Methodology and limitations
Understand how the score is produced before you use it
Genutext converts provider outputs into a consistent AI-oriented scale, exposes passage context in paid scans, and treats every result as indicative.
Current pipeline
From submitted prose to a Genutext result
The current implementation is designed for English text and uses Winston AI as the external analysis provider.
- 01
Validate and prepare the text
A submission must contain at least 300 characters. Control characters are removed and whitespace is normalised for paid scans. Text beyond the selected word limit is trimmed before analysis.
- 02
Request an English-language analysis
The prepared prose is sent to the provider's AI-content endpoint with English selected. Paid scans request sentence-level results; the free preview requests a document-level result only.
- 03
Convert human probability to AI probability
The provider returns a human-oriented score. Genutext subtracts that score from 100 so a higher displayed number consistently means a stronger AI-writing signal.
- 04
Aggregate paid sentence results
When sentence scores are returned, Genutext converts each one to the AI scale and uses their rounded average as the paid scan's overall score. If no sentences are available, it uses the converted document score.
- 05
Apply a readable label
The numeric score is grouped into Likely Human, Mixed Content or Likely AI Generated. These labels describe the model output, not a finding about the writer.
Displayed labels
Thresholds organise the output; they do not define truth
| AI score | Genutext label | What remains unknown |
|---|---|---|
| 0–29 | Likely Human | Whether a person wrote every passage or whether generated text was edited |
| 30–59 | Mixed Content | Why different passages have different patterns and what assistance was used |
| 60–100 | Likely AI Generated | Authorship, intent, permitted use and whether any rule was broken |
Accuracy statement
Genutext does not claim one universal accuracy percentage
Performance can change with the generator, detector version, text length, language, genre, editing and evaluation dataset.
Genutext has not published an independent, representative product benchmark that would justify a single accuracy claim across real educational writing. The responsible position is therefore to avoid a headline percentage and describe the conditions and limits of the current implementation instead.
False positives and false negatives are possible. Short, formulaic, highly edited, translated, technical or mixed-authorship text can be especially difficult to interpret. A future validation report should identify the tested pipeline version, dataset, document types, generators, thresholds, false-positive rate and detection rate rather than report accuracy alone.
No model attribution
The current result does not establish that ChatGPT or any particular model produced the text.
No process observation
The detector sees the submitted prose, not drafts, prompts, revisions or the person who created it.
No policy decision
The pipeline does not know whether AI assistance was permitted or disclosed for the task.
No certainty at either end
A high score is not proof of AI authorship, and a low score is not proof of wholly human authorship.
Source matching
Plagiarism mode reports source evidence separately
Paid users can ask the same analysis provider to check accessible online sources.
Genutext keeps up to ten source candidates returned for the detailed browser result. Accessible sources with a positive match are treated as confirmed matches; inaccessible candidates are marked as unverified. Unverified candidates alone do not increase the displayed confirmed plagiarism score.
The result is still a source-matching report, not a plagiarism decision. A reviewer must open the source, inspect the passage, check quotation and citation, and apply the relevant academic or editorial standard.
Retention boundary
The writing and full breakdown are not stored in scan history
Genutext stores the score summary and operational metadata needed for credits and history.
The submitted text is sent to the external provider for analysis, then discarded by Genutext. Paid sentence and source details are returned to the browser and held temporarily in session storage; the server-side scan record retains scores rather than the writing or detailed passages.
Users can request an emailed report while the temporary result is available. See the privacy policy for the full description of account, payment, analytics and service-provider processing.
Use the result responsibly
Inspect the score, passages and alternative explanations
A Genutext result can identify where to review. The conclusion still belongs to a person with the relevant context.