Our Methodology

AI Lens uses a client-first, zero-external-API forensic detection platform designed to verify the authenticity of digital media. We operate strictly in-memory without relying on paid black-box LLMs, delivering deterministic, auditable, and privacy-preserving synthetic content analysis.

Image Analysis

Images are analyzed via deterministic heuristics and mathematical signatures:

  • EXIF Verification: Authentic photos carry camera sensor data (ISO, Focal Length). Missing data heavily indicates synthetic origins.
  • tEXt/iTXt Chunk Scanning: Diffusion models often embed latent seeds, prompt data, and tool metadata in PNG structures.
  • Header Inspection: Byte-level analysis maps formatting irregularities standard to generator toolkits.

Text Entropy Analysis

We perform computational linguistics without invoking costly LLMs to evaluate natural language:

  • Lexical Diversity (Type-Token Ratio): AI limits vocabulary breadth, resulting in rigid, uniform language structures.
  • Sentence Burstiness: Humans write with highly variable sentence lengths (burstiness). AI defaults to predictable variance.
  • Pattern Matching: Hardcoded transition anomalies (e.g., "In conclusion," "As an AI language model").

Forensic Risk Score

The Forensic Risk Score measures the combined strength of detected synthetic-content indicators. AI Lens aggregates weighted forensic signals into a 0–100 Forensic Risk Score. The score represents the strength of observed synthetic-content indicators, not a statistical probability or certainty.

Confidence

Confidence measures the strength, completeness, and reliability of the available evidence.

Important limitation

Metadata can be stripped, files can be re-encoded, and legitimate human-created content can sometimes exhibit similar statistical or structural characteristics. Results should therefore be interpreted as forensic evidence, not definitive proof of authorship or origin.