questions,
answered.

amige. is an AI content detector. drop in an image, video, or text and it tells you whether the content is AI-generated, then names the most likely model behind it, like ‘looks like Midjourney v6’.

yes. there is a free tier of 3 scans a day, and you can run one without an account. paid plans add higher image and video limits plus unlimited text.

AI-generated images, video, and text in one place. most detectors handle text only. amige. also runs deepfake and model-attribution checks on photos and clips.

most detectors check text only and return a single score. amige. covers image, video, and text in one scan, names the most likely model behind a file, and shows a per-detector breakdown so you see the reasoning behind a verdict.

amige. runs a panel of independent detectors and shows where they agree or disagree, with a per-detector breakdown instead of one black-box score. no detector is perfect, and accuracy drops on heavily edited or re-shared media, so amige. surfaces the reasoning rather than a false promise of certainty.

amige. attributes content across 90+ generators, from Midjourney, GPT-4o, and Nano Banana to Sora and Veo, and reports a confidence level for the most likely model.

yes. amige. runs deepfake-specific checks on images and video alongside the general AI-generation detectors, and flags the most likely manipulation rather than averaging it away.

use it as a signal, not a verdict. independent research shows every text detector carries a false-positive risk, and a 2023 Stanford study flagged a majority of essays by non-native English speakers as AI. treat any detector output as one input, never the sole basis for an academic-integrity decision.

scans are private by default and become a public permalink only if you choose to share one. see the privacy policy for retention details.

look for the details AI still gets wrong: extra fingers or limbs, warped hands, mismatched earrings or teeth, garbled text, and physics errors in reflections, shadows, and backgrounds. check for a visible watermark or C2PA Content Credentials. these checks weaken as models improve, so a detector that reads statistical fingerprints is the more reliable step.

watch for morphing background objects, faces that flicker or smear at the edges, lighting that does not match the scene, and audio that drifts out of sync with the lips. modern video models reduced these tells, so a detector that scores each frame and the audio track catches what the eye misses.

most detectors score between 65% and 90% in independent tests, below the 98-99% some vendors advertise, and OpenAI shut down its own text detector for poor accuracy. accuracy also drops on compressed, cropped, and re-shared media. a panel of independent detectors fails less often than any single one, which is why amige. shows where its detectors agree or disagree.

text detectors lean on perplexity and burstiness, which measure how predictable and how varied the writing is. flat, simple, or formulaic prose can score like AI even when a person wrote it. a 2023 Stanford study found detectors flagged a majority of TOEFL essays by non-native English speakers as AI, and short text makes the error worse.

Google is rolling SynthID detection into Search and Chrome, but it only flags images from Google and partner models that carry the watermark. most AI images in the wild, from Stable Diffusion, Flux, Midjourney, and face-swap tools, carry no SynthID, so a classifier-based detector is needed to cover them.

a screenshot does not remove SynthID, because the watermark lives in the pixels, not the file metadata, so it survives cropping, filters, and compression. C2PA Content Credentials, by contrast, get stripped when most social platforms re-encode an upload. heavy manual editing can weaken both watermark and classifier signals, but it is not a reliable erase.

sometimes. each generator leaves statistical fingerprints, and a watermark like SynthID or C2PA names the source outright when it survives. amige. reports the most likely model with a confidence level. attribution gets harder after heavy editing or when a brand-new architecture has no training reference.