How it works
How AI Detectors Work: Our Method
How AI detectors work, and how our AI detector turns an engine score into a verdict, with the evidence and limits of each check.
Last updated Oct 29, 2026
The short version
This page explains how AI detectors work in general, and exactly how ours turns an engine's answer into a verdict. For a Deep check we send your text or image to a dedicated AI-detection engine, keep its answer exactly as returned, place the score on a scale with a wide inconclusive zone, and show the evidence and limits next to it. The free quick check runs an open detection model on our own server and gives one verdict for an English text, with the same honest middle zone. We do not run a chatbot over your content, combine several engines into one number, or draw evidence an engine did not return.
How AI detectors work
An AI detector is itself a machine-learning model. It is trained on large collections of writing or images whose origin is known, some made by people and some generated by AI models, and it learns the statistical patterns that tend to separate the two. When you submit new content, the detector does not look anything up or find a matching source. It measures how closely your content resembles each group and returns a score.
Text detectors look at patterns in word choice and sentence structure. Text generated by a language model tends to be more predictable from one word to the next, more even in sentence length and rhythm, and more regular in phrasing than most human writing. People write with more surprises: unusual words, uneven sentences, personal detail and small mistakes.
Image detectors look for traces that image generators leave behind: regularities in texture, noise and fine detail, lighting and edges that do not quite agree, and patterns in how pixels are arranged that cameras and editing tools rarely produce. Some images also carry provenance metadata such as Content Credentials, but that metadata is easy to remove, so its absence proves nothing.
Because a detector measures resemblance rather than proving where content came from, its answer is an estimate. Human writing that is formal, formulaic or heavily edited can look machine-like, and AI output that has been paraphrased or edited by hand can look human. That is why we show a wide inconclusive zone and the limits of every check. See accuracy and limits for where detection fails.
How our AI detector works
The rest of this page describes our own checks: which engines we use, what their scores mean and how we choose a verdict.
Engines and versions
| Content | Engine | Version |
|---|---|---|
| Text, Deep check | Winston AI content detection | 5.0 |
| Text, free quick check (English) | Gradient, an open model, on our own server | gradient-int8@c2e8b6d |
| Images | Winston AI image detection | 5 |
Every report records the engine and version it used, so reports stay comparable when an engine is updated. Voice, video and music are not offered until an engine passes our own tests.
What the score means
For the Deep check, the engine returns a human score from 0 to 100: higher means the content looks more like human work. We show this number unchanged. For text we also show an "AI-leaning reading", which is simply 100 minus the human score. It is a different way to read the same number, not a calibrated probability.
For images the engine may also return its own AI probability between 0 and 1. We show it as returned when it is in that range, and leave it out otherwise.
From score to verdict (Deep check, mapping v1)
| Human score | Verdict |
|---|---|
| 0–29 | Likely AI-generated |
| 30–69 | Inconclusive |
| 70–100 | Likely human-made |
The middle zone is deliberately wide. Mixed signals are common with edited, translated or short content, and "inconclusive" is a real answer. If we change the thresholds we publish a new mapping version, and old reports keep the version they were made with.
Free quick check
The quick check gives English text a single verdict for free, without an account. It is a smaller, faster check than the Deep check and does not score sentences.
- Model. Gradient, a DeBERTa-v3 classifier released under the MIT licence by Team DACTYL as part of their PAN at CLEF 2026 system. We run it compressed to 8-bit numbers on our own server in Germany; the compressed model gives the same verdict as the original for 99.7% of our test texts. Your text is not sent to any third-party detector.
- What it reads. English only, 150 to 1,500 words. The model reads the first 512 tokens, about 400 words, and the result says how many words that was. AI-written text after that point is not seen.
- The score. An AI-likeness score from 0 (human-like) to 1 (AI-like). It is the model's own output, not a probability, so we never show it as a percentage.
- One score for the whole text. The model cannot tell which parts of a text were written by AI. A text that mixes your own writing with AI can land in any zone. For sentence-level evidence, use the Deep check.
| AI-likeness score | Verdict (mapping gradient-en-v1) |
|---|---|
| 0.90–1 | Likely AI-generated |
| 0.54–0.90 | Can't tell |
| 0–0.54 | Likely human-written |
We chose the thresholds on our own test set of human-written and AI-written English text. The lower line is set so that about 99% of the everyday human writing in our test lands in "Likely human-written". Only scores of 0.90 and above are called AI, because essays by English learners more often land between the two lines. The test and its results are on the Accuracy and limits page.
What we keep. Nothing of the text itself: it is analysed in memory and discarded. We keep a usage record for 30 days to apply the daily limits and to watch the service: time, verdict, score, word counts and your account or a daily-changing hash of your IP address. Guests are limited to 3 quick checks a day and must pass a Cloudflare Turnstile check; signed-in users get 10 a day.
Evidence
Text
- Sentence scores — the engine scores each sentence; we highlight only the sentences it actually scored.
- Hidden-character tricks — zero-width characters and look-alike letters, which are used to slip text past detectors.
- Detected language — the language the engine read.
Images
- Content Credentials (C2PA) — if the file carries a credential, we show what it claims, such as the tool that made it. We read the credential; we do not yet verify its certificate chain.
- AI watermark — whether a known invisible AI watermark was found. Many generators add none, so "not detected" says little on its own.
- Camera data (EXIF) — present or not. Messaging apps and screenshots strip EXIF from real photos, so missing EXIF is never treated as evidence of AI.
- Region heatmap — the current engine scores the whole image and returns no heatmap, so we do not draw one.
What we analyse
Deep check reports state the coverage: how many words of the text, or that the whole image was analysed at its original size, plus what it cost and how long it took.
Credits
The quick check is free. For the Deep check, text costs 1 credit per 100 words (rounded up); an image costs 3 credits. Credits are reserved when a check starts and charged only when a usable report is ready. You are never charged more than the quote shown before the check.
Known failure cases
See Accuracy and limits.