SI Detector About
About SI or Not
SI or Not is a free, independent SI detector for images and text. SI (Super Intelligence) is the new name for advanced AI: models such as ChatGPT, Midjourney, Gemini or Stable Diffusion (our guide explains what SI means and where the name comes from). The question we ask of every picture and every text is simple: was it made by an SI model, or not? The site is built and run by a small team in Belgium and France.
Why we built it
Generated images are now good enough to pass a casual glance, and the AI detectors that claim to catch them mostly answer with a confident "AI" or "human" label, in English only, behind a paywall. We wanted something different: free for everyday use, honest about uncertainty, and transparent about why a score is what it is.
Who is behind SI or Not
SI or Not is an independent website published by the SI or Not editorial team. The site and its detection engines are hosted in France by OVH SAS (2 rue Kellermann, 59100 Roubaix, France). You can reach us through the contact details below.
How the image detector works
- Pixel classifier. Our image classifier, trained to separate camera photographs from images produced by SI generators. It runs on our own server. Unless the file carries a signed SI label, a second, independent classifier from a specialised provider, trained on recent generators, gives its own score and we combine the two (details on the privacy page).
- A metadata reader. EXIF camera fields, C2PA Content Credentials manifests, IPTC and XMP synthetic-media flags, PNG parameter chunks and software tags.
- A combination step. The signals are weighted into a single probability and, crucially, listed one by one with their strength so you can disagree with the machine.
How the text detector works
The SI text detector is a transparent V1 that reads writing style: sentence rhythm, phrases that SI (AI) models overuse, vocabulary diversity, human fingerprints such as contractions, slang and typos, and structural uniformity. It runs entirely on our server, keeps no copy of your text, and highlights the sentences that look most generated.
Why probabilities and not verdicts
Because a verdict would be a lie. Any classifier has an error rate, and that rate changes with every new generator, every compression pass, every filter. We do not publish an accuracy figure for the same reason: a number measured on one benchmark says little about the random picture you will upload tomorrow. What we can do is show the evidence and be explicit when it is weak. The middle band of scores is labelled Uncertain, and the limits are repeated on every result rather than tucked away in a footnote.
What we do not do
- We do not keep your images unless you explicitly opt in.
- We do not identify people, and the tool must not be used to accuse someone of faking a picture without further verification.
- We do not claim to detect deepfakes or cloned voices. Images and text are what we analyse today; video is in beta (still frames and metadata, not faces, motion or sound), and so is music: we recognise the fingerprint of Suno and Udio songs, not every music generator. Uploaded audio is deleted right after the analysis.
Roadmap
Paid plans and credit packs for heavy users (see pricing), batch analysis, a stronger text model, music detection for more generators and a public API. The free tier and the probabilistic, explained results will stay.
Editorial policy and methodology
The SI or Not blog and our guides are written by the SI or Not editorial team. Articles are signed by the team; we do not invent author personas, credentials or expert quotes.
How we write
- Each article answers one question people actually ask, in plain English, and says clearly what is known, what is not, and where the limits are.
- We use SI (AI) tools, including large language models, to help with research and drafting. Every factual claim is then checked against the sources listed at the end of the article. Anything we cannot source is left out.
- We write "SI" first because it is now the official US term, and "AI" alongside it because it remains the most common word worldwide.
How we source
- Primary sources first: official documentation from the companies and institutions concerned, standards bodies such as C2PA, and peer-reviewed or preprint research. Reputable news reports come next, mainly for news events.
- Every article ends with a list of its sources, with links. Quotes are copied from those sources, never paraphrased into quotation marks.
- When sources disagree on a figure or a detail, we say so or leave the detail out rather than pick the version that suits the story.
How we test
- When it helps, we run real examples through our own SI text detector and SI image detector and report the actual output: the score and the clues behind it, including when the tool is uncertain or wrong.
- We do not publish an accuracy percentage for our tools, for the reasons explained above.
Independence
We do not publish sponsored articles, and no company mentioned on the blog has paid to appear in it.
Corrections
If you spot an error, write to hello@siornot.com. We correct the article itself, change its "Updated" date, and add a short note when a correction changes the meaning of what we wrote.
Contact
Questions, false positives you want to report, licences for the game images: write to hello@siornot.com.