Winston AI reviews writing for AI-generated patterns and copied passages, estimates whether images were generated by AI, and produces shareable reports. Its web app and developer API have different input and credit rules.
Used this tool? Rate it
Used this tool? Rate it
Winston AI is an AI detector for text, with a separate plagiarism checker. Its image workflow estimates whether an image was AI-generated and displays available forensic metadata. Individuals and teams can review documents in the web app; developers can connect text detection to other applications through an API.
Human Score estimates a document’s resemblance to verified human writing on a 0%–100% scale. Near 100% resembles human writing; near 0% resembles AI-generated writing. A middle result can reflect mixed authorship, heavy editing, translation or insufficient text. It is not the exact percentage of words generated by AI.
The AI Prediction Map divides writing into passages with five score bands:
| Score band | Displayed label |
|---|---|
| 0–20 | AI Generated |
| 20–40 | Likely AI |
| 40–60 | Uncertain |
| 60–80 | Mostly Human |
| 80–100 | Human Written |
Highlights indicate passages for review, not proof of authorship. Sentence scores are less reliable than the full-document score; the guidance says not to average them or treat one highlight as proof of AI use.
Explain summarizes scan patterns as a reading aid, not a separate detection result. Readability Score uses Flesch-Kincaid and a US school grade level to measure reading difficulty, not authorship.
The web app accepts pasted text and DOCX, PNG and JPG uploads. OCR extracts text, including handwriting, from scans and pictures; it does not determine their origin. Documents have title and author labels and views by type. Review includes sentence highlights, optional plagiarism checking and printable reports. Human Score does not establish that text is uncopied or correctly attributed.
AI image detection accepts an upload or image URL and estimates whether it is AI-generated. It can show available C2PA, IPTC and Exif data. Winston AI says the analysis can include clues about the generating tool. The result alone cannot verify a depicted person’s identity or establish that an event occurred.
HUMN-1 is the vendor’s human-content certification for websites and blogs. A site connects through a sitemap URL or WordPress plugin; its content is audited and a vendor badge is supplied.
Writing Feedback flags repetitive, overly formal or unclear wording and suggests grammar, syntax, sentence structure and tone changes. Writers choose which to apply.
Anyone receiving a shared report link may be able to view the report and original submission.
Editors can screen freelance articles, agency drafts, syndicated copy and guest posts for AI-likely passages, copied text and readability before assigning, revising or scheduling a draft.
Writing Feedback flags repetitive, overly formal or unclear AI-assisted drafts and suggests edits for writers to review.
Classrooms, departments and institutions can review student work with shareable reports. Educators also have a separate rubric-based essay grader. Enterprise and custom offers describe tailored detection, integrations, dedicated support and implementation assistance.
Alongside the web application, Winston AI lists a Chrome extension, Google Classroom, WordPress, Zapier and an API. The homepage feature list names English, French, Spanish, German, Portuguese, Dutch, Polish, Italian, Romanian, Indonesian, Tagalog, Russian, Bulgarian and simplified Chinese. Its FAQ instead lists 11 languages and omits Tagalog, Russian and Bulgarian.
The API landing page lists Tagalog and simplified Chinese but not Russian or Bulgarian. These list differences are not language-by-language performance results.
The developer area connects detection to other applications; the regular app is designed for daily individual use. It requires valid API credentials and advertises 2,500 introductory credits. The v2 endpoint uses Bearer authorization and JSON.
Its documented input rules distinguish three sources:
The v2 endpoint documents 23 language codes, including Tagalog, Russian, Bulgarian and simplified Chinese: nine more than the homepage feature list, without comparative performance data. It consumes one credit per processed word; web credits follow separate rules.
Pricing HTML captured November 12, 2026 contained Monthly and Annual labels and advertised 50% off with code WINSTON50. The table records that HTML; coupon application and the final checkout charge were not verified.
| Plan | Monthly billing, per month | Annual billing, monthly equivalent | Annual billing, promotional total | Monthly credits |
|---|---|---|---|---|
| Essential | Regular $18, promo $9 | Regular $10.00, promo $5.00 | $60.00 per year | 100,000 per month |
| Advanced | Regular $29, promo $14.50 | Regular $16.00, promo $8.00 | $96.00 per year | 200,000 per month |
| Elite | Regular $49, promo $24.50 | Regular $26.00, promo $13.00 | $156.00 per year | 500,000 per month |
The annual monthly equivalent spreads the yearly price across the year. A help table labels standard prices in USD, in plan order: $18/$29/$49 monthly and $120/$192/$312 annually, or $10/$16/$26 monthly equivalents. Earlier HTML captured at 13:05 UTC on November 2, 2026 also contained Monthly and Annual labels.
Free is $0 with 2,000 credits for a 14-day trial. The terms say the account is not charged and the trial subscription is suspended at expiry unless upgraded to a paid version. Essential targets individual text and image detection with writing feedback; Advanced also targets users needing advanced plagiarism checks, more credits, team access or website certification.
Web credit cost depends on content volume and enabled tools. Monthly credits refresh each billing cycle without rollover. Advanced lists up to five team members; Elite lists unlimited members and credit top-ups. Paid web scans list up to 200,000 characters each and unlimited scans per hour, subject to the plan’s credit allowance.
The terms add sales tax where required and accept USD, EUR, GBP, COP, BRL, IDR, INR, MXN and PHP. Prices may appear in USD. When conversion is needed, the payment provider sets the transaction-time rate; Winston AI disclaims responsibility for that provider’s conversion fees or exchange-rate differences.
Purchases are non-refundable subject to mandatory consumer rights, including Quebec’s Consumer Protection Act for Quebec residents. Cancellation through the account or support takes effect at the current paid term’s end. The pricing FAQ also says annual plans cannot be canceled during their term.
GPTZero presents Google Docs Writing Replay for a document’s writing process, pasted text and typing patterns, plus Canvas and Google Classroom integrations. Writing history differs from a detector score; these vendor features are not an accuracy ranking.
Originality.ai presents QA/QC team roles, optional training-data use, an opt-out and scan-history deletion. These controls establish neither better detection nor matching retention terms.
Web text scans require at least 500 characters; the help center recommends 300 words or more. API v2 has different input thresholds and a request maximum.
The help center rates social posts, bullet lists, translations, heavily edited AI drafts and verbatim transcripts as limited inputs. It calls source code, math formulas, boilerplate legal text, tables and structured data unsuitable; these provide less natural-prose signal.
Paraphrasing or humanization can raise the score of an AI-generated original. Despite training on humanized content, the vendor acknowledges reduced confidence after heavy rewriting. A higher score does not establish human authorship.
The homepage advertises 99.87% accuracy without a breakdown by language, format or editing history. This is a vendor claim. Its help center acknowledges false positives and negatives and treats results as one piece of evidence in high-stakes academic or employment decisions, alongside highlighted text, supporting materials and context.
An independent study published in June 2024 examined work by 153 students in an introductory microbiology course at UW–Madison. The students wrote about 500 words under supervision and also supplied AI and modified AI answers. Survey sessions ran from November 29 to December 14, 2023; those dates are not stated as the detector test dates. The researchers converted Winston’s human score to an AI score by subtracting it from 100, and counted a human-written answer as a false positive when the result was greater than 50% AI. At that threshold the paper reports a 45.8% false-positive rate for Winston on human writing. A Human Score of exactly 50% is not above that threshold. The paper does not specify the Winston model version or exact detector test dates, and this finding does not establish a current general error rate.
A separate residency-application study published in July 2025 analyzed 25 selected texts of about 700 words each in November 2023. They included five AI-generated statements, five pre-ChatGPT applications, five applications from the 2023–2024 cohort and two groups of five literary excerpts. Both applicant groups were convenience samples. Winston AI gave the AI-generated controls 0% human scores and the pre-ChatGPT application controls 100% human scores. For the recent applications, its human scores ranged from 3% to 100%. Researchers could not confirm whether those applicants had used ChatGPT, so that range cannot establish a false-positive rate. The results describe those historical tests, not the current web detector, API v2 model 5.2 or other languages and settings.
Text and images remain stored while their dashboard report exists. The policy gives up to 30 days for active-storage deletion after report deletion, or for all associated content and reports after account deletion. Encrypted backups may retain content up to 90 days after active deletion, with a legal-preservation exception. The backup period starts after active deletion, not upload.
Account data remains for the account’s duration plus up to six months from an idle period’s start, or until a deletion request, whichever comes first. Logs remain up to 12 months before deletion or anonymization; backups may be isolated pending deletion.
The homepage says submitted content is never used to improve the company’s models or capabilities. The privacy policy permits aggregated, de-identified usage analytics to improve analysis accuracy and performance. It separately requires the submitter’s explicit opt-in to build or train AI models on submissions and says submitted text and images are not sold, licensed or otherwise commercially exploited.
For data received through Google Workspace APIs, use is limited to user-requested functions and excludes generalized model training, advertising and sale. Third-party transfers are limited to what is necessary to provide and secure the service, or comply with applicable law.
The policy permits disclosure to contracted providers, during business transfers, for valid legal requests and to protect rights or investigate suspected violations. It describes encryption at rest and in transit and need-to-know access. The terms say the site is hosted in Canada. The policy says no internet transmission can be guaranteed 100% secure. Submissions and reports are private by default, with optional report-link sharing.
Users retain ownership of submitted content and its intellectual-property rights. Processing follows the privacy policy and account choices. The terms require necessary rights, licenses and permissions, including written consent, release or permission from each identifiable person depicted for processing their name or likeness where law requires it. Child sexual abuse material (CSAM) is prohibited, including purported detection or research submissions. Other submissions must comply with the terms and applicable law.
The terms contain a narrow exception to the restriction on submitting non-consensual intimate imagery (NCII), solely for detecting AI generation or other manipulation. They prohibit using the site as a platform to store, host, distribute or disseminate NCII. The privacy policy separately prohibits uploading NCII without consent from the people depicted.
The terms require users to be at least 18. The privacy notice says users represent either being at least 18 or being a parent or guardian consenting to a minor dependent’s use. Automated or non-human site access is restricted except as expressly permitted through the API. A separate clause excepts activity that may result from standard search engine or Internet browser use.
API access is subject to rate limits and fair-use policies. Reselling, sublicensing or redistributing it requires express prior written permission. The 99.9% uptime target has downtime and liability qualifications.
No. The score estimates resemblance to human writing; copied text and attribution need separate analysis.
The web trial lists 2,000 credits for 14 days. The developer area advertises 2,500 introductory credits and requires valid API credentials.
No. Active-storage deletion, encrypted backups and legal preservation have separate timelines from removing a report from the dashboard.
