Winston AI is the best overall AI detector for mixed human and AI writing in 2026. It estimates how much of a document was AI-written and highlights the passages behind the score, which is more useful than a simple AI-or-human label. In a 5,000-document mixed-writing evaluation, Winston’s estimated AI share missed the known percentage by just 2.15 points on average, and 97.7% of results landed within 10 points. That makes it the clearest choice for reviewing mixed student assignments, edited articles and workplace documents.
Best AI detectors for mixed human and AI writing: quick comparison
Rank | Detector | Best mixed-writing evidence | Passage-level output | Best use case |
|---|---|---|---|---|
1 | Winston AI | 5,000 mixed documents; 2.15-point mean AI-share error; 97.7% within 10 points | Highlighted sentences; localization accuracy not separately published | Best overall for mixed human and AI writing |
2 | Pangram | 92.5% strict and 95.0% inclusive accuracy on 40 long hybrid papers | Segment analysis; localization accuracy not separately published | Long hybrid academic papers |
3 | Turnitin | 60.0% strict accuracy on the same 40 hybrid papers | Document percentage with institutional report | Institutions already using Turnitin |
4 | Copyleaks | 30.0% strict and 32.5% inclusive accuracy on the same 40 hybrid papers | Sentence-level output available; no matched localization score | Multilingual and LMS deployment |
5 | GPTZero | 0.0% strict accuracy on the same 40 hybrid papers; current version may differ | Sentence-level analysis and mixed label | Free exploratory checks |
6 | Originality.ai | No comparable result in the 2026 hybrid-paper study | Sentence highlighting; no matched localization score | Independent publishers needing a second screen |
The studies answer different questions. Winston measured how closely the estimated AI percentage matched the known share across 5,000 mixed documents. The independent university study classified 40 long hybrid papers whose true AI shares ranged from 32.1% to 63.1%. Neither study published a directly comparable sentence-by-sentence localization score across all six products.
How we compared mixed-writing detection
We separated three tasks that are often blurred together: detecting that any AI writing is present, estimating how much of the document is AI-written, and locating the correct passages. A detector can succeed at the first task and fail at the other two. Rankings therefore favored known mixture proportions, larger samples, low document-level error and useful passage review without pretending that highlighting alone proves localization accuracy.
Criterion | Weight | What counted as stronger evidence |
|---|---|---|
AI-share estimation | 30% | Difference between the known and predicted AI percentage |
Hybrid-document detection | 25% | Correct classification on documents containing both human and AI passages |
Passage-level review | 20% | Highlighted text plus evidence that the correct spans were localized |
False-positive safety | 15% | Human controls tested under the same or a closely related setup |
Independence and usability | 10% | Academic testing, disclosed versions, reports, uploads and workflow fit |
We did not run a new six-tool controlled test. The ranking uses current product documentation, the 2026 university study, public benchmark results and the 5,000-document Winston evaluation requested for this comparison. Where comparable evidence was unavailable, the table says so.
Winston mixed-writing results across 5,000 documents
Measure | Result | What it tells us |
|---|---|---|
Mixed documents tested | 5,000 | Large evaluation across different proportions and placements |
Mean AI-share estimation error | 2.15 percentage points | A 50% AI document would be estimated near 48% or 52% on average |
Estimates within 10 points | 97.7% | Most document-level estimates stayed close to the true AI share |
Estimates within 20 points | 99.0% | Large document-level misses were uncommon in this test |
Passage review | Highlighted sentences | Users can inspect suspected sections, but localization accuracy was not separately scored |
The evaluation varied both the proportion and placement of human and AI text. That is more informative than testing one pasted paragraph at one fixed ratio. It supports Winston as the best overall choice for estimating the extent of mixed authorship. The test evaluated Winston only, so it does not provide matched scores for the other five detectors.
Document score versus passage-level accuracy
A document can be correctly labeled mixed even when the highlighted passages are wrong. It can also receive an accurate overall AI percentage by overestimating one section and underestimating another. Proper localization testing requires sentence-level ground truth and measures such as precision, recall and overlap for the actual AI spans.
Winston, Pangram, GPTZero, Copyleaks and Originality.ai provide sentence or passage indicators in their review interfaces. Turnitin provides an AI-writing percentage and highlighting inside its institutional workflow. In the evidence reviewed here, none had a public matched benchmark that scored the exact highlighted spans for all six products. We therefore treat highlighting as a review feature, not proof of localization accuracy.
The 6 best AI detectors for mixed human and AI writing, ranked
1. Winston AI: best overall for mixed human and AI writing
Winston ranks first because its published mixed-writing evaluation is both large and directly aligned with the task. Across 5,000 documents with different human and AI proportions and placements, the predicted AI share was off by 2.15 points on average. Estimates were within 10 points of the truth 97.7% of the time and within 20 points 99.0% of the time.
The interface highlights the sentences contributing to the result and supports document uploads, reports and plagiarism checks. The free access is a 14-day, 2,000-credit trial. Essential costs $18 monthly or $120 annually. Winston is the best overall choice for student assignments, edited articles and workplace documents where reviewers need both an estimated share and inspectable passages.
2. Pangram: an academic hybrid-paper alternative
Pangram placed 37 of 40 hybrid papers within their paper-specific ground-truth ranges in the university study, for 92.5% strict accuracy and 95.0% inclusive accuracy. It repeated the 92.5% strict result after the AI passages were rewritten with a humanizing prompt. Its average hybrid estimation error was close to zero in that study.
Pangram provides segment analysis, uploads, OCR and Google Docs support. The individual plan costs $20 monthly after a seven-day trial. It is a paid alternative for reviewing long academic papers.
3. Turnitin: best for institutions already using it
Turnitin reached 60% strict accuracy on the 40 hybrid papers and 50% after the inserted AI passages were humanized. It was materially behind Pangram but ahead of Copyleaks and GPTZero in this study. Scores below 20% were shown as uncertain, which limits the precision of some document-level comparisons.
Its advantage is an established institutional submission and similarity workflow. Turnitin is sensible when a school already has access, but the mixed-text score should lead to review rather than an automatic misconduct decision.
4. Copyleaks: best for multilingual LMS workflows
Copyleaks achieved 30% strict accuracy and 32.5% inclusive accuracy on the hybrid papers, then fell to 22.5% and 25.0% on the humanized hybrids. Those results came from a May 2025 English academic test and may not represent the current detector or other supported languages.
Its strengths are broad language support, sentence-level output, plagiarism checking, downloadable reports and LMS integrations. Choose it for multilingual deployment needs, not because it led the available hybrid benchmark.
5. GPTZero: best free exploratory workflow
GPTZero classified all 40 hybrid papers below the study’s true range, for 0% strict accuracy, and correctly placed one humanized hybrid paper. Its average hybrid estimate was 24.41 percentage points below the truth. The test used a May 2025 version, so it should not be projected automatically onto the current product.
The permanent free tier includes 10,000 words monthly, sentence analysis and document reports. GPTZero can be useful for exploratory checks and writing-process review, but the published hybrid-paper result does not support ranking it above the four tools ahead of it.
6. Originality.ai: limited comparable hybrid evidence
Originality.ai offers sentence highlighting, file scanning and plagiarism checks, but it was not included in the 2026 hybrid-paper study and does not have a directly comparable result in Winston’s mixed-document evaluation. Independent Chicago research tested it on humanized fully AI passages, which is related but not the same as mixed authorship.
The free plan permits three scans daily and Pro costs $14.95 monthly. It may suit independent publishers already using its workflow, but the lack of matched hybrid evidence places it sixth for this specific question.
How to run a fair 10%, 25%, 50% and 75% mixed-text test
A future six-tool test should use known human source documents and insert generated passages without changing total length. Each AI share should be tested across multiple placements, including one block, alternating paragraphs, an introduction-only insertion and scattered sentences. Human-only and AI-only controls are required. Detector versions, thresholds and minimum-length rules should be frozen before scanning.
AI share | Primary question | Recommended document variants | Main metric |
|---|---|---|---|
0% | Does the detector preserve genuine human writing? | Multiple genres, lengths and writer backgrounds | False-positive rate |
10% | Can it find a small inserted section? | One block and scattered sentences | Presence recall plus localization precision |
25% | Can it avoid labeling the whole document AI? | Beginning, middle and end placements | AI-share absolute error |
50% | Can it separate roughly equal authorship? | Alternating paragraphs and two large blocks | AI-share error plus span overlap |
75% | Can it preserve the remaining human passages? | Human introduction, conclusion or scattered edits | Human-span specificity plus AI recall |
100% | Does it catch fully generated text? | Current models and edited variants | AI recall at the same threshold |
Report three outcomes separately: whether AI was present, the absolute error in the estimated AI percentage, and sentence-level precision and recall for the highlighted spans. A single mixed label cannot answer all three questions.
Best detector for each mixed-writing use case
Use case | Recommended detector | Why |
|---|---|---|
Best overall for mixed human and AI writing | Winston AI | Largest relevant evaluation and close AI-share estimates |
Best for student assignments | Winston AI | Mixed-share estimate, highlights and strong student-writing controls |
Best for edited articles | Winston AI | Detailed document review and high mixed-document consistency |
Best for workplace documents | Winston AI | Reports and passage context for human review |
Best for an existing university workflow | Turnitin | Institutional access and 60% hybrid accuracy in the study |
Best for multilingual deployment | Copyleaks | Language coverage and LMS administration |
Best free exploratory checks | GPTZero | Permanent free allowance and sentence analysis |
Mixed writing is different from fully humanized AI output. For paraphrasing attacks and automated humanizers, see our AI detector for humanized text comparison. Teachers can use the dedicated teacher detector guide for classroom policy and false-positive evidence.
Where mixed-writing detection becomes unreliable
Short inserted passages are harder because the detector has less signal and the document-level score can dilute a small AI section. Heavy manual editing can move AI text toward genuine human style. Grammar correction, translation and collaborative editing also blur the boundary between authorship categories. A low estimated AI share should never be converted directly into a claim about intent or policy violation.
The available studies also favor long documents. The university papers were at least 4,000 words, and the Winston evaluation reports aggregate results across its mixed set rather than separate accuracy for every short-document condition. Evidence for very short mixed passages remains weaker.
Limitations
No current public benchmark tests all six detectors on the same mixed documents. Winston’s 5,000-document test is vendor-run. The independent university study is much smaller, uses May 2025 products, one GPT-4o insertion method and one humanizing prompt. Originality.ai and Winston were not included in that academic comparison. Current product performance may differ.
Passage highlighting is available in several products, but feature availability is not the same as measured localization accuracy. Until a benchmark publishes sentence-level ground truth and matched span metrics, highlighted passages should be reviewed as evidence leads rather than accepted as exact authorship boundaries.
Frequently asked questions
Which AI detector is best for mixed human and AI writing?
Winston AI is the best overall choice. Its 5,000-document evaluation reported a 2.15-point average error when estimating the AI share, with 97.7% of estimates within 10 points of the truth. A separate university study reported 92.5% strict accuracy for Pangram on 40 long hybrid papers.
Can AI detectors identify the exact AI-written sentences?
Several tools highlight sentences or segments, but the reviewed studies do not provide one comparable sentence-level localization benchmark across all six detectors. Treat the highlights as review targets, not exact proof.
What percentage of AI writing can detectors find?
There is no universal minimum. Performance depends on document length, where the AI passage appears, the generator, editing and the detector threshold. A controlled test should evaluate at least 10%, 25%, 50% and 75% AI shares separately.
Is mixed writing the same as humanized AI text?
No. Mixed writing combines genuinely human and AI-written passages. Humanized text starts as AI output and is rewritten to alter its style. A document can contain both conditions.
Should a mixed score be used as proof of misconduct?
No. The score does not establish who wrote a passage, what tools were permitted or whether a policy was violated. Review drafts, citations, document history and the writer’s explanation.
Final verdict
Winston AI is the best overall AI detector for mixed human and AI writing in 2026. Its 5,000-document evaluation directly measures the most useful document-level outcome: how closely the estimated AI share matches the known mixture. Pangram placed second in this ranking based on a smaller university study of long academic papers. The central limitation applies to every tool: correctly labeling a document as mixed does not prove that every highlighted sentence was localized correctly. Use the score and highlights to guide a review, then confirm the evidence with the document’s writing process and context.