The term 'fake photo' covers two meaningfully different types: images that are entirely AI-generated, and images that are photographs manipulated through compositing, cloning, or other editing. Each type leaves different forensic traces and requires different detection approaches. AI-generated images are detectable through metadata absence, sensor noise patterns, and dimension constraints. Composited images are detectable through compression inconsistency analysis (ELA) and lighting or shadow discontinuities.

Detecting fully AI-generated fake photos

Fully synthetic images, those generated entirely by AI from text prompts, have no camera metadata, no sensor noise, and dimensions constrained by the generator's architecture. These are the easiest to detect forensically because they were never captured by a physical device. EXIF analysis, sensor noise measurement, and dimension checking are the primary signals. Missing EXIF on a JPEG at standard AI output dimensions is near-definitive evidence of AI generation.

Detecting composited and manipulated photos

Composited images, real photos with faces swapped, backgrounds replaced, or elements added, are harder to detect because they contain some authentic camera data. The primary detection technique is Error Level Analysis (ELA): re-encoding the JPEG at a known compression quality and measuring pixel-level differences from the original. Composited regions that were inserted at a different compression quality show as brighter areas in the ELA visualization, revealing the boundaries of manipulation. Lighting and shadow analysis can also reveal inconsistencies between composited elements.

Face-swap detection

Face-swap deepfakes (the most harmful category for personal harm) are a hybrid: they apply AI to composite a synthetic face onto an existing photograph or video. Detection looks for soft blending at face boundaries, lighting mismatches between face and body, and skin texture discontinuities. Metadata analysis is less effective if the base image is authentic, since camera EXIF will be present. Sensor noise analysis can still be useful if the composited face region shows lower noise than the surrounding authentic image.

When forensic results matter legally

For TAKE IT DOWN Act removal notices, you do not need to prove a photo is fake, you only need to assert that it depicts you in an intimate context without your consent. However, forensic detection results can strengthen a legal submission and support any criminal complaint. The ScanErase free detector generates a probability score and signal breakdown that can be included in documentation.

Frequently asked questions