Forensics Expert Demonstrates Geometry Test For AI Media

Farid, identified in the YourTango piece as a leading digital forensics expert at the University of California, Berkeley, told the outlet he often verifies whether photos or videos have been manipulated. According to YourTango, the article says Farid helped to found the field of digital forensics more than 20 years ago. YourTango quotes him explaining that earlier AI fakes were detectable from unrealistic sensor noise and statistical artifacts, but modern generative models reproduce those low-level patterns much better. The article reports Farid's clear diagnostic: check the scene's geometry and perspective for inconsistencies, because, in his words, "Generative AI doesn't know about physics, doesn't know about geometry."

Editorial analysis: Geometry and perspective checks are a classic forensic heuristic because real-world imaging follows strict geometric constraints: consistent vanishing points, coherent object scale across depth, and physically plausible occlusion. Models trained on large image corpora learn visual correlations but do not encode explicit 3D physical laws, so they can reproduce local texture while violating global spatial rules. For practitioners, that means visual heuristics targeting multi-object relationships and perspective consistency remain useful complements to pixel-statistics and noise-based detectors.

Editorial analysis: As generative models close the gap on sensor-level artifacts, detection is shifting toward higher-level inconsistencies. This move mirrors a broader pattern in detection arms races where improvements in synthesis push defenders to exploit structural, semantic, or temporal constraints that are harder to mimic. For image-forensics teams, geometry checks are computationally cheap and interpretable compared with some ML-based detectors, making them practical first-line triage tools.

Editorial analysis: Observers should monitor two trends: whether generative models progressively learn to respect geometric consistency through 3D-aware training or architecture changes, and whether automated detectors begin incorporating explicit 3D priors or multi-view consistency checks. Forensic workflows that combine statistical, semantic, and geometric signals will likely remain more robust than single-signal detectors.