UnmarkPro
Forensics & Verification

AI Detectors Are Mostly Guesswork — What's Actually in the File

Websites claiming to detect AI images with "98% accuracy" flood search results, but in practice they frequently flag real camera photos as synthetic. Here is why probabilistic checkers fail, and what verifiable evidence actually exists inside the file.

Verify Real File Origin Factually

Inspect C2PA manifests, camera hardware headers, and IPTC claims locally.

Open What Made This

Why Probabilistic "AI Label Check Software" Fails

Most consumer-facing AI detectors use a lightweight convolutional neural network trained on a dataset of Midjourney and real photos. These models look for statistical artifacts: high-frequency noise gradients, overly smooth skin textures, or unnatural lighting symmetry.

This approach is fundamentally flawed:

  • A real photograph taken on a modern iPhone with Portrait Mode or studio ring lights often scores as "85% AI" because the camera's computational photography smoothed the background.
  • An actual generative AI image with added film grain or heavy compression often passes as "100% Human."

Probabilistic guessing cannot serve as a reliable foundation for professional publishing, journalism, or academic compliance.

The Binary Truth: What Is Actually Inside the File

Instead of guessing based on visual appearance, digital forensic experts inspect the immutable container structures of the file:

  1. Signed C2PA Manifests: The Coalition for Content Provenance and Authenticity (C2PA) specifies cryptographic container blocks. When OpenAI, Adobe Firefly, or Microsoft Designer generates an asset, they sign a manifest detailing the exact software tool, parent assets, and creation timestamp. This cannot be forged without the private cryptographic key.
  2. IPTC Generative Tags: The International Press Telecommunications Council standardizes metadata fields like digitalSourceType: trainedAlgorithmicMedia. This is the exact tag social platforms read to inject automated AI badges.
  3. Hardware EXIF Tags: Authentic camera captures carry physical lens serial numbers, shutter timings, sensor temperatures, and camera manufacturer payloads that synthetic generators do not fabricate.

Deterministic Verification: How UnmarkPro Works

Our What Made This? engine does not invent probability scores. It parses the binary bytes of your file directly in your browser:

  • If a signed C2PA manifest is present, it reports the verified generator (e.g., "OpenAI DALL-E 3").
  • If genuine camera hardware headers are found, it identifies the physical camera equipment (e.g., "Authentic Camera Capture").
  • If no verifiable signatures exist, it reports "Undetermined" honestly, rather than inventing a statistical guess.

Frequently Asked Questions

Why do AI image detectors give conflicting percentage scores?

Most online AI detectors use probabilistic computer vision models that guess likelihood based on surface textures and blur patterns. Because authentic photography with smooth lighting or studio grading mimics AI textures, probabilistic models generate frequent false positives.

How can software reliably check if a file was made by AI?

The only definitive proof comes from inspectable file metadata: signed C2PA provenance manifests, IPTC digitalSourceType assertions, and camera hardware EXIF records. These provide factual evidence rather than statistical guesses.

What is the difference between an AI detector and an AI provenance inspector?

An AI detector guesses based on visual heuristics. An AI provenance inspector reads the binary file container to verify whether the file was cryptographically signed by an AI generator (like OpenAI, Adobe, or Google) or captured on physical camera hardware.

Can I inspect an image's provenance without uploading it to a cloud server?

Yes. In-browser tools like UnmarkPro read C2PA manifests, EXIF metadata, and IPTC tags directly in your local browser memory with zero server uploads.