# What your AI detection score means

How to read a ScanForAI result: what the bands mean, why a score is a signal not a verdict, and how false positives on formal or non-native writing happen.

- Source: https://scanforai.com/docs/interpreting-scores
- ScanForAI — AI-writing detection. A score is a signal, not a verdict.

---

# Interpreting your score

Read this before acting on any number this tool gives you.

## The score is a density, not a probability

The 0–100 score measures how densely your text carries patterns typical of *raw, unedited* model output. It is **not** “the probability this was written by AI”, and **not** “the percentage written by AI”.

Higher means denser AI-style patterning — **not** “more likely to be AI”. A high score on formal or edited human writing is a known, explainable outcome (below), never proof.

Low reads human — nothing to do. Moderate shows some machine patterns; normal for edited or formal writing, worth a second look only if it shouldn’t be there.

Elevated and High mean dense AI-style patterning. Open the breakdown, check the coverage and confidence lines, and treat it as a reason to **ask** — for drafts, notes, or version history — never as a verdict. Formal and non-native writing land here for innocent reasons, so a high score is a starting point for a conversation, not an answer.

## Check the coverage line first

Each check refuses to judge when there isn’t enough text — you’ll see something like “9 of 11 checks had enough text to judge”. On short texts most checks abstain and the score rests on almost nothing, so treat anything under ~300 words as weak evidence, whatever the number says.

## Confidence — how far to trust the read

Every result carries a `confidence` of high, moderate, or low, with a one-line reason. It is deliberately cautious: short text, or a score in the ambiguous middle (roughly 30–55 — “some patterns, but not clearly one way”), returns **low** — the same abstain-rather-than-guess rule the per-check signals follow. **High** means a decisive score across enough text with the trained model behind it — and still is not proof.

## Text types

Formal writing reads “AI-ish” for reasons that have nothing to do with AI: few contractions, even sentence lengths, a careful register. So the scanner lets you set the **text type**, and it re-baselines — easing the formality signals for types where they would otherwise over-flag, instead of counting them against you. Each type below is its own setting, even where two share the same underlying register.

- **Student essay / assignment** — academic register; the formality patterns that over-flag careful coursework are relaxed.
- **Research paper / article** — the same academic register; dense citations and formal structure aren’t held against you.
- **Blog / web article** — expository web writing; a conversational-but-structured register.
- **Creative / fiction** — narrative voice; literary devices (aphorism, the reused simile) are expected there, not penalised.
- **General (auto-detect)** — the scanner infers the register from the text when you don’t pick one. Other document types — business writing, emails, cover letters, reports, technical docs — ride on this.

Picking the right type gives a fairer score; it never invents signs of AI writing that aren’t there.

## Non-native & ESL writing — a separate thing

This isn’t a text type — it’s about *who wrote it*, and it cuts across every type above. Learned, careful English shares the surface habits of machine text (even sentences, formal word choice, few contractions), so non-native English is the *most documented false-positive class* across every detector — independent tests have measured other tools flagging non-native writers above 60%. ScanForAI eases exactly those formality signals whichever text type you’re in, so second-language writing isn’t punished for being careful. Choosing an academic type, or leaving auto-detect on, applies the same easing. [Why non-native English gets flagged →](https://scanforai.com/esl-writers)

## Known false positives

Human writing scores elevated when it is: formal or academic (few contractions, uniform sentences), written by non-native English speakers (the most documented harm across all detection tools — other tools have been measured falsely flagging non-native writers above 60%), heavily edited or house-styled, or simply short. If you’re contesting a score, open the per-signal breakdown: every claim this tool makes comes with its reason, which is the point.

## Known false negatives

A light paraphrase or human editing pass removes most surface patterns. A low score means the text *reads* human — it cannot certify how it was made.

## Never use this to accuse anyone

No text detector — free or paid — is reliable enough to be the sole basis for an accusation, a grade penalty, or an employment decision. Vendors advertising “99% accuracy” are routinely measured far lower by independent tests.

**Use the score as one signal among many, keep a human in the loop, and give writers the breakdown so they can respond to specifics.**