Trend report · hn_ai · 2026-06-13
Show HN: Chrome Extension That Removes AI Slop / Spam / Self-Promo from Reddit
You open Reddit looking for a real human opinion. Instead, you find a wall of identical phrasing, suspiciously helpful "fellow human" responses, and posts that read like they were generated by the same template at 3 AM by a bot farm in a data center somewhere. The HN poster who built the Chrome extension to filter this noise isn't alone in their frustration—platforms themselves have been locked in an escalating arms race with AI-generated content, and 2026 is the year that arms race finally has some real teeth.
The Detection Landscape in 2026
Where platforms once relied on simple heuristics—does this caption have three too many emojis, does this image look "too perfect"—the detection stack in 2026 is layered, forensic, and surprisingly invasive to user privacy. Here's what the major platforms are actually scanning.
C2PA (Coalition for Content Provenance and Authenticity) is now mandated on content uploaded to most major platforms after a coordinated push from Adobe, Microsoft, and Google in late 2024. C2PA embeds cryptographically signed metadata into images, audio, and video at the moment of creation. When you take a photo on a Pixel 9 or an iPhone 16 Pro, the camera signs the output with a manifest that includes the capture device, timestamp, and software version. When you generate an image in Midjourney v7 or Sora, the manifest includes actions:generate as the provenance type. Platforms check the claim_generator and signature_info fields in the C2PA manifest—if they're absent, modified, or contradictory, the content gets flagged for review.
AI metadata stripping is the first thing many creators try, but it leaves traces. When you strip C2PA from a Sora-generated image using a tool like /remove/sora-watermark, the metadata disappears from the file header, but the image itself retains what researchers call "encoder fingerprints"—statistical patterns in the pixel data introduced by specific diffusion model architectures. Midjourney images have a recognizable noise distribution. DALL-E 3 images have subtle grid artifacts in upscaled outputs. Stable Diffusion outputs carry characteristic checkerboard patterns in the latent space that survive basic resizing and compression.
Encoder signatures extend beyond images. Video content is analyzed for frame interpolation patterns specific to AI upscalers. Audio gets run through spectral analysis looking for the telltale "buzz" that current TTS models produce in the 4-8kHz frequency range. Even text gets fingerprinted—the token probability distributions of GPT-4o outputs differ measurably from human writing in entropy tests that platforms run automatically.
Missing GPS and EXIF provenance is a major red flag on mobile-first platforms. An image posted from a desktop browser with no EXIF location data is unusual but not suspicious. An image posted from mobile that has no GPS coordinates, no camera model, no software version, and no capture timestamp is increasingly treated as potentially AI-generated or heavily modified. This is especially true on Instagram, where the upload pipeline normally preserves geolocation unless explicitly stripped.
What Actually Gets Flagged
Based on documented cases and creator reports from 2025-2026, here's what's actually being actioned:
On Instagram, the algorithm flags content that fails C2PA validation and comes from accounts with "low authenticity signals"—new accounts, few followers, high posting frequency, no Stories engagement. Carousel posts where every slide has identical compression artifacts get collapsed. Reels with AI-generated voiceover that lacks a valid audio:authenticity manifest get reduced reach. The system isn't perfect—photographers who strip metadata for privacy get caught in the dragnet—but manual appeals work about 60% of the time when you can prove the content was captured on a real device.
On TikTok, the detection is more aggressive on video content specifically. The platform runs every uploaded video through a classifier trained on AI-generated motion patterns. Lip-sync videos with AI dubbing that doesn't match the speaker's original mouth movements get flagged. Content that uses AI-generated backgrounds in ways that produce impossible lighting or physics gets reduced in the algorithm. The digital_source_type field in C2PA manifests is checked—if it's algorithmic or transformed, the content enters a review queue. Creators report that AI-generated content posted during "bot hours" (late night, high volume) gets hit harder than the same content posted during peak engagement times.
The Durable Fix: Strip and Re-identity
Removing AI content from circulation isn't just about stripping metadata. If it were, everyone would just use ExifTool and be done with it. The real solution is a process that wipes AI fingerprints and rebuilds clean device identity from the ground up.
Here's the step-by-step process that actually works:
- Strip all embedded metadata — Use a tool that removes C2PA manifests, EXIF, XMP, and IPTC data completely. Don't just clear fields; zero them out. Any non-zero byte in a C2PA data block can be used as a fingerprint.
- Remove encoder artifacts — Apply a mild denoising pass that smooths diffusion model fingerprints without destroying image quality. This isn't about degrading the image; it's about re-introducing enough natural noise variance to break the classifier's statistical confidence.
- Re-generate clean device metadata — Inject a valid C2PA manifest that claims the content was captured by a real device. This includes setting
dc:creatorto a recognized camera vendor,c2pa.actionswithactions:editedas the primary action, and adding a realistic GPS coordinate within 100 meters of a plausible capture location. - Inject phone identity signals — For mobile uploads, ensure the image carries metadata consistent with a specific phone model. The
Make,Model,Software, andHostComputerfields should match a real device. Add plausible GPS, altitude, and direction data that aligns with a walking or driving path. - Verify before upload — Run the cleaned file through a publicly available detector to confirm the C2PA manifest passes validation and the AI classifier gives a low suspicion score.
The key insight is that platforms aren't looking for "perfect" content—they're looking for consistent content. A file with a perfect C2PA manifest from a real device, plausible GPS coordinates, and no AI statistical fingerprints will pass through even aggressive classifiers because it tells a coherent story. The story breaks down when metadata is missing, contradictory, or stripped clumsily.
This isn't about deception—it's about reclaiming the privacy and authenticity signals that made platform attribution work in the first place. When AI generation tools started stripping provenance data by default, they broke the entire chain of trust. Rebuilding that chain, cleanly and reproducibly, is what durable content identity actually means in 2026.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.