Document value
An estimated proportion or probability for AI-like passages in the analysed text.
IB · Academic integrity · AI writing detection
Interpret AI-writing detection correctly.
What Turnitin, GPTZero, Copyleaks and other detectors actually measure—with URLs, access models, result types and clear limits for educational decisions.
Two separate reports
Turnitin Similarity and AI Writing answer different questions.
The Similarity Report finds text overlap with comparison sources. The AI Writing Report evaluates language patterns in qualifying prose and estimates which passages may be AI-generated or AI-revised.
A paper can therefore have low similarity and a high AI-writing indicator—or the reverse. The percentages have different denominators, data sources and meanings; they must not be added or compared directly.
What the specialist detector measures
Language patterns—not the actual writing process.
Providers train classification models on human and machine-generated text. Depending on the tool, predictability, sentence variation, word choice, style and other statistical features are combined.
Passage indicators
Sentences or paragraphs the model regards as particularly conspicuous.
Categories
Depending on provider: human, AI-generated, AI-refined or mixed.
Confidence
How confident the model is in a classification; not the same as the real amount of AI use.
Use precise languageWrite “AI proportion estimated by the tool” or “text proportion classified as AI-like”—not “actual AI proportion”.
Tool comparison · checked 27 August 2026
AI text detectors for controlled counterchecks.
The table describes currently visible access. Plans, free allowances, supported languages and models change frequently; recheck details on the linked provider page before an investigation.
| Tool and URL | Access | What the specialist tool returns | Interpretation for IB work |
|---|---|---|---|
| Turnitin AI Writing Report | Institutional · only when enabled by school/licence | The proportion of qualifying long-form prose classified as likely AI-generated or AI-revised, with highlights. Very low results may be shown without an exact percentage. | Most relevant to the school's workflow, but technically separate from the Similarity Score and explicitly not sufficient evidence by itself. |
| GPTZero | Limited free use · larger scans/allowances require an account or plan | Estimated AI proportion at document level and highlighted phrases/sentences; writing-process and plagiarism features vary by access. | Useful as a second education-focused check. Longer coherent text is more informative than isolated sentences. |
| Copyleaks AI Detector | Limited free use · basic scan/starting credits, regular use is paid | AI classification, highlighted passages and explanatory AI Logic information; multilingual. Plagiarism checking is separate. | Useful for multilingual counterchecks, but provider accuracy claims do not prove authorship in an individual case. |
| Originality.ai | Limited free use · currently up to three short free scans daily; full access is paid | Confidence/classification values for likely human or AI text and highlighted sections, depending on model and plan. | A useful independent countercheck, designed primarily for publishing and content workflows. |
| Pangram | Free plan · limited daily checks; more features are paid | AI, human and AI-assisted assessments with interpretable section indicators. Plagiarism checking is an additional paid feature. | Interesting for mixed or revised text; the result remains a model estimate. |
| Grammarly AI Detector | Free web check · expanded insight in Pro | A percentage estimating how much scanned text appears AI-generated; Pro provides more detailed section guidance. | A quick second opinion. Do not confuse it with Grammarly's separate paid plagiarism checker. |
| QuillBot AI Detector | Free web check | AI score, sentence-level highlighting and categories such as AI-generated or AI-refined. | A low-friction comparison value, particularly when the identical version is used in other tools. |
| Scribbr AI Detector | Free · currently up to 1,200 words per check without sign-up | Human, AI-generated or AI-refined classification with paragraph-level feedback; supports several languages. | Useful for quick student checks. Scribbr's separate plagiarism checker measures something else. |
| Sapling AI Detector | Limited free use · currently 2,000 characters; longer text in Pro/Enterprise | Overall assessment and highlighted conspicuous regions based on the detection model. | Suitable only for short counterchecks; the character cap obstructs direct comparison of a complete IA or EE. |
| Winston AI | Trial access · limited credits for 14 days; paid thereafter | AI assessment, highlights and document reports; plagiarism and image detection are additional product functions. | Useful for file and report comparisons, but not the school's Turnitin benchmark. |
A reproducible detector comparison
- Identical version: enter exactly the same text in every tool without intermediate revision.
- Adequate length: do not infer a whole document from one sentence; follow each provider's minimum length.
- Record metadata: tool, full URL, date, language, word count, account type, settings and visible model version.
- Preserve raw results: record percentage, classification, highlighted passages and errors separately.
- Do not average into “truth”: range, median and disagreement can describe detector variation but cannot create a real AI proportion.
- Use controls: where possible, compare known human, known AI-generated and mixed texts of similar length and language.
- Discuss errors: examine false positives, false negatives and possible effects of translation, editing and short text.
What matters more for IB work than a detector value
Current IB and school rules on AI use are authoritative. Permitted support should be documented, while external or AI-generated material should be acknowledged and referenced transparently. Drafts, notes, source logs, data, calculations and version history also help demonstrate the student's actual process.
A detector value replaces neither source work nor a conversation about the real writing and research process.
Primary sources