What this tool is for
A transparent first review, not an automated accusation
AI-assisted writing can be difficult to discuss because polished structure, repeated transitions and generic wording also appear in templates, translations, formal reports and writing by people using a second language. A black-box percentage hides that ambiguity. This tool instead exposes the measurable features that moved its index.
Use it to inspect a draft, compare two versions of the same text or identify passages that deserve closer reading. The report can help an editor ask better questions about sources, revision history and factual support. It cannot establish who wrote the text, which model may have been used or whether assistance violated any rule.
The detector deliberately runs without a remote AI model. Its formula is deterministic: the same text, language and release produce the same result. That makes the output reproducible, but does not make the heuristic universally accurate.
What the score actually means
A valid sample starts at a neutral base of 45. Visible contributions are added for features such as unusually even sentences, repeated short sequences and dense template-like phrasing. Contributions are subtracted for features such as strong variation, concrete details and personal context. The total is bounded between 5 and 95.
The result is an AI-pattern index, not a probability. A score of 80 does not mean there is an 80% chance that AI wrote the text. Confidence describes only whether the sample is long enough to measure its rhythm consistently; it does not certify that the classification is correct.
- Below 36: few measured AI-like patterns.
- 36–56: mixed profile.
- 57–69: some AI-like patterns.
- 70–95: many AI-like patterns.
- No score: fewer than 80 units or four sentences.
A responsible review workflow
- Paste text or load a local TXT/Markdown file.
- Choose the language that best matches the sample.
- Check the sample counters before running the analysis.
- Run the detector and read the score formula.
- Open the strongest weighted signals and compare them with the source.
- Ask for drafts, sources or revision history when context requires it.
- Export the measurement report without the original text if you need an audit note.
Reproducible diagnostic
Two samples show what the formula rewards and penalizes
The built-in English samples are fixed in release 20260723-ai-detector4. Anyone can load them and reproduce the values below. They demonstrate the algorithm; they are not a benchmark of real-world accuracy.
| Sample | Dominant measurements | Result |
|---|---|---|
| Pattern-heavy example, 122 units and 8 sentences | Very uniform sentences +14; repeated sequences +14; template phrases +12; transitions +8; vocabulary −5 | 88/100, many AI-like patterns, medium sample confidence |
| Varied workshop narrative, 150 units and 11 sentences | Concrete details −8; vocabulary variation −5; first-person context −5 | 27/100, few AI-like patterns, medium sample confidence |
Measurements used by the local heuristic
All measurements are calculated from the pasted text. Thresholds are fixed in the public browser code and are not learned from visitors.
| Measurement | How it is calculated | Possible effect |
|---|---|---|
| Sentence rhythm | Coefficient of variation of sentence lengths. | Very even rhythm adds up to 14 points; strong variation subtracts 8. |
| Vocabulary variation | Average unique-unit ratio in moving windows of 50 units, or 40 for Chinese. | Low variation adds up to 10; high variation subtracts 5. |
| Repeated sequences | Repeated 3-unit sequences, or 4-character sequences for Chinese. | High repetition adds up to 14. |
| Template phrases | Exact matches from a small published list for the selected language. | Density can add up to 12. |
| Transitions | Language-specific formal transition phrases per 100 units. | High density can add 8. |
| Punctuation and paragraphs | Number of punctuation types and variation of paragraph lengths. | Uniformity adds points; variation may subtract points. |
| Concrete and personal context | Numbers, quotations, URLs and first-person phrases per 100 units. | Can subtract up to 13 points. |
How to use the report without harming people
Detection errors have real consequences in education, publishing and employment. A score should never replace a conversation or evidence about the writing process. The safest use is editorial: locate regular patterns, verify factual claims and compare drafts.
- Do not accuse a student or employee from this result alone.
- Allow the writer to explain tools, translation, templates and accessibility support.
- Prefer version history, notes, cited sources and oral explanation over detector output.
- Document the tool version and exact sample if a report is retained.
- Apply the same review process consistently to comparable writers.
Language, genre and fairness matter
The phrase lists are localized, but the thresholds are general demonstrations rather than validated language-specific models. Short, formal or translated writing can look highly regular. Research has also found that detectors may falsely flag writing by non-native English writers.
Poetry, legal clauses, product templates, code documentation, bullet lists and heavily edited prose are outside the intended comparison range. When a sample mixes languages, analyze coherent sections separately and interpret both cautiously.
- Choose the page language that matches most of the text.
- Avoid comparing scores across different languages or genres.
- Do not rewrite natural language merely to reduce a detector score.
- Treat accessibility tools, grammar correction and translation as contextual facts, not evidence of misconduct.
Text stays on the device
Pasted text and loaded TXT/Markdown files are read only by this browser tab. The detector does not send the source text to Xephory or an AI provider, and it does not store the source in localStorage.
- The 50,000-character limit bounds browser work.
- The 1 MB file limit is checked before reading.
- Copied and downloaded reports contain measurements, not the source text.
- Language preference may be stored under the shared Xephory language key.
- Analytics and Auto ads follow the shared consent system but are not given the pasted text by this tool.
Known limitations
Any writing style can trigger or avoid these thresholds. The detector sees surface patterns, not intent, provenance or the writing process.
- Paraphrasing and human editing can sharply change the score.
- A human can write regular, generic prose and receive a high score.
- An AI system can produce varied, specific prose and receive a low score.
- The phrase dictionaries are intentionally small and cannot represent every language variety.
- The tool does not identify a model, plagiarism, factual accuracy or policy compliance.
- Scores from different releases should not be treated as directly comparable.
- No threshold is suitable as an automatic disciplinary or employment rule.
Research and official context
Sources reviewed on 23 July 2026. They support the limitations and responsible-use guidance, not the accuracy of this Xephory heuristic.
- OpenAI: retired AI text classifierOpenAI withdrew its classifier because of its low rate of accuracy and documented important limitations.
- Liang et al.: bias against non-native English writersPrimary research on false positives and language-related bias in several detectors.
- Sadasivan et al.: Can AI-Generated Text be Reliably Detected?Primary research on practical detector reliability and paraphrasing attacks.
- NIST AI 100-4: synthetic content transparencyOfficial overview of detection, provenance, watermarking and transparency approaches.
Frequently asked questions
Can this detector prove that AI wrote a text?
No. It measures surface patterns and cannot observe authorship, prompts, drafts or editing history.
Is the score an AI probability?
No. It is a bounded index built from a base score and visible fixed contributions.
How much text should I use?
At least 80 units and four sentences are required. Around 250 units or more makes rhythm measurements more stable, but not necessarily correct.
Why can a human text receive a high score?
Templates, formal structure, translation and second-language writing can be regular or generic in the same ways measured by detectors.
Why can AI text receive a low score?
Prompts, paraphrasing and human editing can add variation, details and personal style that lower surface-pattern scores.
Is my pasted text uploaded or saved?
No. The analysis is local, source text is not stored, and exported reports exclude it.
Can I compare scores across languages?
That is not recommended. Tokenization, writing conventions and phrase lists differ, and the thresholds are not a cross-language benchmark.
What should I do with a high score?
Review the visible signals, verify facts and sources, and ask for process evidence. Do not make a high-stakes decision from the score alone.