Test log
Winston AI test data
Raw results for the Winston AI review. The verdict is in the review.
- Tested
- 6 November 2026
- Tester
- Miles Foster
- Account
- Free demo account, signed up by RankDude, nothing paid to the brand
- Tests
- 8
- Screenshots
- 7
- Method
- Published methodology
- Environment
- Google Chrome on desktop
Results
| ID | Test | Key measurement | Score |
|---|---|---|---|
| T01 | Accuracy on AI-Written Text | AI passages caught: 5 of 5 | 5/5 |
| T02 | False Positives on Human Text | Human passages flagged as AI: 0 of 5 | 5/5 |
| T03 | Mixed & Lightly-Edited Text | Mixed passages with the AI part highlighted: 2 of 2 | 4/5 |
| T04 | Non-English Text | Passages correct: 4 of 4 | 5/5 |
| T05 | Document & File Support | Ways in that worked: 6 of 10 | 3/5 |
| T06 | Score Clarity & Evidence | Points met: 9 of 10 | 4/5 |
| T07 | Scan Speed | Median per scan: 3.8 s | 5/5 |
| T08 | Pricing & Free-Tier Limits | Trial words in total: 2,000 | 3/5 |
| Overall (in the review) | 4.9/5 | ||
Accuracy on AI-Written Text
5/5Tested- Method
- Five AI-written passages of 78 to 91 words (remote work, the Industrial Revolution, sleep, a cover letter, electric cars) were pasted once each; the first verdict counted, with 50 percent human as the line.
- AI passages caught
- 5 of 5
- Lowest AI score
- 0% human
- Remark
- Scale 5: 47 to 50 of 50 caught. 5 of 5 is 100 percent, so the grade is 5; the sample is five passages, not fifty.
Raw notes
Five AI-written passages of 78 to 91 words (remote work, the Industrial Revolution, sleep, a cover letter, electric cars) were pasted once each; the first verdict counted, with 50 percent human as the line.
All five scored 0% human, with every sentence highlighted as likely AI and the verdict "We are highly confident that this material was produced using an AI language model."
A sixth and seventh AI passage in Spanish and German also scored 0% human (see the non-English test).
- Five passages from one model (Claude), at default settings, not 50 from six models.
False Positives on Human Text
5/5Tested- Method
- Five passages from books published before 1900 (Pride and Prejudice, Walden, On the Origin of Species, Moby-Dick and the Autobiography of Benjamin Franklin) were pasted once each, so any AI flag would be a false positive.
- Human passages flagged as AI
- 0 of 5
- Lowest human score
- 98%
- Remark
- Scale 5: at most 1 of 50 flagged. 0 of 5 flagged is 0 percent.
Raw notes
Five passages from books published before 1900 (Pride and Prejudice, Walden, On the Origin of Species, Moby-Dick and the Autobiography of Benjamin Franklin) were pasted once each, so any AI flag would be a false positive.
Scores were 100, 100, 100, 98 and 100 percent human. No sentence was highlighted as AI.
The two non-English human passages (Don Quijote, Les Miserables) scored 98 and 100 percent human.
- Old literary text is far from the student and web writing that detectors wrongly accuse; a clean result here is a minimum. Five passages, not fifty.
Mixed & Lightly-Edited Text
4/5Tested- Method
- Two passages each joined a human paragraph (from Great Expectations and A Tale of Two Cities) to an AI paragraph, in both orders, and were scanned once.
- Mixed passages with the AI part highlighted
- 2 of 2
- Scores
- 67% and 64% human
- Remark
- Counted as flagged because the AI sentences were marked. Scale 5 is 17 of 20 with accurate highlights; the headline score leaning human costs one grade.
Raw notes
Two passages each joined a human paragraph (from Great Expectations and A Tale of Two Cities) to an AI paragraph, in both orders, and were scanned once.
Scores were 67 and 64 percent human. The highlights landed on the AI sentences in both: the AI half red, the human half green, and the bars read 50 percent likely AI and 50 percent human.
The verdict sentence said human authorship was more probable with some AI-like traits, so by the 50 percent line alone both would pass as human.
- Two passages, not twenty. The AI-draft-with-human-edits kind was not run.
Non-English Text
5/5Tested- Method
- Four passages were scanned once: an AI passage in Spanish, an AI passage in German, a human Spanish passage (Don Quijote) and a human French passage (Les Miserables, imported as a text file).
- Passages correct
- 4 of 4
- Human passages flagged
- 0
- Remark
- Scale 5: 22 to 24 of 24 correct with at most one human passage flagged. 4 of 4 is 100 percent.
Raw notes
Four passages were scanned once: an AI passage in Spanish, an AI passage in German, a human Spanish passage (Don Quijote) and a human French passage (Les Miserables, imported as a text file).
Both AI passages scored 0% human. The human Spanish passage scored 98% and the human French passage 100%.
No human passage was flagged.
- Four passages, three languages. Chinese was not run.
Document & File Support
3/5Tested- Method
- Each way in was tried on the trial: paste, a text file, a PDF, a Word file, a batch of files and a web address, then the cap, extension, API and export points were checked.
- Ways in that worked
- 6 of 10
- Remark
- Scale 3: 6 or 7 of 10 points. Import was tested without scanning where scanning would spend credits.
Raw notes
Each way in was tried on the trial: paste, a text file, a PDF, a Word file, a batch of files and a web address, then the cap, extension, API and export points were checked.
Worked: paste; a .txt file (scanned, 106 words); a PDF (98 words imported); a Word file (106 words imported); the batch box (files listed as Ready with word counts, up to 50 files); a web address (3,218 words imported from a web page). That is 6 of 10.
Not counted: a 2,000-word scan (the trial could not pay for one); the Chrome extension (listed, not installed); the API (it sits on a separate developer platform with its own account); export (Download is locked behind a paid plan).
- The batch and the PDF and Word imports were not scanned.
Score Clarity & Evidence
4/5Tested- Method
- Each of the ten points was checked on the scans already run, with one passage scanned twice for repeatability.
- Points met
- 9 of 10
- Remark
- Scale 4: 8 or 9 of 10 points.
Raw notes
Each of the ten points was checked on the scans already run, with one passage scanned twice for repeatability.
Met: a plain verdict sentence; a percentage; sentence highlighting; a readable explanation with a how-to-read link; a share link (a Link sharing switch on each scan); a document history; repeatability (the same passage scored 0% human twice with the same sentence); a stated limit (it says the sentence map is less accurate than the overall score); no upsell blocking the result. That is 9 of 10.
Missed: a stated margin or confidence band. The verdict says "highly confident" but gives no range.
- Repeatability used two scans, not three, to stay inside the credit allowance. Explain and Download are locked on the trial.
Scan Speed
5/5Tested- Method
- All 17 scans were timed from pressing Scan to the verdict sentence on screen, checking ten times a second.
- Median per scan
- 3.8 s
- Slowest scan
- 4.7 s
- Remark
- Scale 5: a median under 5 seconds.
Raw notes
All 17 scans were timed from pressing Scan to the verdict sentence on screen, checking ten times a second.
Median 3.8 seconds, fastest 3.1, slowest 4.7. No scan failed or queued.
- Passages of 75 to 125 words; a long document takes longer. Free-trial speed only.
Pricing & Free-Tier Limits
3/5Tested- Method
- The trial was used until I reached 82 percent of the allowance; the plans page in the app and the public pricing page were read on the test date.
- Trial words in total
- 2,000
- Cost of 1,000 words on Essential
- about $0.18
- Remark
- Scale 3: a tight free tier. The paid price is in the $10 to $19.99 band.
Raw notes
The trial was used until I reached 82 percent of the allowance; the plans page in the app and the public pricing page were read on the test date.
Trial: 2,000 credits for 14 days, one credit a word, so 2,000 words once, not a daily allowance. It needed a Google sign-in and no card.
The cheapest paid plan, Essential AI, is $9 a month, or $5 a month billed yearly, for 100,000 words a month. That is about $0.18 for 1,000 words monthly and $0.10 billed yearly.
Advanced is $14.50 and Elite $24.50 a month. A WINSTON50 code showed 50 percent off at checkout; I did not use it.
- No plan was bought; checkout was not opened.
Coverage
Tested
- 17 scans: 5 AI passages, 5 human passages from books published before 1900, 2 mixed passages, 4 non-English passages and one repeat
- Time from pressing Scan to the verdict, for all 17 scans
- Imports of a text file, a PDF, a Word file and a web page, and the batch box
- Share, Download and Explain buttons; document history
- The plans page and the public pricing page read on the test date
- The Trustpilot page read in the same browser
Not tested or limited
- The trial's 2,000 credits paid for 17 short passages of 75 to 125 words, not 50 + 50 + 20 + 20 + 24 full passages
- Resistance to humanisers and AI model coverage were not run: no humaniser account was bought and every AI passage was written by one model
- Every human passage is from an old book, which is easier to tell from AI than a modern essay
- Chinese was left out of the non-English test to save credits
- A batch of files was imported and listed but not scanned; a per-scan cap of 2,000 words was not tested because the trial could not pay for one
- No plan was bought: Download, Explain, plagiarism and team features were seen locked, not run
Still to test
- The other 45 AI and 45 human passages on a paid plan
- Resistance to humanisers and paraphrasers
- AI model coverage across six models
- A full-length essay and a 2,000-word scan
- A batch of files scanned end to end