Forensic analysis looks for places where an image is internally inconsistent. It's genuinely useful, and it's also the area where confident amateurs most often reach wrong conclusions.
Physical Consistency (the Most Reliable)
Physics doesn't negotiate, so these checks are the strongest:
- Shadow direction. All shadows from a single light source should converge consistently. Objects with shadows pointing different ways indicate composition from multiple sources.
- Shadow softness. Shadows from the same source share edge hardness. A hard-edged shadow beside a soft one is suspicious.
- Reflections. Mirrors, windows, water and glasses should show what's actually in front of them, at the right angle.
- Perspective and scale. Added objects often sit at a slightly wrong scale for their apparent distance, or their vanishing lines don't match the scene.
- Lighting on the subject vs the scene. A person lit from the left composited into a scene lit from the right is the classic tell.
Statistical Signals (Useful but Noisy)
- Noise inconsistency. Every camera and ISO setting produces a characteristic noise pattern. A region that's noticeably cleaner or noisier than its surroundings may have been added, generated or heavily retouched.
- Compression inconsistency. Re-saved regions may show different JPEG block artifacts than the rest of the image.
- Error Level Analysis (ELA) visualizes how much different regions change when re-compressed. Widely used and widely misread — it highlights edges, text and high-detail areas in *unedited* images too, so a bright region is not proof of anything.
- Resampling traces. Scaled or rotated regions can carry detectable interpolation patterns.
- Duplicate regions. Cloned areas produce identical pixel blocks elsewhere in the image; automated clone detection finds these well.
Why False Positives Are So Common
Almost every published photograph has been edited: cropped, colour-corrected, sharpened, resized, re-compressed by a platform. All of these change the statistical fingerprint. Forensic tools flag them enthusiastically.
Additional confounders:
- Screenshots and social media re-encoding destroy the original compression signature entirely.
- HDR and computational photography merge multiple exposures by design, producing region-level differences that look like compositing.
- Denoising and AI enhancement create the smooth regions that noise analysis flags.
So the correct conclusion from a forensic anomaly is usually "this region was processed differently", not "this image is fake".
A Sensible Method
- Start with context, not pixels — source, earliest appearance, corroboration.
- Check physical logic: shadows, reflections, perspective. These need no tools and produce the most reliable findings.
- Then look for statistical anomalies, treating them as questions rather than answers.
- Weigh the stakes. Accusing someone of fabrication requires far more than an ELA screenshot.
- Consult a professional for anything legal or evidentiary. Real forensic analysis is a specialist discipline with rigorous methodology, and courtroom standards are far above the tools available online.
The most useful mindset: forensics tells you *where to ask questions*, while verification of sources tells you *what the answer is*.