You're scrolling. You pause on a photo. The lighting looks consistent, the shadows fall correctly, nothing is obviously off. You move on. But an AI flagged that image before you even registered it, and the reason has nothing to do with magic.
It's forensics. It runs on physics.
The Noise That Isn't Random
Every digital camera sensor introduces a faint, unique pattern of errors into every image it captures. Researchers call it Photo Response Non-Uniformity, or PRNU. Think of it as a fingerprint baked into the static: each pixel in a sensor responds to light slightly differently from its neighbors, and that microscopic inconsistency repeats across every single photo the device ever takes.
When someone copies a face from one image and pastes it onto another, the donor region carries its own PRNU signature. The background carries a different one. An AI trained on sensor noise doesn't see a portrait with one consistent origin. It sees two incompatible fingerprints sharing the same frame, like two radio stations bleeding into each other at exactly the same frequency.
Swap out someone's eyes. Smooth a scar. Brighten one cheek. The sensor noise in that small region still won't match the surrounding pixels. The AI doesn't need to know what a face looks like. It just needs to know that the noise pattern broke.
What Compression Knows
There's a second layer most people never consider: JPEG compression artifacts.
When a camera saves a JPEG, it divides the image into 8x8 pixel blocks and compresses each one. That process leaves a grid-like signature across the entire file. Open the image in a photo editor, clone-stamp a region, and save it again. The cloned area gets compressed a second time, on a slightly different grid alignment. Double-compressed regions show a characteristic artifact pattern that differs from singly-compressed areas, and AI systems trained on Error Level Analysis (ELA) map this directly, producing a heat map where tampered zones glow visibly brighter.
Consider a concrete example: a product photo where a competitor's logo has been quietly replaced with a client's own. The rest of the image has been compressed once. The logo region, pasted in from a separate file, has been compressed twice. ELA lights it up immediately. A human art director reviewing the final image might never catch it. The filter catches it in under a second.
That asymmetry is the whole problem, honestly.
The Geometry That Lies
Beyond noise and compression, there's geometric consistency. Light has a source. Reflections obey it. Shadows have to point away from that source at the correct angle.
Neural networks trained on millions of real photographs learn these relationships deeply enough to flag violations. A manipulated image where one person's catch-light (the small reflection in the eye) implies a window to the left, while the shadows under their chin imply a light source from above, fails a consistency check the human visual system often glosses over entirely. We're wired to process faces fast, not to audit their lighting geometry. AI does the opposite, and it's better at it.
This is where the most sophisticated detectors operate now: not on noise or compression alone, but on learned physical plausibility. Models underlying research like Google's SynthID work and Adobe's Content Authenticity Initiative tooling use convolutional neural networks that have essentially memorized what real light does, at scale.
What People Assume That Isn't True
The common assumption is that a good enough manipulation will fool any detector. For any single method in isolation, that's technically not wrong. But it misses how layered detection actually works.
Scaling an image by 1% destroys PRNU. Recompressing at maximum quality reduces ELA signals. Almost no real-world manipulation, though, destroys all forensic signals simultaneously without leaving its own new artifacts. Scrubbing one fingerprint tends to smear another. The cover-up creates evidence.
Also worth knowing, and this matters: AI detectors produce false positives. A heavily processed but entirely genuine image can trip noise-based detectors. Forensic AI is a triage tool, not a verdict. Courts and serious editorial teams treat it as one signal among several.
So what should you actually do with a flagged image? If the ELA map shows a single isolated hot zone over a face or logo, that's meaningful. If the entire image lights up uniformly, you're probably just looking at an aggressively processed original.
Manipulating an image leaves physical consequences in the data the same way pressing your thumb into clay leaves a ridge. The manipulation is never truly invisible. There's just a question of whether anyone is reading the ridges.