most detectors hand you a number. amige. shows you how it got there. it reads what you drop, sends it to the detectors built to read that kind of thing, weighs what they say, then settles on one honest estimate — likely AI, likely human, or uncertain. two quick looks help steer it, and neither one ever decides the verdict. a probability, never a proof.
a scan is read, routed, weighed by a panel, then fused & calibrated — landing on a probability, never a proof. watch one go through.
how it works, in full →How an amige. scan works, step by step. First, the content you drop — an image, video, or text — is read for signal. Before any verdict, amige. takes two quick looks that guide the scan but never score it: a routing hint that points toward the detectors most likely to read this kind of content well, and a novelty check that notices when something is unusual so the engine treads more carefully. The novelty check only flags that the content is unusual; it never claims what made it, and it is not a new-generator detector. Next, a panel of independent detectors — built by different teams on different data — each weigh in on their own. They do not all agree, and the disagreement is kept on display, not averaged away. Their readings are then fused and calibrated into a single estimate: combining only tempers confidence, it never inflates it, and the result is capped so a verdict never reads 0 percent or 100 percent. The outcome is one of three honest calls — likely AI, likely human, or uncertain — always a probability, never a proof. When the panel splits or the content is unusual, amige. holds the verdict back as uncertain rather than guess. The readings shown in the animation are illustrative of a single scan and are not accuracy claims.
amige. trained its own eye for the marks generators leave, so it can say what a thing resembles — not just that it’s synthetic.
different detectors are strong at different things. amige. routes what you dropped to the ones built to read it best.
when the evidence splits or the content is unfamiliar, it holds back and says uncertain — rather than guess.
amige. watches itself over time and recalibrates as generators change. honesty isn’t a launch feature; it’s a habit.
one pass per scan, two loops that never stop correcting it. follow the spine left to right. the two quick looks feed routing and the hold-back gate — but neither ever enters the verdict itself.
every answer is a probability, never a proof. we never read 0% or 100%. and when we're not sure, we say so.
we don't put a single accuracy number on amige., because no one honest can. what we can show you is the work: which detectors ran, how they leaned, the model a thing resembles, and where amige. chose to hold back. that's the whole posture — it's real? or is it?