The rapid proliferation of state-of-the-art latent diffusion models—including Midjourney v6, OpenAI’s DALL-E 3, Stable Diffusion XL, and Black Forest Labs’ Flux—has radically degraded public trust in visual evidence. Synthetic photorealistic imagery is now routinely deployed in election interference, financial stock market manipulation, and wartime propaganda.

While early AI models were notorious for rendering mangled hands with seven fingers or melted facial features, current frontier generators produce skin pores, subsurface light scattering, and intricate textile textures that effortlessly deceive the naked eye.

However, generative models do not capture light through physical glass lenses onto silicon sensors; they statistically predict pixel values based on learned probability distributions. This fundamental architectural difference leaves indelible mathematical and physical artifacts that human forensic investigators can reliably identify.

The 7 Diagnostic Tests for AI-Generated Imagery

┌────────────────────────────────────────────────────────────────────────┐
│                   THE 7 SYNTHETIC FORENSIC DIAGNOSTIC TESTS            │
├────────────────────┬────────────────────┬──────────────────────────────┤
│ 1. PUPIL REFLECTION│ 2. ELA COMPRESSION │ 3. HIGH-FREQUENCY RESIDUALS  │
│ Corneal catchlights│ Uniform error level│ Unnatural noise distribution │
│ mismatch across eyes. across entire scene. across color channels.      │
├────────────────────┼────────────────────┼──────────────────────────────┤
│ 4. TEXT GLYPHS     │ 5. EAR ARCHITECTURE│ 6. FIBER & JEWELRY LOGIC     │
│ Pseudo-alphabetic  │ Asymmetrical tragus│ Necklaces fading into skin;  │
│ alien typography.  │ and lobule folds.  │ eyewear melting into temples.│
├────────────────────┼────────────────────┼──────────────────────────────┤
│ 7. LIGHT COHERENCE │                    │                              │
│ Shadows projecting │                    │                              │
│ toward light source!                    │                              │
└─────────────────────────────────────────┴──────────────────────────────┘

Test 1: Corneal Light Reflection (The Catchlight Test)

In a real photograph of a human subject, the cornea of the eye acts as a spherical mirror. The catchlight (the bright reflection of the light source, window, or flash) must be mathematically identical in both eyes: * In synthetic imagery, diffusion models render each eye as an independent cluster of probabilities. * The Glitch: Zoom into both pupils at 400%. If the left eye reflects a square window while the right eye reflects a circular umbrella flash, or if the reflection angle does not align with the subject’s gaze, the image is synthetic. * Pupil Shape: Real human pupils are near-perfect circles. Diffusion models frequently produce oblong, teardrop, or irregularly notched pupil perimeters.

Test 2: Error Level Analysis (ELA) Uniformity

Error Level Analysis works by measuring the compression differences across an image file. When an authentic camera saves a JPEG, different visual complexities (fine hair vs. a flat blue sky) compress at differing mathematical error rates: * When a human photo is manipulated in Photoshop (e.g., pasting a politician’s head onto another body), the pasted element displays a dramatically higher or lower ELA error level than the surrounding canvas. * In a pure AI-generated image directly exported from a generator, the entire canvas exhibits an artificially uniform error level across all pixel blocks, completely unlike the dynamic compression spectrum of a physical optical sensor.

Test 3: High-Frequency Residuals and Noise Spectra

Physical camera sensors produce fixed-pattern noise (PRNU: Photo-Response Non-Uniformity) caused by microscopic physical imperfections in the silicon sensor array. This noise signature acts as a unique camera fingerprint: * AI-generated images have zero PRNU. * Instead, diffusion models leave characteristic frequency-domain artifacts—subtle periodic grid patterns visible when the image is transformed via Fast Fourier Transform (FFT). * If an FFT spectrum reveals star-shaped circular ringing spikes in the high-frequency corners, it indicates synthetic upscaling or latent diffusion sampling.

Test 4: Text, Glyphs, and Numeral Topology

While frontier models (such as DALL-E 3) have improved at rendering short English words, generative models still fail at consistent background micro-text: * Inspect street signs, storefront awnings, book spines, military uniform patches, and vehicle license plates. * Diffusion models frequently produce pseudo-alphabetic glyphs: characters that resemble a chimeric hybrid of Cyrillic, Latin, and Greek letters, or words that spontaneously morph into gibberish midway through a syllable.

Test 5: Anatomical Micro-Structure (Ears, Teeth, and Nails)

Human anatomy contains rigid cartilaginous structures that AI models frequently approximate without understanding underlying biological mechanics: * Ears: Real human ears possess a defined helix, antihelix, tragus, and lobule. In AI generations, the inner folds of the ear frequently loop into impossible closed mobius strips or blend smoothly into the side of the jaw. * Teeth: Look for the “infinite toothline”. AI often generates an unnatural row of 12 or 14 visible incisors with zero distinction between premolars, or teeth that share a common gum line without individual roots. * Fingernails: While hand counts have improved, nail beds remain a primary failure mode. Look for fingernails that wrap around the side of the finger or emerge from the knuckle joint.

Test 6: Textile Coherence, Eyewear, and Jewelry

Inspect objects that require precise structural continuity: * Eyewear Frames: Follow the line of spectacle frames from the nose bridge to the ear. In synthetic portraits, the frame often passes through the eye socket or changes thickness abruptly behind the ear. * Jewelry & Necklaces: Chains around a neck frequently enter a collarbone and fail to emerge on the other side, or two distinct chains merge seamlessly into a single strand. * Fabric Weave: In real clothing, fabric threads follow orthogonal warp and weft lines. AI textiles often dissolve into chaotic swirls upon close inspection.

Test 7: Lighting Vector and Shadow Incoherence

Physics-based lighting requires that every object in a scene casts a shadow directly opposite the primary illuminator: * Identify the primary light source by looking at the specular highlights on the subject’s forehead or nose. * Trace the shadows cast by utility poles, vehicles, or nearby pedestrians. * In synthetic composites, shadows frequently project at conflicting angles, or objects hover weightlessly with zero contact ambient occlusion shadows beneath their base.

Forensic Verification Workflow: Real vs. AI Comparison

By methodically auditing corneal catchlights, Fourier noise residuals, and anatomical topology, researchers dismantle synthetic illusions and defend visual truth.