How the AI writing checker scores your text

What the score means

The style signal index describes wording patterns under six editing rules. It does not estimate the probability that a person or AI wrote a text. A low score is not a certificate of originality, quality or factual accuracy.

We check each sentence in its paragraph context. Nearby paragraphs are included when available. Four sentence rules produce separate judgments. Two additional rules examine paragraph structure; these findings appear separately so they do not add the same evidence to the document score twice.

Six rules, with exceptions

Generic opening

Flag interchangeable background that adds no information needed by this reader. “In today’s fast-paced world, productivity is more important than ever” can often be removed. “In today’s release, invoice exports include a date column” provides specific information and should remain.

Unsupported emphasis

Flag sweeping claims of value or transformation without concrete support in the supplied context. A documented, limited claim is different from “a revolutionary solution for everyone.” Missing support in the text does not prove the claim false.

Unclear source

Flag appeals to experts, research or consensus when the supplied context contains no identifiable source. A nearby citation, named report or link counts. The checker does not visit links or establish that a citation supports the claim.

Redundant wording

Flag synonyms, modifiers and framing that can be removed without losing information. Keep uncertainty, negation, technical distinctions and conditions. “May help” and “will help” make different claims.

Repeated idea

At paragraph level, look for sentences that restate the same point without adding a fact, example, condition or explanation. A useful summary or necessary restatement is an exception.

Mechanical structure

At paragraph level, look for repeated syntax that causes redundancy or forced transitions. Steps, feature lists and intentional parallel structure are not problems by themselves.

Severity and display thresholds

Each rule uses five descriptive levels: no issue; optional minor preference; a clear local problem; a problem that dominates the target; and a target almost entirely composed of the problem. The model returns a distribution over these levels. The expected level ranges from zero to four.

A suggestion is displayed when the combined model weight on levels two through four is at least 0.8. It is marked priority when levels three and four together reach 0.8. These are initial display thresholds, not measured accuracy rates. Model confidence describes the distribution, not a guarantee of correctness.

For each sentence, we take the largest of its four rule scores. The document index is 25 times the word-weighted average of those sentence scores, rounded to an integer. We do not add correlated rule scores together. If more than a quarter of the sentence word weight has an elevated but uncertain leading judgment, the interface withholds the numerical score and shows mixed signals.

How highlighting works

A tinted sentence means there is a suggestion to review. The checker can then examine exact word or phrase candidates inside that sentence. Underlines appear only where a contextual judgment supports a more precise clue. No banned-word list decides that a term is wrong. Sometimes the whole sentence is the most honest location.

Click a sentence or a paragraph issue to read the rule and editing advice. Dismissing a suggestion hides it; it does not improve the model score. Any edit invalidates the old highlights. Run a new check before relying on the result.

Limits and versioning

The first version is designed for English prose. We keep the original character positions and version the rules and scoring logic. Current rule version: 2026-10-28.1. The checker does not verify facts, inspect linked sources, detect plagiarism or guarantee a result on another detector. Very short, quoted, technical or deliberately repetitive writing needs your judgment.

Try the editor, learn the article workflow, or review meaning-preservation checks.

Long paragraphs and word clues

For very long paragraphs, each sentence is reviewed with a nearby context window to keep checks responsive. Paragraph-level rules still read the complete paragraph. A source or explanation outside the sentence window can therefore be missed; review attribution suggestions against your full document.

Word and phrase clues use a separate judgment over exact spans of the original text. A clue appears only when the probability assigned to the two strongest evidence levels totals at least 80%. Ambiguous clues are omitted. These thresholds are product settings, not a calibrated probability that AI wrote the text.

Local rewrites receive a limited nearby context window. They are charged for the selected words only. If a suggestion depends on a distant paragraph, edit the text yourself or review the complete article instead.