What an AI detector score can—and cannot—tell you

Read the definition before the percentage

An “AI score” has no universal meaning. One service may report a classifier confidence. Another may estimate the share of sentences that resemble examples in its training data. A percentage is useful only when you know what the provider says it measures.

Do not assume that a score of 80 means an 80% chance that a particular person used AI. The interpretation depends on how the model was trained, calibrated, and evaluated. It also depends on whether your text resembles the material used in that evaluation.

Expect errors in both directions

A false positive occurs when human writing is flagged as AI-generated. A false negative occurs when generated writing is classified as human. Both can happen. Predictable structure, common phrasing, a constrained assignment, or substantial editing can affect the signals available to a detector.

A published accuracy result describes a particular test setup. It does not automatically transfer to a new model, a different language, a specialist subject, or a very short passage. Ask what was tested, how many examples were included, and which errors were counted.

Short or unusual text has less useful context

A few sentences give a detector little material to analyze. Lists, citations, templates, source code, and heavily quoted passages may also behave differently from ordinary prose. Follow a provider’s documented minimum length and supported languages.

Adding unrelated paragraphs just to reach a minimum can create a new problem: the result may describe the combined text rather than the passage you care about. Keep the assessment tied to a clearly defined document and report its limitations with the result.

Disagreement is a reason to inspect the method

Different services use different data, models, thresholds, and score labels. Their results can disagree without either score being a direct measurement of authorship. Running a text through more tools does not turn those outputs into independent witnesses.

Repeatedly editing until every detector gives a favorable result can also damage writing quality. It may remove useful terminology, introduce awkward phrasing, or change a careful claim. Evaluate the text for its intended reader and verify its factual content.

Use evidence about the writing process

For an important authorship question, look at drafts, notes, version history, sources, and the author’s explanation of their process. Apply a clear policy that distinguishes permitted assistance from prohibited use. Give the writer a meaningful way to explain the evidence.

A detector output should never be the sole basis for a grade penalty, misconduct allegation, hiring decision, or other consequential action. Save the exact provider, date, score definition, text, and known limitations if an assessment needs to be documented.

EasyModRewrite’s local writing signals are descriptive checks, such as repetition and sentence length. They are not a model-based AI probability. If no genuine detection provider is connected, the site does not create a percentage from those rules.