how AI image detection works: probability, not proof

the certainty trap

binary certainty is the first thing to unlearn. a detector that returns a flat “real” or “fake” hasn't solved the problem, it has hidden the interesting part: how sure, based on what, and wrong how often. the Reuters Institute's guidance for newsrooms (April 2024) is blunt about it: detection tools return a probability, their signal is degraded by cropping and re-compression, and a clean result does not prove a file is genuine. and when a vendor does promise certainty, regulators have started checking. in 2025 the U.S. FTC ordered Workado to stop advertising its AI Content Detector as roughly 98% accurate after its complaint alleged independent testing put it at 53% on general-purpose content.

the asymmetry matters too. a missed AI image is bad; a real photographer's work flagged as machine-made is worse, because that's an accusation. honest detection tunes against the false positive, which means accepting more “uncertain” verdicts, not fewer.

what a classifier actually reads

the era of counting fingers is mostly over. as generators improved, the visible glitches (warped hands, melted text, impossible fences) got rarer, and the reliable signal moved somewhere your eye can't follow: statistical regularities in how a generator renders noise, texture, lighting and detail. classifiers are trained on enormous piles of known-real and known-generated images until they learn those regularities well enough to score a new image against them.

two honest caveats come with that. first, the traces are fragile, not indestructible: screenshots, crops, filters and every re-compression pass erode them, which is why a heavily reshared meme is genuinely harder to read than a fresh upload. second, the patterns are generator-shaped. a Midjourney image and a Stable Diffusion image tend to leave different statistical residue, which is what makes model attribution possible at all. attribution is always a resemblance claim (“this looks most like Midjourney”), never an identification, because a detector can only name models it was trained to recognize.

why a panel beats a single model

a single classifier is a single point of failure: one training set, so one blind spot. the sturdier shape is an ensemble — several independent detectors that fail on different inputs, so one model's blind spot gets caught by another's read.

that's the shape amige. runs. a panel of independent detectors scores each image; their reads are weighed by each detector's track record, fused into one estimate, and capped so a verdict never reads 0% or 100%. before any of that, a trained routing layer points the scan toward the detectors most likely to read it well; attribution coverage spans 90+ generative models. the routing hint never enters the verdict itself. when the panel splits, amige. says uncertain rather than guess. disagreement is reported, not averaged away: you can see every detector's read on the result page.

the scan modes, honestly

• →auto. the fast consensus: a couple of detectors, and a verdict only when they agree. built to keep a single scan quick. it is not a bulk-moderation pipeline, and we won’t pretend otherwise.
• →deep. the full panel. more detectors, stronger consensus required before the verdict lands. use it when the answer matters more than the wait.
• →custom. you pick which detectors run. useful when you already know which ones you trust for a given kind of image.
• →prioritize deepfake. a separate toggle, not a mode: it switches the panel to deepfake-capable detectors and asks a different question, ‘is this a doctored real person?’ rather than ‘is this generated?’

video gets the same panel logic plus one extra problem: time. generators still find it hard to keep a scene coherent frame-to-frame, so video detection leans on temporal consistency as well as per-frame signal. and for any file, the strongest evidence isn't statistical at all: if an image carries signed Content Credentials, that provenance outranks any classifier's guess. most files in the wild don't, which is why the probabilistic layer exists.

what to do with a probability

treat it like a weather forecast, not a court ruling. an 80% likely-AI read on a stranger's profile photo is a reason to check further (reverse-search it, ask for another shot, look for provenance), not a conviction. an “uncertain” is the tool telling you the truth about its own limits, which is more useful than a confident number that folds under one crop. that's the standard we hold our own machine to, and the standard worth demanding from anyone else's.