Introducing Winston 5.0: A Breakthrough in AI Detection Accuracy

Winston AI is now the most accurate AI detector across leading standardized public benchmarks

The release of Winston 5.0 marks a major step forward in AI content detection. In a held-out evaluation conducted in-house on 12,000 English documents, Winston 5.0 achieved 99.87% overall accuracy, correctly classifying 11,984 samples with just 16 errors.

We have also evaluated Winston 5.0 on established public benchmarks. Results for PERSUADE 2.0, the DetectionAI benchmark from researchers at the University of Chicago Booth School of Business, and DetectRL-X are included below. Together, these benchmarks test Winston across student essays, several writing domains, and AI output from multiple current models.

For a direct comparison with other leading AI detectors, see how Winston 5.0 performed against GPTZero, Pangram, Grammarly, and QuillBot across standardized public benchmarks, including RAID, Chicago Booth, and PERSUADE 2.0.

Key results

• Winston correctly recognized 99.88% of human writing, with only seven false positives among 6,000 documents, or roughly one false flag for every 857 human documents.
• Winston identified 99.85% of AI-generated writing, correctly detecting 5,991 of 6,000 documents produced by 20 current AI models.
• Twelve of the 20 AI generators were detected with 100% accuracy. Even the lowest individual result was 99.36%.
• On the standardized public PERSUADE 2.0 benchmark, Winston correctly recognized all 25,996 student essays as human, with zero false positives.
• On the University of Chicago Booth DetectionAI benchmark, Winston correctly classified all 8,697 documents, with no false positives or false negatives.
• On DetectRL-X, Winston detected 15,598 of 15,599 AI documents (99.99%), recognized 97.08% of human documents, and achieved 98.54% overall accuracy across 31,196 texts.

These results cover Winston 5.0’s English-language model. The multilingual model has not yet been released and will be added shortly.

Performance on standardized public benchmarks

Our 12,000-document evaluation shows how Winston 5.0 performs on a balanced test set. Public benchmarks add another useful check: they show whether the same performance carries over to datasets created for academic research and available for others to inspect.

We have completed evaluations on PERSUADE 2.0, DetectionAI from the University of Chicago Booth School of Business, and the English portion of DetectRL-X.

*PERSUADE 2.0 contains human writing only, so it measures false positives rather than balanced overall accuracy or AI detection.

PERSUADE 2.0: zero false positives across 25,996 student essays

PERSUADE 2.0 is a public corpus of argumentative essays written by U.S. students in grades 6–12. Winston classified every essay as human. There were no false positives across all 25,996 documents.

The scores were not close to the 0.5 classification threshold. The average AI score was 0.0027, and the 99th-percentile score was 0.0170. Winston also achieved 100% human accuracy in every demographic group included with the corpus, covering grade level, gender, race and ethnicity, economic disadvantage, and disability status.

University of Chicago Booth benchmark: 100% accuracy across 8,697 documents

The DetectionAI benchmark was created by researchers Alex Imas and Brian Jabarian at the University of Chicago Booth School of Business. It contains human and AI writing drawn from six types of content: novels, news, résumés, blogs, Amazon reviews, and restaurant reviews. The evaluation excluded 1371 documents that fell below Winston’s 500-character minimum, leaving 8,697 texts for scoring.

Winston classified every document correctly. All 1,747 human texts were recognized as human, and all 6,950 AI texts were detected as AI. That produced 100% accuracy, 100% AI recall, and 100% human specificity, with no false positives or false negatives.

The scores were not clustered near Winston’s 0.5 classification threshold. The highest AI score assigned to any human document was 0.2158. The lowest score assigned to an AI document was 0.8633. Even the two closest cases therefore remained on the correct side of the threshold, with a sizeable gap between them.

The 6,950 AI documents covered four models from Anthropic, Google, and OpenAI. Winston recorded 100% detection for each one:

Accuracy was also 100% in each of the six content domains. The tested set included 5,000 novel passages, 1,427 news texts, 968 résumés, 579 blog posts, 436 Amazon reviews, and 287 restaurant reviews. This makes the result broader than a test built around one writing style or one AI provider.

DetectRL-X: 99.99% of AI writing detected

DetectRL-X is a public independent benchmark built from paired human and AI texts. The source file contained 31,200 English texts. Four fell below Winston’s 500-character minimum and were excluded before scoring, leaving 31,196 documents across academic writing, novels, news, SEO content, general web text, and Wikipedia-style writing. The AI texts came from DeepSeek V3, GPT-4o, Qwen Max, and Gemini 2.5 Flash.

Winston achieved 98.54% overall accuracy. It detected 15,598 of the 15,599 AI documents, missing only one. On the human side, it correctly recognized 15,142 of 15,597 documents.

Accuracy exceeded 99% in four of the six domains, and AI recall across the full set was 99.99%. Results varied more on human web and SEO texts, which gives us a specific area to work on.

See the difference across real-world writing

We tested more than essays. The evaluation covered ten kinds of human writing and output from 20 current AI models.

Human writing stayed human

The 6,000 human documents included college essays, books, news articles, political speeches, movie reviews, Reddit posts, Wikipedia entries, poetry, recipes, and medical writing. Winston correctly classified every document in seven of those ten categories.

Only seven human documents were incorrectly flagged across the full set. Three came from the medical category, two were poems, and two were recipes. All 600 college essays were correctly recognized as human, as were all 600 news texts, 600 Reddit posts, 600 Wikipedia entries, 600 movie reviews, 600 political texts, and 600 book excerpts.

Testing several kinds of human writing reduces the chance that one easy category is driving the overall result and we received a 0% false positive rate on the most common writing categories in English.

AI writing from current models was consistently detected

The AI portion contained 6,000 documents produced by 20 generators from OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, and Qwen. Winston correctly detected 5,991 of them.

Twelve of the 20 generators had a perfect detection result: Claude Sonnet 4.6, Claude Sonnet 5, Gemini 3.5 Flash, Gemini 3.5 Flash-Lite, GPT-4o, GPT-5.6 Luna, GPT-5.6 Terra, Mistral Nemo, Qwen 3.7 Flash, Qwen 3.7 Plus, Qwen 3.8 27B, and Qwen 3.8 Max. Even the lowest result, for Grok 4.5, was 99.36%, with 312 of 314 documents correctly detected.

Longer documents produced even stronger results

Every document in the evaluation contained at least 500 characters. Across the complete set, accuracy was 99.87%. That increased to 99.98% among the 5,369 documents with at least 2,000 characters.

Winston correctly classified all 3,358 documents containing at least 3,000 characters.

The result is consistent across very different kinds of writing: Winston identifies current AI output while leaving genuine human work overwhelmingly untouched.

Why accuracy and false positives matter in the classroom

For a teacher, an incorrect result creates work immediately.

When a genuine essay is flagged as AI, the teacher has to pause grading, review earlier assignments, check the document history, speak with the student, and sometimes involve parents or administrators.

The student may also feel that their work and honesty have been questioned before any conversation begins.

AI-generated writing that goes undetected creates a different problem. If a generated assignment is classified as human, a teacher may grade and provide detailed feedback on work that does not reflect the student’s understanding. Over time, that makes it harder to know who has developed the required skills and who needs additional support.

The Winston 5.0 evaluation produced seven false positives among 6,000 human documents and nine false negatives among 6,000 AI documents. Those results matter together. A detector needs to identify AI-generated work while remaining highly cautious with authentic human writing.

False positives create extra investigations for teachers, while missed AI writing makes the detector less useful. Both errors were rare in this evaluation, allowing teachers to focus on the small number of submissions that deserve a closer look.

How we tested Winston 5.0

Our internal evaluation used a held-out dataset of 12,000 English documents: 6,000 written by people and 6,000 generated by AI.. Winston has a minimum input requirement of 500 characters, matching the limit used in the application. Shorter passages provide too little linguistic context for a reliable document-level assessment, so this requirement was applied before scoring. The test therefore focused on document-level writing rather than isolated sentences or short answers.

The human set contained 600 documents from each of ten categories: books, college essays, medical writing, movie reviews, news, poetry, political speeches, recipes, Reddit posts, and Wikipedia articles. Using the same number from each category prevented a large source from dominating the result.

The AI set covered 20 generators from OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, and Qwen. Sample sizes ranged from 140 to 330 documents per generator, with 15 of the 20 generators contributing between 310 and 330. This spread allowed us to measure whether performance remained stable across different model families rather than depending on the writing patterns of a single chatbot.

How each document was scored

Winston produced a human score and an AI score for every document. The class with the higher score became the predicted result. A document with an AI score above 0.5 was classified as AI, while a score below 0.5 was classified as human.

The score summary below shows how sharply the results separated. Human documents had a median AI score of 0.0016, and 99% scored below 0.0503. AI documents had a mean AI score of 0.9972 and a median of 1.0.

How classification errors were counted

Each result falls into one of four groups:

• Correctly identified human: Human writing classified as human.
• False positive: Human writing incorrectly classified as AI.
• False negative: AI-generated writing incorrectly classified as human.
• Correctly identified AI: AI-generated writing classified as AI.

The confusion matrix presents all four outcomes together. Reading across each row shows how the documents in that known class were categorized by Winston.

Mixed/Hybrid writing: measuring AI within human documents

Writing is increasingly hybrid. A person may draft most of a document, use AI for an introduction or conclusion, or move between human- and AI-written passages. In these cases, a simple human-or-AI label does not describe the document very well. The more useful question is how much of the text appears to be AI-generated.

Hybrid writing is especially relevant in education, where acceptable AI use can vary by assignment. A student might use AI to help with an introduction, rewrite one difficult paragraph, or complete a conclusion while writing the rest themselves. A document-level human-or-AI label cannot capture those differences.

We tested Winston 5.0 on 5,000 English documents containing both human and AI writing. The set covered seven mixing patterns, including human text followed by AI text, AI text followed by human text, alternating sections, and AI-written introductions, middle paragraphs, and conclusions.

Across 5,000 mixed documents, 99.0% of Winston’s estimates were within 20 percentage points of the true AI-written share. The average difference was 2.15 percentage points.

To put that into context, if a document were 50% AI-written, a difference of 2.15 points would mean an estimate of roughly 48% or 52%. Individual results varied, but 4,885 estimates (97.7%) were within 10 percentage points of the true mix.

For mixed documents, Winston estimates the overall human-written share and highlights individual sentences that appear more AI-generated or human-written.

In the example above, the two AI-written opening sentences were marked as likely AI and the four human-written sentences were highlighted as human. The sentence view gives the reviewer specific passages to examine alongside the student’s earlier work, the assignment rules, and the document history.

The evaluation placed AI-written material in different parts of the document. Across the seven mixing patterns, the share of estimates within 20 percentage points of the true AI-written share ranged from 96.3% to 99.8%. That range covered AI appearing before or after the human writing, between human passages, or in an introduction, middle paragraph, or conclusion. The result did not depend on generated text appearing in one predictable place.

Final verdict

Winston 5.0 correctly classified 11,984 of 12,000 documents, including 5,993 of 6,000 human texts and 5,991 of 6,000 AI texts. It also recorded zero false positives across all 25,996 PERSUADE 2.0 essays and correctly classified every document in our 8,697-text Chicago Booth evaluation.

AI writing will keep changing, so we will keep testing Winston against new models and new methods of rewriting generated text. Based on this evaluation and the standardized public benchmarks, Winston 5.0 is the most accurate AI detector we have tested.