AI Photo Detector

The obvious signal is texture, and it is the least reliable one on its own. Over-smooth surfaces, uniform backgrounds, and repetitive micro-detail all read as synthetic — but so do heavy beauty filters, aggressive denoising, and low-light phone processing. Any detector resting on texture alone will produce false positives on real, heavily edited photos.

The stronger signals are physical. Sensor noise in a real capture has a structure that stays consistent across the frame while its amplitude varies; generated noise tends toward uniformity or toward an uncorrelated pattern that resembles film grain without the underlying sensor statistics. Lens vignetting falls off smoothly and predictably. Motion blur follows the subject. Specular highlights obey the light source that supposedly cast them. Generators reproduce the look of these effects but frequently break their internal consistency.

Structural coherence matters as much as any single cue. In a genuine photograph, the geometry holds: perspective converges correctly, repeated elements like brickwork or foliage keep their shape, text on signs and labels resolves into legible characters. Generators still collapse on these under enough scrutiny, especially in small regions of the frame that received little attention during generation.

What the detector cannot know matters just as much. It does not know who uploaded the file, whether the uploader is who they claim to be, or whether a watermark or metadata block was stripped in transit. Content Credentials (C2PA) and SynthID answer a different question — check those separately when present, and treat their absence as neutral rather than as evidence.

Zoom to full resolution and look where generators still struggle. Hands and fingers merge, gain extra joints, or terminate in soft blobs. Ears and teeth blur into shapes that almost resolve but never quite do. Glasses produce reflections inconsistent with the room behind them. Text inside the frame — signage, labels, logos, book spines — looks plausible at thumbnail size and collapses into letterform nonsense when you actually read it.

Then check the physics. Shadows falling in two directions. Reflections showing a scene that does not match the one visible. Objects floating without contact, or intersecting without occlusion. Liquids frozen mid-splash in a way no shutter timing would produce. Fabric behaving like moulded plastic. Backgrounds dissolving into soft mush where depth should be sharp.

Two more tells worth a look: repeated textures stamped across a frame with the same exact shape, and the general absence of sensor noise in shadow regions — deep shadow in a real capture is where noise is most visible, and generated images tend to keep it suspiciously clean.

The critical caveat: newer models fix many of these at social-media output sizes, because the artifacts live at frequencies that resampling removes. Absence of visible tells is not evidence of authenticity. Treat these signs as a reason to run a scored check, never as a substitute for one.

Camera captures carry a physical fingerprint that is difficult to fake. Noise that stays structurally consistent while its amplitude varies with light. Motion blur where moving things are sharp and static things are not. Slight vignetting in the corners. Small optical imperfections — chromatic fringing at high-contrast edges, mild softness in one corner from a cheap lens.

Real scenes are also messy in believable ways. Cables tangle and droop with correct gravity. Dust accumulates in a way consistent with use. Paint wears where hands touch. Scuffs appear where objects actually get handled. Generators tend toward a tidiness that reads as artificial precisely because real environments are never so well organised.

Light is the strongest single confirmation. In a genuine photograph, every shadow, reflection, and highlight traces back to one consistent source or a small number of plausible sources. Look for the moment that breaks: a shadow pointing away from the brightest surface, a highlight on something that should be in shadow, two incompatible light directions in one frame.

Provenance beats pixel inspection whenever it exists. An original camera file with intact EXIF, a C2PA Content Credentials record, a burst of adjacent frames, or a second photo of the same scene from another angle all support a conclusion far more strongly than any single-image analysis. When those are available, use them first.

High scores mean strong synthetic cues under the current model. Treat them as a reason to pause before approving a refund, publishing a news image, or trusting a profile picture — not as a verdict. Mid-range scores mean the evidence is genuinely mixed: heavy editing, low resolution, a hybrid pipeline, or a file that has been through several recompressions. That band is where most false positives live and where a human decision is worth more than a debate about the number.

Low scores mean no strong generative cues were found in this file. They do not mean the image is authentic, that its story is true, or that nobody manipulated it outside the generative pipeline. Editing tools, colour grading, and ordinary photography all produce camera-like results. A clean score on a degraded upload is weaker evidence than the same score on an untouched original — which is why file quality is part of reading the result, not an afterthought.

If two versions of the same photo disagree, do not average them. The highest-quality file is the authoritative read, and disagreement between versions is itself a signal that the file was processed. Escalate rather than splitting the difference into a confident-sounding number that means nothing.

Detection ages fast. A score from an older engine is not a permanent label, and a result worth acting on today is worth re-checking with current tools if the decision is challenged months later. Keep the original file — not a screenshot — so a later re-check is meaningful.

Marketplace operators can screen high-value listings, delivery evidence, and refund claims before a human spends time on them. Synthetic product shots and staged damage images slip through visual review consistently at volume. The score is most useful here as a routing signal — flag the file, do not auto-reject it.

Newsrooms and creators can check a viral image before amplifying it. Ask for the original file, reverse-search for prior publication, then run the detector. When the stakes are high — elections, disasters, public figures — treat even a moderate score as a reason to wait for corroboration rather than a reason to publish.

Educators and hiring teams can use a scored pass as the first step on portfolio pieces, assignment submissions, or candidate photos. Be transparent that automated screening is a first pass, and keep a human in the loop whenever academic or employment consequences attach.

For any of these workflows the policy matters more than the threshold. Auto-approve low scores with low stakes, route the middle band to review, hold high scores until the source is confirmed — and record what you decided, because that record is what makes the policy consistent between reviewers.

Photography has a physical baseline that art does not. A camera leaves sensor noise, lens behaviour, and lighting physics behind, and a human airbrushed gradient has no equivalent trail to follow. This gives photo detection real anchors: the engine knows what a genuine capture looks like statistically, and can measure how far a file departs from it.

Illustration has no such anchor. Flat colour fields, clean digital inking, cel shading, and airbrushed gradients are smooth by design, which pushes human and generated work toward the same readings. Style also compresses the signal — the texture and noise cues that separate the two collapse together under stylisation. This is why art-focused checks carry a wider uncertainty band and lean on structural repetition and detail coherence instead.

The practical consequence for how you use these tools: treat a photo score as a meaningful read on origin, and treat an art score as a prompt to look closer. The same percentage means different things on different pages, and reading either without knowing which question the page answers is how false confidence happens.