Trend report · hn_ai · 2026-06-12
Show HN: LinkedIn Unfiltered – Turn AI slop into readable posts
The rise of AI-generated content has created an unexpected side effect: platforms are getting remarkably good at detecting it. Tools like LinkedIn Unfiltered, which transform AI slop into readable posts, highlight a growing tension between synthetic content and platform trust systems. But there's a deeper problem lurking beneath the surface—one that affects anyone trying to publish AI-assisted work without getting flagged, throttled, or shadowbanned.
In 2026, content moderation isn't just about text anymore. It's about provenance. Here's what platforms actually scan for, and more importantly, how to address the detection vectors that matter.
What Platforms Scan For in 2026
Modern AI detection operates at the metadata and signal level, not just visual analysis. Here's the technical stack platforms are running:
C2PA: The Content Provenance Standard
The Coalition for Content Provenance and Authenticity (C2PA) has become the industry baseline. C2PA embeds cryptographically signed metadata into images and videos, indicating their origin. If an image lacks C2PA provenance data, or if the data claims it came from an AI generator without proper signing, that's a flag.
Key fields platforms check:
- uapf:hash – A content hash that changes if the image is modified
- c2pa.actions – A chain showing the image's creation history
- c2pa.assertions[].label – Specifically looks for "stds.schema-org" or "c2pa.ai-generated" assertions
- xmpGImg:guide – Some generators add this to indicate AI involvement
When an image comes from Midjourney, DALL-E 3, or Sora, it typically carries PNG ztxt chunks or XMP metadata identifying the generator. Platforms parse these with libraries like exempi or custom parsers. Absence of expected C2PA blocks—or presence of AI-specific ones—triggers review.
Encoder Signatures: The Invisible Fingerprint
AI image generators don't just produce pixels; they produce pixels with specific statistical fingerprints. Convolutional neural networks trained on diffusion models leave detectable patterns in:
- Frequency domain artifacts – Fourier transforms reveal consistent high-frequency noise patterns that differ from camera sensors
- DCT coefficient distributions – JPEG compression artifacts from AI generators cluster differently than natural photographs
- Color histogram anomalies – AI images often show unnatural uniformity in certain color channels
Platforms like TikTok run these through classifiers trained on millions of AI-generated images. The accuracy varies, but false positives against AI-assisted photography are increasingly common.
Missing GPS and EXIF Patterns
Here's a subtle one: authentic smartphone photos carry specific EXIF patterns. A freshly captured iPhone 15 Pro image contains:
- GPSLatitude/GPSLongitude – Real coordinates from the GPS chip
- MakerApple – Identifies the device manufacturer
- ExifIFD:FNumber – Lens aperture from the optical system
- Device properties – Serial numbers, lens identifiers, firmware versions
AI-generated images, even when stripped, often lack these fields entirely or contain placeholder values. Platforms have learned to flag images with suspicious EXIF gaps—not just missing GPS, but missing all device metadata. A 2400x1800 image with no EXIF data whatsoever screams "not a real photo" to modern classifiers.
What Gets Flagged on Instagram and TikTok
Based on platform enforcement patterns observed through 2025-2026:
Instagram typically triggers review when:
- A newly uploaded image contains AI-specific PNG chunks (look for "parameters" or "sd-metadata" in ztxt)
- The C2PA manifest claims "GenerateTool" without a valid signing certificate
- Image dimensions match common AI output sizes (1024x1024, 512x512, 1792x1024)
- No EXIF data and no apparent compression history
TikTok is more aggressive with video. It flags when:
- Video frames show frequency artifacts consistent with AI upscaling or generation
- Audio lacks expected microphone characteristics or carries synthetic speech markers
- Upload metadata shows mismatched creation dates (e.g., file created in 2024 but EXIF says 2026)
- Device fingerprint doesn't match the claimed platform origin
The result? Legitimate creators using AI-assisted workflows get caught in the dragnet. A photographer who uses AI upscaling, an artist who composites AI elements, or a marketer who runs copy through an AI editor—all face platform friction.
The Durable Fix: Strip and Inject Clean Phone Identity
Most "AI removal" tools only strip metadata. That's not enough. Platforms don't just check for bad metadata—they check for presence of authentic device identity. The only reliable fix is a two-step process:
- Strip – Remove all AI-specific metadata, C2PA assertions, PNG chunks, and XMP data
- Inject – Replace it with genuine phone camera metadata that passes platform verification
This isn't about deception—it's about authenticity. You're creating a genuine camera-captured artifact from your phone. The metadata should reflect that reality.
Step-by-Step: How to Properly Clean AI Content
Here's the concrete process for durable results:
- Extract original phone metadata – Before touching AI tools, photograph a test chart with your actual phone. Save the RAW or HEIC file. This preserves the device fingerprint: Make, Model, SerialNumber, LensInfo, Firmware.
- Strip all AI artifacts – Use a metadata cleaner that removes PNG ztxt chunks, XMP data blocks, and C2PA manifests entirely. Don't just clear EXIF—nullify the entire metadata structure.
- Inject phone identity – Take the device metadata from your test chart and apply it to the AI-generated image. Critical fields: Make = "Apple", Model = "iPhone 15 Pro", Software = "Adobe Lightroom", DateTimeOriginal = a realistic timestamp.
- Add GPS coordinates – Use your actual location or a plausible one. Include GPSLatitude, GPSLongitude, GPSAltitude, and GPSTimeStamp. Platforms flag missing GPS more aggressively than almost any other signal.
- Simulate lens characteristics – AI images lack lens distortion profiles. Inject LensModel, LensMake, and LensSerialNumber matching your phone's actual lens.
- Add compression history – Authentic photos pass through ISP pipelines. Add ProcessingSoftware tags, realistic Quality values, and standard JPEG quantization tables that match your phone's output.
- Verify before upload – Run the image through a metadata viewer (exiftool, exiv2) and check that no AI signatures remain. The file should look indistinguishable from a real phone photo.
The key insight: platforms aren't looking for "bad" images. They're looking for "not from a real device" images. Injecting authentic phone identity satisfies their verification checks without triggering detection classifiers.
The Bottom Line
LinkedIn Unfiltered solves the text problem—making AI-generated copy readable. But if you're publishing AI-assisted visuals, the metadata war is won or lost before you hit upload. C2PA standards, encoder fingerprints, and device identity verification aren't going away; they're intensifying.
Strip-only solutions are a temporary fix. Only complete metadata replacement with authentic device identity provides durable protection against platform detection. This is the approach professional creators and marketers are moving toward.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.