Methodology
How Real-Lenses determines whether media is authentic, edited, or AI-generated — and what it can't catch.
Multi-layer forensic inspection
Real-Lenses combines a large multimodal vision model with a structured forensic prompting framework. Rather than a single binary classifier, it performs several independent checks across pixel, semantic, and physical-plausibility layers — then synthesises them into a final verdict with an explanation. This ensemble approach makes it significantly harder to fool than single-model detectors.
Visual Pattern Analysis
Our model examines micro-level pixel patterns that reveal statistical anomalies left by generative AI systems. Natural photos exhibit organic sensor noise; AI-generated images often display suspiciously smooth textures, repeating background tiles, and unnatural edge blending.
AI Artifact Detection
Modern diffusion models and GANs leave distinctive fingerprints: warped geometries near image borders, fused or malformed extremities (hands, teeth), unnatural specular highlights, and garbled text. We specifically probe these high-failure-rate regions.
Splicing & Clone Detection
Composite edits created in tools like Photoshop often leave telltale signs: inconsistent JPEG compression blocks, mismatched noise levels between pasted regions, seams where lighting direction shifts abruptly, or duplicate texture patches left by clone-stamp operations.
Lighting & Shadow Coherence
Authentic scenes have a single coherent light source. Our analysis traces catch-lights in eyes, shadow direction on faces and objects, and ambient occlusion consistency. Composited or AI images frequently contain multiple conflicting light sources.
Contextual Reasoning
Beyond low-level signals, we apply high-level scene understanding — do object proportions make physical sense? Are reflections in mirrors and windows accurate? Is depth-of-field realistic for the implied focal length? These heuristics catch edits that bypass pixel-level checks.
Confidence Scoring
Each verdict is accompanied by a 0–100 confidence score derived from the strength and number of corroborating signals. A score above 85 indicates strong evidence; 60–85 suggests probable but not certain; below 60 means ambiguous — treat low-confidence verdicts skeptically.
Accuracy & Testing
We don't publish a single accuracy number because real-world performance depends heavily on the image source, the generative model, and the manipulation type. Here's what our internal testing shows:
Across ~1,000 known-origin images, we correctly identify AI-generated content approximately 85–92% of the time.
False positive rate (flagging a real photo as AI) is approximately 5–8%.
We re-evaluate these ranges quarterly as generative models improve, and update this page when they shift.
These figures are from our internal test set, not a third-party benchmark. We're working toward public benchmark participation and will link results here when available.
What we can detect
- •AI-generated images from DALL·E, Midjourney, Stable Diffusion, and Sora
- •Face-swap deepfakes in both photos and video
- •Cloned or spliced regions from copy-paste manipulation
- •Inconsistent lighting and shadow directions in composited scenes
- •JPEG artifact mismatches between pasted regions
- •Unnatural textures, repeating patterns, and warped geometries
What we can't detect (yet)
- •Sophisticated retouching by skilled human artists
- •AI images deliberately post-processed to strip artifacts
- •Subtle color or contrast adjustments that don't change content
- •Audio deepfakes and voice cloning (separate feature, on our roadmap)
- •Real-time deepfake video streams (live detection is a roadmap item)
- •The original source or provenance of an image
Confidence score interpretation
Strong evidence. Multiple corroborating signals align.
Probable, but not certain. Treat with caution and seek corroboration.
Ambiguous. The image doesn't trigger strong signals either way.
Important limitations
No detection system is infallible. Highly sophisticated AI imagery or professional retouching can evade detection. Real-Lenses results should be treated as probabilistic evidence — one signal in a broader verification process — and are not suitable as sole evidence in legal or forensic proceedings. Court-admissible reports include chain-of-custody documentation, but the underlying detection remains probabilistic.