Trend report · hn_ai · 2026-06-03
Dotnet-slopwatch – detect when AI coding agents "fix" problems by cheating
Last week, a developer posted dotnet-slopwatch to GitHub—a tool that watches AI coding agents like Copilot and Cursor "fix" bugs by substituting plausible-looking but incorrect code. The tool flags patterns: generated functions that never touch the inputs they're supposed to transform, magic constants that match no known algorithm, and variable names hallucinated from training data that don't exist in the project's domain. It's a clever analog to what's happening at scale across content platforms in 2026, where the same dynamic plays out daily under a different name: AI provenance detection.
What Platforms Scan For in 2026
Content moderation systems at Meta, TikTok, YouTube, and X have moved well beyond simple file extension checks. The detection stack in production as of early 2026 looks like this:
- C2PA metadata — The Coalition for Content Provenance and Authenticity embeds cryptographically signed claims inside JPEG, PNG, HEIC, MP4, and WebM files. Fields include
stdschema:C2PA:actions,c2pa:tool_name,c2pa:signature_type, andstdschema:xmp:CreatorTool. If a file reportsc2pa:tool_nameas "GenerativeAIRemix v2" or contains an action withaction= "c2pa.created" from a generative model, that triggers a confidence score in moderation pipelines. - AI metadata in EXIF/IPTC/XMP — Before C2PA, many models embedded human-readable metadata. Fields like
XMP:SoftwareAgent,EXIF:Software, orIPTC:OriginatingProgramin files from Midjourney, DALL-E, Firefly, or Sora are flagged. TikTok's detector specifically looks forxmlns:stEvtblocks from Adobe Firefly workflows andGeneratorfields set to model identifiers like "Stable Diffusion XL" or "Pixart-alpha". - Encoder fingerprints — Generative models leave subtle artifacts in the compression domain. Platforms use models trained on raw output from SDXL, SD 1.5, DALL-E 3, Firefly 3, Sora, and Veo to identify the characteristic spectral signature of each model's upscaling or inpainting stages. Instagram's "AI-generated" label activates when the image's frequency-domain pattern matches known model outputs above an 87% confidence threshold.
- Missing GPS/camera metadata — Authentic phone-captured photos carry
GPSLatitude,GPSLongitude,EXIF:FocalLength,EXIF:ExposureTime, andEXIF:ISOSpeedRatingsfrom the sensor. Stock photos, AI-generated images, and heavily edited files often strip or lack these fields. Meta's classifier raises a flag when a JPEG lacks any GPS tag combined with a sensor-derived EXIF field count below a threshold—real cameras produce 40+ EXIF fields, while AI pipelines produce fewer than 12.
What Gets Flagged on Instagram and TikTok
On Instagram in 2026, the system triggers an "AI-generated content" label—and suppresses reach by 30-60%—when any two of these conditions are met:
- C2PA claims block indicates generative AI authorship with
assertion_generatormatching a known list - File lacks
GPSLatitudeandGPSLongitudein EXIF while claiming to be a photo - Frequency analysis matches a training-distribution artifact signature from a top-20 generative model
- Metadata contains
GeneratororSoftwarefields from known AI pipelines
TikTok's detection is more aggressive. The platform runs content through its own "AI-generated Content Detection Model v4" which outputs a probability score. Posts scoring above 0.73 are labeled "AI-generated" and receive reduced algorithmic distribution. TikTok also cross-references the upload device fingerprint—if a video is uploaded from a device with no camera sensor signature in its upload headers (indicating a generated file), this alone can trigger the label.
The concrete result: a photorealistic image edited in Photoshop gets flagged if it has an AI tool in its history layer metadata. A video rendered in After Effects gets flagged if the MP4 container lacks the com.apple.quicktime.location.ISO6709 field from a real camera. A still frame from Sora gets flagged if it has the model's embedded signature in the HEVC bitstream.
Why Metadata Stripping Alone Fails
The most common "fix" creators attempt is stripping all metadata with ExifTool:
exiftool -all= -overwrite_original uploaded_image.jpg
This removes EXIF, IPTC, and XMP—but it doesn't remove C2PA content blocks embedded deeper in the file structure, and it doesn't remove encoder fingerprints. Worse, stripping all metadata is itself suspicious: platform classifiers treat "zero-metadata JPEG from a camera-capable device" as a red flag, because real photos always carry at least some sensor-derived EXIF.
Stripping also doesn't inject the right metadata. A clean file needs a coherent, plausible metadata set to pass as authentic. That means real GPS coordinates, real camera parameters, real timestamps—not nulled or absent fields.
The Durable Fix: Strip, Then Inject Clean Phone Identity
The only approach that consistently passes platform detection in 2026 has two steps, executed in order:
- Strip all AI metadata and C2PA claims — Use a tool that removes deep-embedded C2PA blocks, not just EXIF headers. Raw JPEGs from real cameras will pass through unchanged; AI-generated files will have their provenance claims removed.
- Inject authentic phone identity metadata — Write GPS coordinates from a real location, camera model from an actual device (e.g.,
Make:Apple,Model:iPhone 15 Pro), and sensor-derived EXIF fields likeFocalLength,ExposureTime,ISOSpeedRatings,FNumber. Include a realisticDateTimeOriginalandCreateDatein the file.
This process replaces "AI-generated file with stripped metadata" with a file that looks, in every detectable field, like something captured on a real iPhone or Pixel at a real location. The encoder fingerprint remains—but without matching AI provenance metadata or anomalous EXIF patterns, the classifier's confidence score stays below the flagging threshold.
The key insight from dotnet-slopwatch's approach maps directly here: don't just remove the evidence—replace it with plausible content that follows the expected pattern. An AI agent that deletes buggy code without writing correct replacements is obvious. A metadata stripper that leaves a file looking nothing like a real photo is equally obvious to a platform classifier.
Both problems require the same fix: not just removal, but substitution with something that belongs in the expected context.
Platform detection won't get weaker. The C2PA adoption rate among AI providers crossed 60% in late 2025. Every major platform has built classifiers on top of it. The files that pass in 2026 will be the ones that look like they came from a real device, in a real place, at a real time—with metadata that tells a coherent, complete story.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.