AI-Generated Images vs Real Photos: 15 Forensic Signals Investigators Look For

Why a Checklist Matters

Generative image models in 2026 produce photographs that fool most viewers on first glance. A scrollable feed, a low-resolution social-media re-upload, and the human tendency to trust faces all conspire against casual verification. Gold Stone Intelligence uses a structured 15-signal checklist, applied in order, to determine whether an image is camera-captured, edited, or fully synthetic.

Key Takeaways

• No single signal is decisive — forensic conclusions come from convergence across multiple categories.
• Signals fall into four buckets: biometric/anatomical, physical-world consistency, generative artifacts, and metadata/provenance.
• Even a "clean" AI image will leak traces in at least three of these categories under expert analysis.

Biometric and Anatomical Signals

1. Catchlight Symmetry in the Eyes

Real photographs of a person illuminated by a single dominant light source show matching catchlights in both eyes — same shape, same position, same intensity. Diffusion models still struggle to keep these consistent, especially in three-quarter poses.

2. Pupil and Iris Geometry

Authentic pupils are circular and concentric with the iris. AI images frequently produce elliptical pupils, off-center placement, or pupils with slight color bleed into the iris.

3. Ear Asymmetry Patterns

Ears are biometrically unique and notoriously hard to generate. AI-generated portraits often render two visibly mismatched ears or attach earrings to a malformed lobe.

4. Hair Strand Continuity

Real hair has continuous strands that obey gravity and follow consistent flow. Synthetic hair tends to break into mid-air, merge unnaturally, or show repeating texture patches at high zoom.

5. Dental Structure

Teeth in AI images often appear as a uniform white surface without individual tooth boundaries, or as a row of identical, repeating shapes — a signature of decoder hallucination.

Physical-World Consistency Signals

6. Shadow Direction Coherence

Every shadow in a real scene points away from the dominant light source by the same angle. AI-generated scenes frequently mix multiple incompatible shadow directions, especially around chairs, table edges, and feet.

7. Reflection Consistency

Mirrors, windows, and polished surfaces in real photographs reflect the scene with geometric accuracy. Generative models often produce reflections that do not match the subject's pose or simply do not exist where they should.

8. Perspective Vanishing Points

Authentic architectural images show clean vanishing points where parallel lines converge. AI images often have subtly diverging "parallel" lines, betraying that no real camera ever recorded the scene.

9. Hand and Finger Anatomy

The classic deepfake tell. Even in 2026, hands remain a weakness: extra fingers, fused joints, impossible thumb angles, or rings that pass through skin all signal generative origin.

10. Text in the Scene

Signs, book covers, T-shirt logos, and license plates in AI images often contain gibberish letterforms or impossible glyphs. Real-world text obeys typography and language rules.

Generative Model Artifacts

11. Diffusion Noise Fingerprint

Diffusion-based generators leave a characteristic high-frequency noise pattern visible in the FFT (Fast Fourier Transform). Specialized detectors compute this fingerprint and compare it against a known library of model signatures.

12. Texture Repetition at Patch Boundaries

Latent diffusion models often produce subtle repeating textures along 64×64 or 128×128 pixel patch boundaries. At extreme zoom these grids become visible.

13. Skin Smoothness Anomaly

Real skin has pores, micro-shadows, and irregular highlights. AI-generated skin tends to be hyper-smooth in some patches and over-textured in others, never matching the controlled randomness of a real sensor capture.

Metadata and Provenance Signals

14. EXIF and Container Trail

A photograph from a real camera carries EXIF data: make, model, exposure, GPS, software string, and crucially a thumbnail preview that must match the main image. AI images either lack EXIF entirely, contain inconsistent fields, or carry a thumbnail that does not match the file's content.

15. C2PA Content Credentials

Increasingly, both real cameras and ethical generative tools embed C2PA-signed credentials describing creation and edits. Absence of any provenance signature in a high-stakes image is not proof of forgery, but its presence — when cryptographically valid — is strong evidence of authenticity.

How GoldStone Applies the Checklist

Our analysts score each signal on a four-point scale (consistent, weak inconsistency, strong inconsistency, decisive). The final report includes per-signal evidence, an overall confidence interval, and a recommended verdict. This methodology is reproducible, auditable, and court-admissible — and it is the foundation of every Certificate of Authenticity we issue.

Frequently Asked Questions

Can I check these signals myself?

Several signals — hand anatomy, scene text, shadow direction — are accessible to attentive non-experts. The deeper signals (diffusion fingerprints, EXIF thumbnail mismatch) require specialized tools and trained analysts.

Why don't free online detectors give a clear answer?

Free detectors usually rely on a single neural classifier trained on a limited sample of generators. They produce a probability score, not a methodology. Court evidence requires the kind of layered, documented analysis described above.

Will these signals still work in 2027?

Some will weaken as models improve — particularly hand anatomy and dental structure. Others, especially provenance signals and frequency-domain fingerprints, will grow stronger as C2PA adoption and detector calibration improve.

Need a Verified Answer?

Submit your image for a forensic Certificate of Authenticity via our Request Analysis form, or read our companion guide on what makes content a deepfake.