Inspect an image's metadata for AI-generation signals: camera data, software tags, provenance hints.
AI-generated images are detected in two broad ways, and it matters which one any given tool uses. Metadata analysis reads the file's embedded information: generator watermarks (visible or invisible), C2PA content credentials recording the creation tool, EXIF quirks that cameras produce and generators usually do not. Signal analysis examines the pixels: diffusion models leave statistical fingerprints — texture regularity, frequency-domain patterns, impossible physics in reflections and hands — that classifiers can learn to spot. Metadata is conclusive when present but trivially stripped; pixel analysis survives stripping but is probabilistic and degrades with resizing and compression.
| Signal | What to look for | Reliability |
|---|---|---|
| Metadata presence | Real camera photos carry EXIF; many AI images have none | Moderate — strippable |
| Content credentials | C2PA manifest naming the creating tool | High when present |
| Hands and text | Malformed fingers, garbled signage | High when visible |
| Reflections and shadows | Inconsistent light sources, missing reflections | Moderate |
| Texture uniformity | Over-smooth skin, repeating patterns | Moderate |
The button above runs the metadata half locally: it reads the file's embedded blocks and reports what creation information survives — camera data, software tags, C2PA hints. A file with full Canon EXIF and consistent timestamps is probably a camera photo; a pristine image with no provenance at all is ambiguous, because screenshots and stripped files look identical to generated ones. For pixel-level classification, dedicated detector models exist, and honest ones publish accuracy on adversarial benchmarks rather than marketing pages.
Every public detector's method eventually feeds back into generator training, and each generation of models erases the previous fingerprints — the tell-tale hands of a few years ago are refined away, frequency signatures diminish with better samplers. Provenance approaches sidestep the race entirely: rather than detecting fakes after the fact, cameras and editing tools cryptographically sign what they produce, so authenticity becomes verifiable rather than inferred. The two approaches are converging in practice: detection for the existing image sea, provenance for everything created going forward.
For photographers and creators, AI imagery is a commercial threat on two fronts — flooded stock and print-on-demand markets undercutting real work, and synthesized content damaging reputations. Provenance tooling has become part of the professional stack: signing your own work, monitoring for unauthorized use, and acting on what monitoring finds. That protection layer — provenance plus detection plus enforcement — is authAspect's territory, built specifically for image creators navigating exactly this landscape.
The hardest class of image to assess is the stripped one: no metadata, no credentials, no history. Detection tools return a probability, visual checks return impressions, and the truthful summary is "inconclusive leaning generated" or the reverse. What breaks the tie is context — where the image appeared, who posted it, what it claims to show — and provenance going forward: the ecosystem is moving toward cameras and tools signing their outputs so that the question "where did this come from" has an answer built into the file. Until coverage is universal, the practical stack remains: metadata when present, visual signals when visible, context always.
Editors, buyers, and investigators process unknown images in volume and need a repeatable pass: pull metadata first (fast, free, sometimes conclusive), run a reverse image search second (an image that has circulated for years with an established origin is usually its own verification), check visual signals third, and only then pay for pixel-level detection on the cases that still matter. This orders the checks by cost, and the cheap ones resolve most images. The residual hard cases — fresh, stripped, plausible — are precisely what provenance systems are being built to eliminate.
Know when you have enough. For casual skepticism — is this viral photo real — the free stack settles most cases: metadata plus reverse search plus visual checks reaches a defensible conclusion in minutes. For consequential decisions — publication, purchase, legal — inconclusive means inconclusive: do not squeeze certainty from tools that honestly report probability, and either find the provenance or act on the uncertainty itself. The discipline of stopping is underrated; most misjudgments around AI imagery come not from bad tools but from demanding binary answers the current technology does not give.
The fastest way to calibrate your judgment: run the analyzer on one photo you know is genuine (from your own camera) and one you suspect. Seeing what real metadata looks like versus ambiguous absence, once, does more for your detection instincts than any description. From there the workflow above takes over, and your conclusions get faster with every image.
Detection instincts decay as generators improve, so refresh them periodically: run a known-real and a known-generated image through your checks every few months and notice what changed. The signal that mattered last year may be gone; a new one may have appeared. Five minutes of deliberate recalibration keeps your conclusions anchored to the current generation rather than the one you first learned — and that is the difference between a skill and a memory.