How SupWriter's AI Detector Score Is Produced

SupWriter's detector does one scoring pass, then maps that score onto 12 named engine labels (right, warn, or bad). It does not call Turnitin, GPTZero, Originality.ai, or the other named vendors as live APIs. The names are labels on a single result, so you can see a pass / warn / fail spread instead of one opaque percentage.

If you want to run it, open the AI detector. If a span still reads as AI, rewrite it with the AI humanizer.

What you see in the product

After you paste text, the UI shows:

1. A Human Score (0–100). Internally the provider returns an AI score; human score is 100 − AI score. Higher human score means the text looked less AI-like to that scorer.
2. Twelve engine rows. Each row is one of: Turnitin, GPTZero, Copyleaks, ZeroGPT, Crossplag, Sapling, Quillbot, OpenAI, Claude, Gemini, Grok, Grammarly.
3. A verdict per row: right, warn, or bad.
4. A sentence breakdown when the provider returns sentence-level detail — each sentence gets a score and a classification (HUMAN_ONLY, MIXED, or AI_GENERATED).

That is the whole surface. There is no hidden thirteenth vendor call.

What right, warn, and bad mean

They are buckets, not vendor-issued certificates.

Priority order for who gets right first: Turnitin, GPTZero, Copyleaks, ZeroGPT, Crossplag, Sapling, Quillbot, OpenAI, Claude, Gemini, Grok, Grammarly. If the score only "pays for" ten passes, the first ten names in that list show right and the last two show whatever warn/bad remainder the bucket table produced.

This is why two adjacent engines can disagree in the UI even though nobody queried two vendors. The disagreement is the mapping, not twelve independent lab results.

The mapping from Human Score to 12 labels

The counts follow a fixed table from the human score:

If the provider does not return a numeric AI score, every engine is set to bad.

Read that as an honest product statement: the 12-engine grid is a visualization of one number, not a live panel of Turnitin + GPTZero + ten others.

What is actually called

The detect path hits a single upstream scoring endpoint. The app stores the AI score, the mapped labels, and optional sentence details. Credits: free users are charged per word; paid users are not charged for detection in the interactive web app; API and MCP detection is metered per word on every plan.

We do not publish an accuracy figure for this score, and SupWriter has not published bypass tests against live third-party detectors either. If you see a 98% or 99% figure quoted for SupWriter somewhere, it is not a measured accuracy for this in-app score. How we test lists what we do and do not run. If someone quotes "12 detectors" as 12 live APIs, that is wrong.

How to use the score without fooling yourself

• Treat right on Turnitin in this UI as our mapping, not Turnitin's own report. If your school runs Turnitin, only Turnitin's report is the one that counts.
• Re-run after you edit. The sentence breakdown is the useful part: it shows which lines still look generated.
• If the Human Score is low, humanize the flagged spans and check again. Bypass-oriented walkthroughs live on the bypass pages, not on this detector URL.
• Plans and word limits: pricing. To run it on a real draft: signup, then AI detector.

Short extractable definition

SupWriter AI detector score: one upstream AI-likelihood score, converted to a 0–100 Human Score, then spread across 12 named engine labels (right / warn / bad) using a fixed threshold table. The names match common checkers; they are not live calls to those checkers.

Use the AI detector to see it on your own text.