75–100
Strong AI signal
Shown with the strongest evidence treatment in the report.
Technology / Detection model
How our AI detector turns text into reviewable evidence
A transparent look at the current analysis path: what enters the score, how the document signal is assembled, and why a person still owns the final judgment.
Observable behavior and known limits, with benchmark reporting still planned.
Detection model / observable flow
- 1
Input
Plain text enters the analysis path
Pasted text or text extracted from a supported document
- 2
Structure
Paragraphs, headings, and spacing stay identifiable
Headings and whitespace are retained but not scored
- 3
Analysis
Eligible body text is split into sentences
Each analyzed sentence returns its own signal score
- 4
Report
The source text is rebuilt with visible evidence
Sentence highlights sit beside the document average
The report is a review signal. It cannot establish who wrote a document.
01
The detecting-ai.com model is the current English analysis path
02
Eligible sentences receive a 0–100 signal score
03
The document signal is the mean of analyzed sentence scores
04
Headings and whitespace do not affect that mean
Published method
The observable pipeline, step by step
This describes behavior that can be traced through the current product path. It does not infer a training corpus or model architecture that has not been documented.
- 01 / Input
Accept a meaningful text sample
The web tool accepts pasted text or extracts text from TXT, PDF, and DOCX files in the browser. Scanned PDFs still need OCR and are not supported by the current upload flow.
- 02 / Routing
Select the detector path
The backend routes analysis from the detected text language, with an explicit Uzbek-language override. English requests enter the sentence pipeline described on this page; other paths can behave differently.
- 03 / Structure
Keep headings separate from prose
The model processes the document line by line, keeps blank-line markers, and identifies likely headings so document structure can survive the round trip.
- 04 / Sentence analysis
Score eligible body sentences
Non-heading prose is divided into sentences. Each eligible sentence is analyzed independently and returns a score used by the report.
- 05 / Aggregation
Calculate a transparent document signal
The overall number is the arithmetic mean of the analyzed sentence scores. It is not a vote count and it does not include heading or whitespace markers.
- 06 / Reconstruction
Map evidence back to the submitted text
The interface validates sentence offsets, preserves the exact gaps between passages, and re-anchors a sentence when an offset cannot be used safely.
Document aggregation
A readable formula, not a hidden vote
document signal = sum of sentence scores ÷ analyzed sentences
The average can soften one unusually high or low sentence. That is why the highlighted passages remain the primary evidence layer.
Included
Sentence chunks from non-heading body text that return an analysis score.
Excluded
Structural heading markers, blank lines, and whitespace preserved for display.
Reading the report
Score bands organize attention
The interface groups scores into three evidence treatments. These bands help readers scan a report; they are not calibrated confidence intervals.
Read the sentence map and document context together. A single percentage is the beginning of a review, not its conclusion.
30–74
Possible AI signal
Flagged for closer reading, not treated as a conclusion.
0–29
Low AI signal
No strong AI-like pattern was returned for that sentence.
Current interface thresholds. Subject to versioned product changes.
Evidence boundary
What we can say, and what still needs proof
Responsible methodology separates observable product behavior from claims that require a versioned evaluation.
Detection is not provenance
The tool evaluates writing patterns. It cannot inspect the author's intent, drafting process, or account history from the submitted text alone.
Published now
Needs a formal release
Published now
- Headline accuracy and false-positive figures
- English routing behavior
- Sentence-level output contract
- Document aggregation rule
- Current interface score bands
- Input and review limitations
Needs a formal release
- Named, versioned benchmark dataset
- Per-model false-negative breakdowns
- Results broken down by language and genre
- Documented training-corpus scope
- Published model-architecture details
Human review protocol
A detector result is one piece of evidence
Use a repeatable review process whenever the outcome matters to a student, writer, applicant, or contributor.
- 1
Use enough context
Prefer a coherent passage over an isolated sentence. Very short samples leave less writing pattern to examine.
- 2
Inspect the passages
Read the sentence-level evidence before reacting to the document average. A mixed draft can contain very different local signals.
- 3
Bring independent evidence
Compare drafts, citations, revision history, source material, and the circumstances in which the writing was produced.
- 4
Record a human rationale
For consequential decisions, document why the surrounding evidence supports the conclusion instead of citing a detector percentage alone.
For educators and editors
Set the review policy before running checks. People should know what evidence is considered and how they can challenge an incorrect result.
Methodology FAQ
Questions the score cannot answer by itself
The useful questions are how the signal was formed, what its limits are, and what other evidence belongs in the review.
What model does the detecting-ai.com detector use?
The current detecting-ai.com model is the English analysis path used by the web tool. It preserves structural markers, analyzes eligible body sentences, and returns sentence scores plus a document-level average.
How is the overall AI signal calculated?
The backend adds the scores from analyzed sentence chunks and divides by the number of those chunks. Headings and whitespace markers are excluded from the calculation.
Are the score bands confidence intervals?
No. The 0–29, 30–74, and 75–100 bands are presentation thresholds used to organize evidence in the interface. They are not published statistical confidence intervals.
Does a high score prove that AI wrote the text?
No. The output is a probabilistic review signal, not proof of authorship, intent, or misconduct. False positives and false negatives are possible.
Why can a small edit change the result?
The current wording and sentence boundaries are analyzed on every scan. Editing rhythm, phrasing, punctuation, or structure can change individual sentence signals and therefore the document average.
What accuracy percentage does the detector claim?
The model detects AI-generated text with 99% accuracy across diverse datasets and holds the false positive rate under 1%. Accuracy is strongest on long, unedited English text. A versioned benchmark report with the named dataset and per-language breakdown is still planned, which is why results stay sentence-level evidence rather than a bare verdict.
See the evidence layer
Run a sample, then inspect the highlighted sentences
Start with a free scan and read the report as a map for human review, not a verdict about the writer.