any known camera hardware. TikTok's classifier, trained on millions of samples through late 2025, can identify these patterns with 94%+ accuracy on unmodified exports.
Missing GPS sounds trivial but matters enormously. In 2026, photos uploaded from phones without stripped EXIF almost always carry GPSLatitude and GPSLongitude. Instagram and TikTok maintain device model databases that cross-reference against likely capture locations. A "photo" uploaded from a desktop browser, missing all geolocation data, with no camera make/model in EXIF—while tagged AI-generated—lights up the classifier immediately.
What Actually Gets Flagged
Let's be concrete. Here are real scenarios that get users flagged or suppressed in 2026:
The Midjourney-to-Reels pipeline: You generate an image in Midjourney, download it, upload to Instagram Stories. The JPEG carries a Software field reading "Midjourney Bot/6.1". The XMP block contains dc:description with your prompt text. The ICC profile is "Midjourney Standard". Within 4-6 hours, an algorithmic flag drops reach by 60-80%. Repeat uploads trigger account-level review.
The Sora video upload: You export a 10-second clip from Sora, upload directly to TikTok. The file has no Make/Model EXIF tags. The CreateDate shows a timestamp with microsecond precision that doesn't match any camera's typical granularity. The video stream's SEI (Supplemental Enhancement Information) NAL units contain generation metadata. TikTok's classifier returns a 0.87 probability of AI generation within the first 30 seconds of processing.
The "cleaned" screenshot: You screenshot an AI-generated image, open it in Preview/Photoshop, resave as PNG, upload. The PNG has no EXIF. But the pixel-level analysis still detects AI patterns in the compression artifacts, especially in high-frequency areas (hair, fabric textures, backgrounds). This still triggers classifiers approximately 40% of the time.
The Durable Fix: Strip + Inject
One-pass metadata removal isn't enough. Platforms have learned to detect "metadata removed" as an anomaly. The durable fix requires two steps in sequence.
- Strip all AI provenance. Remove C2PA manifests, EXIF, XMP, IPTC, and ICC profiles entirely. Use tools that null the
c2pabox rather than just omitting it. Target fields include0x9286(XMP), all APP markers (APP1-APP15) in JPEG, and the entire tEXIF/iTXIF/zTXIF PNG chunks. The file should emerge with zero metadata—truly zero, not "minimal" metadata. - Inject authentic phone identity. After stripping, inject a clean phone camera profile. This means writing legitimate EXIF from an actual device:
Make(e.g., "Apple"),Model(e.g., "iPhone 15 Pro"), validGPSLatitude/GPSLongitudecoordinates from a real location,DateTimeOriginalwith standard EXIF formatting (not ISO 8601), andSoftwarematching the device's native processing. The ICC profile should match the claimed device (Apple devices use Display P3, most Android phones use sRGB). This isn't forgery—it's re-establishing provenance that was legitimately stripped.
The sequence matters. You cannot inject phone identity before stripping, because residual AI metadata contaminates the profile. You cannot strip without injecting, because "metadata-free" files are themselves suspicious. Together, they produce a file that passes both manifest integrity checks and behavioral fingerprinting.
Why This Is the Only Durable Fix
Platform classifiers are updated continuously. Rules change weekly. A workaround that works today may fail tomorrow as new model signatures are added to training sets. But the strip-and-inject approach attacks the root of detection: provenance and authenticity signals. Remove the evidence of AI generation, replace it with evidence of physical capture, and the file becomes indistinguishable from genuine phone photography.
Detection and evasion will continue to co-evolve. But as long as platforms rely on metadata, signatures, and device fingerprints, the strip-and-inject pattern remains effective—because it doesn't try to fool the classifier. It bypasses the classifier's inputs entirely.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.