About clAIm
Understanding the science behind our forensic detection approach.
Our Mission
As AI-generated content becomes increasingly sophisticated, the ability to distinguish authentic media from synthetic content is critical for journalists, legal professionals, researchers, and everyday citizens. clAIm provides transparent, research-backed analysis tools that give users actionable confidence scores rather than opaque binary verdicts.
Technical Approach
Our analysis engine uses multiple independent forensic signals rather than a single classifier. This multi-signal approach provides resilience against adversarial attacks and reduces false positives. Each media type is analyzed using techniques specifically designed for its characteristics.
Analysis Techniques
Image Forensics
- Error Level Analysis (ELA) — Detects inconsistent compression artifacts
- Frequency Domain Analysis — DCT-based inspection of frequency content
- Color Distribution Analysis — Statistical histogram anomaly detection
- Edge Consistency — Sobel-based edge distribution profiling
- Noise Pattern Analysis — Sensor noise uniformity detection
- Metadata Forensics — File format and EXIF data inspection
Audio Forensics
- Spectral Flatness — Geometric/arithmetic mean ratio analysis
- Temporal Consistency — RMS energy variation and periodicity detection
- Zero-Crossing Rate — Frame-level spectral characteristic analysis
- Transition Pattern Analysis — Splice and synthesis artifact detection
Video Forensics
- Frame Temporal Consistency — Inter-frame difference spike detection
- Color Stability — Cross-frame color drift measurement
- Edge Sharpness Tracking — Laplacian-based focus consistency
- Compression Artifact Analysis — Double-compression detection
Text Forensics
- Vocabulary Diversity (Perplexity Proxy) — Type-token ratio analysis
- Burstiness Detection — Sentence length variation measurement
- AI Pattern Matching — Common AI phrase and structure detection
- Repetition Analysis — N-gram and sentence opening repetition
- Readability Profiling — Flesch-Kincaid grade level assessment
Research Foundation
Our audio-visual fusion architecture is inspired by Single-Stream Audio-Visual Deepfake Detection (SS-AVD) research, which demonstrates that iterative cross-modal attention fusion outperforms traditional late-fusion approaches. Key innovations include Latent-Shuffle Augmentation (LSA) for mismatch robustness and Multi-Modal Style-Shuffle Augmentation (MMSSA) for compression-invariant detection.
The architecture is designed to support future integration of trained deep learning models for even higher accuracy AV fusion scoring, while the current client-side engines provide immediate value through classical forensic analysis techniques.
Important Limitations
- No detection tool achieves 100% accuracy. State-of-the-art systems reach 95-98% AUC on benchmarks, with lower real-world performance.
- Scores represent probability estimates, not definitive verdicts. They should inform, not replace, expert forensic evaluation.
- AI generation technology evolves rapidly. Detection techniques must be continuously updated to remain effective.
- Heavy post-processing, re-encoding, or intentional adversarial manipulation can reduce detection accuracy.