What Does an AI Detector Score Mean?

Detector Checker’s 0–100 AI-likeness score is an ordinal pattern signal: a higher number means the passage matched more of the AI-associated patterns measured by the current document classifier. It is not the probability that AI wrote the text, the percentage of AI-written words, or proof of authorship.

  • Score firstRead the document-level signal and its band.
  • Context secondReview agreement and sentence cues separately.
  • Never a verdictUse provenance and human review for permitted revision.
On this page

Result anatomy

The three parts of an AI detector result

The layers answer different questions. The document score is primary; agreement and highlights add independent context for review.

Document-level result

69 / 100

Higher AI-like signal

01

AI-likeness score

Means: how strongly the whole passage matched patterns measured by the primary document classifier.

Does not mean: probability, percentage of AI-written words, authorship, intent, or misconduct.

02

Auxiliary signal agreement

Means: how closely the visible perplexity, semantic, and statistical indicators align with one another.

Does not mean: confidence, accuracy, ambiguity, or validation of the document score.

03

Sentence-level cues

Means: where separate sentence-level thresholds suggest closer reading may be useful.

Does not mean: that a highlighted sentence was proved AI-written or caused the document score.

Illustrative layout only. The values above show how to read the interface; they are not an authorship finding or an accuracy claim.

Worked example

How to interpret one illustrative sample response

The values below mirror a recorded product API response and are included only to explain the fields. This is not an accuracy test, a benchmark row, or an authorship finding.

Score69 / 100
BandHigher AI-like signal
Auxiliary agreement83 / 100
Perplexity indicator42 / 100
Semantic indicator35 / 100
Statistical indicator52 / 100
Sentence cues3 Indeterminate
Analyzed3 sentences
Detected languageEnglish
  1. Start with 69 and the Higher band. The whole passage matched document-level AI-associated patterns strongly enough for the product to return its Higher band. “69” is not a 69% probability and does not establish who or what wrote it.
  2. Read 83 agreement separately. The visible auxiliary indicators were 42 for perplexity, 35 for semantic, and 52 for statistical. Agreement summarizes how those three indicators align; it does not turn 69 into a more certain result.
  3. Keep sentence thresholds separate from document bands. The three sentence cues in this response were each labeled Indeterminate, with sentence values of 44, 49, and 67. Their separate labels do not decompose, average into, validate, or contradict the document-level Higher result.
  4. Do not infer ground truth from the response. To investigate authorship or process, bring in drafts, revision history, notes, sources, citations, task instructions, and the writer’s explanation.

Why the layers may look different: the overall score, the three visible auxiliary indicators, and sentence cues are produced separately. Read the product’s displayed label for each layer instead of applying the document-level score bands to sentence cues.

Score bands

How to read Lower, Indeterminate, and Higher results

Use the descriptive label shown with the score. The label tells you which current review band the product returned; none of the bands proves authorship.

0–39

Lower AI-like signal

The passage matched fewer of the detector’s AI-associated document patterns.

Do not conclude: “This text is definitely human.” Edited or atypical generated writing can receive a Lower result.

40–49

Indeterminate

The current evidence did not support a clear Lower or Higher band assignment.

Do not conclude: “The text is a known human–AI mix.” Indeterminate describes the detector’s result, not the writing workflow.

50–100

Higher AI-like signal

The passage matched more of the AI-associated patterns measured by the document classifier.

Do not conclude: “AI use is proved.” Human formal, technical, translated, or template-driven text can produce a false positive.

Boundary note: use the category displayed beside your result. The public benchmark documents a small historical score-mapping audit at band boundaries; rounded scores should not be reverse-engineered into a probability. See the benchmark score-mapping disclosure.

Supporting context

What does Auxiliary signal agreement mean?

It describes only how closely the three visible auxiliary indicators point together. It is intentionally separate from the primary document score.

Perplexity indicator

Supporting context related to predictability of word choice across the sample.

Semantic indicator

Supporting context related to uniformity of meaning, rhythm, and sentence relationships.

Statistical indicator

Supporting context related to repetition and distributional patterns in the text.

The same AI-likeness score has the same product meaning whether auxiliary agreement is high or low. Agreement does not increase, decrease, reverse, or validate the document-level result. For a process overview, read how Detector Checker analyzes text; for versioned technical context, see the evaluation methodology and model card.

Passages to review

What do highlighted sentences mean?

Sentence highlights are independent review cues. They tell you where a separate sentence-level threshold was crossed, so you can begin a closer read.

Illustrative cue—not an analyzed claim

A highlighted sentence belongs on a review list; it does not become proof of machine authorship.

Read the surrounding paragraph

Check whether the line is unusually generic, formulaic, smooth, repetitive, or stylistically different from nearby writing.

Check evidence and specificity

Look for sources, examples, reasoning, quotations, first-hand detail, and accurate attribution—not merely a preferred detector score.

Compare the writing process

Use drafts, revision history, notes, and known samples when authorship matters. The detector cannot reconstruct those facts.

Expect score/highlight differences

Sentence cues do not decompose the document score. A document and its individual lines can produce different-looking signals.

Mixed or inconclusive evidence

How to read an Indeterminate AI detector result

An Indeterminate result is a valid product outcome, not a technical error. It means the current analysis did not support a clear Lower or Higher assignment.

Why it can happen

  • The passage is short or contains little connected prose.
  • The style is highly formal, technical, translated, or template-driven.
  • The draft has been heavily edited, paraphrased, or assembled from different sources.
  • The measured document patterns sit close to the current decision boundary.

What to do next

  1. Confirm that you submitted a complete, coherent passage.
  2. Review any sentence cues and the surrounding context.
  3. Check provenance, drafts, citations, and revision history.
  4. If you legitimately revise the text, analyze the revised version as a new result.

Do not convert Indeterminate into Human, AI, or “50/50.” A technical failure or a result that was not run is also a separate state and should never be silently relabeled as Indeterminate or Lower.

Scenario matrix

How to read common score and agreement combinations

Agreement tells you whether the visible supporting indicators align with each other. It never tells you whether the document classification is correct.

Lower score · closer agreement

Fewer document-level AI-like patterns were measured, while the three auxiliary indicators were more similar to one another. This does not prove human authorship.

Lower score · less agreement

The document result remains Lower. The auxiliary indicators simply diverged more; do not reinterpret that divergence as hidden uncertainty or confidence.

Higher score · closer agreement

More document-level AI-like patterns were measured and the auxiliary indicators aligned more closely. Review the text, but do not treat the combination as proof.

Higher score · less agreement

The document result is still Higher. Less auxiliary alignment neither cancels the score nor proves it unreliable; start with highlighted passages and external context.

Evidence and limits

Can an AI detector score be wrong?

Yes. Pattern classifiers can produce false positives, false negatives, and Indeterminate outcomes. The error risk changes with the sample, language, domain, model, prompt, length, translation, and editing history.

False positive

Human-written text receives a Higher AI-like signal. Formal, technical, translated, polished, non-native, or template-driven prose can share measured patterns with generated text.

False negative

AI-generated or AI-assisted text receives a Lower signal. Editing, paraphrasing, creative variation, mixed workflows, and model differences can weaken measured cues.

What the published benchmark actually establishes

Detector Checker’s first-party benchmark evaluated release rel-2026-08-26-00ad353f8f70 on 880 English texts in 440 source-matched human–AI pairs from a public RAID train_none subset. It reported 87.3% strict balanced accuracy, a 7.5% human false-positive rate, an 11.1% AI false-negative rate, and 3.4% Indeterminate outcomes.

Those figures describe that disclosed release, sample, and protocol. They do not guarantee performance for every current model, language, document type, adversarial rewrite, or editing condition. Read the full AI detector benchmark, including denominators, uncertainty intervals, tested domains, data files, and limitations.

  • Short fragments: headlines, quotations, lists, and isolated sentences provide limited document context.
  • Translations and multilingual writing: input is accepted, but per-language accuracy has not been validated for every language, dialect, or genre. See supported input and language limitations.
  • Mixed authorship: one score cannot recover how much assistance occurred or which parts came from which source.
  • Source-model identification: the result cannot prove that text came from ChatGPT, Claude, Gemini, or another specific system.
  • Quality and truth: an AI-likeness score does not measure factual accuracy, originality, plagiarism, usefulness, or writing quality.

For edge cases and error handling, read AI detector limitations and false positives. AI detection and source matching are different tasks; see AI detection vs. plagiarism checking.

Responsible workflow

What should you do after reading the result?

Treat the detector as a low-stakes review aid for your own text or text submitted with informed consent. It can point to passages worth revising or fact-checking, but it is not a triage tool for people, submissions, cases, or eligibility.

  1. 1

    Check the input

    Use a coherent passage and confirm the correct language setting. Exclude navigation, references, boilerplate, or copied prompts if they are not part of the writing under review.

  2. 2

    Read the score and band

    Record the displayed result without converting it into a probability or percentage of AI-written content.

  3. 3

    Inspect sentence cues

    Review highlighted passages in context. Ask whether wording is generic, repetitive, unsupported, or inconsistent with the rest of the draft.

  4. 4

    Gather external evidence

    Consult drafts, revision history, citations, sources, task requirements, and the writer’s explanation. These can bear directly on process; the score cannot.

  5. 5

    Choose a low-stakes next step

    Use the result only to guide revision, fact-checking, source verification, or a consented editorial discussion. Do not use it to evaluate a person or trigger any adverse action.

If your text was unexpectedly flagged, follow the practical steps in what to do when an AI detector flags your writing. Improve the work for clarity, evidence, accuracy, and authentic voice—not merely to chase a lower detector score.

Privacy: do not submit confidential, regulated, medical, financial, identity, or other sensitive material. Review the Privacy Policy and security guidance before using the tool in a routine workflow.

Quick answers

AI detector score FAQ

Is an AI detector score the percentage of text written by AI?

No. Detector Checker’s score is an ordinal document-level signal. A score of 80 does not mean that AI wrote 80% of the words or that there is an 80% probability of AI authorship.

Does a Higher AI-like signal prove that AI wrote the text?

No. It means the passage matched more of the patterns measured by the current document classifier. Human writing can receive a Higher result, so authorship requires broader evidence.

Does a Lower score prove the text is human-written?

No. Generated text may receive a Lower result after editing, paraphrasing, creative variation, or because its patterns differ from those emphasized by the detector.

Is Auxiliary signal agreement the detector’s confidence?

No. It describes only how closely the three visible auxiliary indicators align with one another. It is not confidence, probability, accuracy, ambiguity, or validation of the primary score.

Do highlighted sentences identify the AI-written parts?

No. A highlight marks a separate sentence-level review cue. It does not prove the sentence’s source, and sentence cues do not decompose or determine the document-level score.

What does an Indeterminate score mean?

It means the current analysis did not support a clear Lower or Higher assignment. It does not mean the text is known to be 50% human and 50% AI, and it is not the same as a technical failure.

Can the score identify ChatGPT, Claude, Gemini, or another model?

No. Detector Checker reports general AI-likeness signals; it cannot prove which person or model produced the text.

Should I run the detector again after editing?

A second scan can be useful after legitimate revision. Treat it as a new review signal and keep the original result, edits, and context together if the review needs an audit trail.

Need a broader answer? Visit the complete AI Detector Checker FAQ.

Use the result in context

Review text for AI-like writing signals

Run a first-pass check, then use the score, auxiliary agreement, sentence cues, and external evidence as separate parts of a responsible review.

Last reviewed: