The Tells of an AI Image: A Visual Forensic Walkthrough

Spotting an AI image used to be easy. Six fingers, melted faces, garbled signage. The current generation — Flux Pro, Imagen 4, GPT-Image-1, Sora 2 — has fixed most of those obvious tells. But forensic signals remain, even when the image looks photorealistic. Here's how to read them.

The visible-eye checklist (what's left)

• Hair near the boundary. Most generators still struggle where hair meets a complex background. Strands fuse, blur, or terminate inside an object.
• Reflections that don't match. Glasses, mirrors, and water reflections rarely line up with the source geometry.
• Crowd faces. Generators handle one or two faces well, then start producing morphing mannequins in the back rows.
• Text in the scene. Even Imagen 4 produces uncanny, half-correct text on signs, packaging, and posters.
• Hand–object interactions. The model may render the hand and the held object correctly in isolation but fudge the contact geometry.

Beyond the eye — the four forensic signals

Visual checks fail more and more often. The reliable signals are forensic. We run four of them in our image detector, all in your browser.

Error Level Analysis (ELA)

Re-encode the image as JPEG at a known quality, decode it, and subtract from the original. Areas with high delta were either edited or carry naturally-noisy detail. A uniformly dark ELA — almost no delta anywhere — is suspicious: it suggests the image was born at one compression level, which is very common with diffusion-model output decoded straight from latent space.

ELA is not magic. It misfires on screenshots, fresh JPEGs from your phone, and heavily processed real photos. Read it as one signal, not a verdict.

FFT magnitude spectrum

Take the 2D Fourier transform of the luminance channel and view the log-magnitude. Real photos give a smooth, roughly-radial falloff from the center. AI samplers — especially older diffusion variants and GANs — leave bright spots off-center, periodic checkerboard patterns, or anisotropic streaks. The newest models have flattened these artifacts, but they haven't eliminated them.

Channel decomposition

Split the image into R, G, B (and optionally YCbCr or HSV) channels and view each as grayscale. Real cameras have well-known color correlations driven by their Bayer-array sensors and demosaicing pipelines. AI generators learn approximations of these correlations, but rarely match them exactly. Look for: blue channel that's too clean, green channel that doesn't carry the expected luminance weight, mismatched grain across channels.

Noise residual (PRNU proxy)

Subtract a blurred version of the image from itself, leaving only high-frequency detail. Real cameras leave a structured sensor-noise fingerprint (Photo-Response Non-Uniformity — PRNU) that's statistically unique to each device. AI images either have no such fingerprint or have a learned, repetitive pattern that looks wrong under close inspection.

The signal that beats them all: C2PA

When an image is signed under the C2PA standard (OpenAI DALL-E 3, Adobe Firefly, Sony cameras, BBC News), the signature is cryptographic. It either verifies or it doesn't — no heuristics involved. We read C2PA before running any forensic transforms, because nothing else matches its reliability.

C2PA isn't universal yet. Older AI images and most consumer cameras don't carry it. But the trajectory is clear: in three years, the question will be "is this signed?" not "is this AI?"

Why combined evidence beats any single classifier

A binary classifier outputs one number. An ensemble of forensic signals outputs five or six, each visualizable. The combined signal is harder to evade — adversarial editing that defeats ELA usually doesn't fix the FFT spectrum; adding sensor noise to fool PRNU doesn't restore correct channel statistics. This is why our detector shows you every signal: not for vanity, but because the breakdown is the moat.

What you can't detect

• AI images that have been re-photographed off a screen — analog attacks defeat most digital forensics.
• AI output that's been heavily edited and re-saved through a full processing pipeline.
• Latest-generation outputs at small resolutions — the artifacts get quieter as compute scales.

The honest framing: forensic detection raises the cost of generating undetectable AI images. It doesn't make them impossible. The long-term answer is provenance, not detection.