AI image generators have reached a level of photorealism that makes visual identification nearly impossible for human eyes alone. Midjourney v7, DALL-E 3, and Stable Diffusion XL can produce portrait photos, product shots, and news-style images that fool even experienced journalists.
This guide explains the forensic signals that give AI images away — and how automated tools catch what our eyes miss.
Why Visual Detection Fails
The human visual system is optimized to recognize faces and natural scenes, not pixel-level statistical anomalies. AI images exploit this: they produce textures, lighting, and composition that feel natural even when the underlying mathematics are completely artificial.
In 2024, researchers at MIT found that people correctly identified AI images only 61% of the time — barely above random chance. By 2026, that number has likely dropped further as models improve.
Forensic Signals That Reveal AI Images
1. Frequency Domain Anomalies
Every image contains a hidden fingerprint in its frequency domain — the distribution of pixel intensity changes across different scales. AI generators leave characteristic patterns here.
Tools like Fourier Transform analysis convert an image to frequency space and look for the regular grid artifacts that GANs and diffusion models produce. Real photographs have natural noise patterns; AI images have suspiciously smooth or periodically structured noise.
2. Facial Geometry Inconsistencies
Despite major improvements, AI still struggles with:
- Teeth: Often unnaturally uniform, too many or too few, or blending into gums
- Ears: Frequently asymmetric in ways that don't occur in real humans
- Eyes: Catchlights (the reflection of light sources) are often identical in both eyes — physically impossible
- Hair strands: Individual strands near the hairline often merge into each other or disappear entirely
3. Background Coherence Failures
AI generates the foreground subject and background somewhat independently. Signs of this:
- Reflections in windows or mirrors that don't match the scene
- Shadows cast in inconsistent directions relative to the apparent light source
- Text in the background is almost always garbled or nonsensical
- Objects in the distance become morphologically impossible (buildings with wrong geometries, trees with impossible branching)
4. Metadata Analysis
Real photographs carry EXIF metadata: camera make and model, lens focal length, GPS coordinates, shutter speed. AI-generated images are born without this data. A "photograph" with no EXIF metadata is a significant red flag.
Tools like exiftool or online EXIF viewers can reveal this instantly.
5. JPEG Compression Artifacts
When a real photo is saved as JPEG, compression artifacts appear in predictable patterns based on the camera's compression algorithm. AI images saved as JPEG have different compression signatures — the blocking artifacts appear in unusual locations because the underlying image structure is fundamentally different.
How Aiscern Detects AI Images
Aiscern's image detection pipeline combines multiple signals:
- EfficientNet-based classifier trained on 2M+ real/AI image pairs, fine-tuned on recent Midjourney and DALL-E 3 outputs
- Frequency spectrum analysis looking for GAN fingerprints and diffusion model artifacts
- Facial landmark analysis (for images containing faces) checking geometric consistency
- Metadata examination checking for missing or suspicious EXIF data
- Ensemble confidence scoring combining all signals into a single verdict
Practical Tips for Manual Inspection
When examining a suspicious image:
- Zoom into the background at 200%+ and look for text or repeated patterns
- Check the edges of hair where it meets the background
- Look at hands — AI still frequently produces six fingers or impossible finger geometry
- Examine reflective surfaces — eyes, glasses, wet surfaces — for inconsistent reflections
- Run it through a reverse image search — if it's a real stolen photograph, it'll appear
- Check the EXIF data — a modern smartphone photo without EXIF is suspicious
Limitations
No detection method is perfect. Current limitations:
- Heavily compressed images lose the frequency artifacts that signal AI generation
- Cropped images may remove the background inconsistencies
- AI images that incorporate real photo elements (inpainting/outpainting) are harder to detect
- Very small images (under 256×256) lack the detail needed for reliable analysis