Test log
ZeroGPT test data
Raw results for the ZeroGPT 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
- 10
- Screenshots
- 9
- Method
- Published methodology
- Environment
- Google Chrome on desktop
Results
| ID | Test | Key measurement | Score |
|---|---|---|---|
| T01 | Resistance to Humanisers & Paraphrasers | – | n/a |
| T02 | AI Model Coverage | – | n/a |
| T03 | Accuracy on AI-Written Text | AI passages caught: 20 of 20 | 5/5 |
| T04 | False Positives on Human Text | Human passages flagged: 2 of 20 | 3/5 |
| T05 | Mixed & Lightly-Edited Text | Mixed passages flagged: 9 of 10 | 4/5 |
| T06 | Non-English Text | Correct labels: 20 of 20 | 5/5 |
| T07 | Document & File Support | Points that worked: 5 of 10 | 2/5 |
| T08 | Score Clarity & Evidence | Points that passed: 6 of 10 | 3/5 |
| T09 | Scan Speed | Median scan time: 1.2 seconds | 5/5 |
| T10 | Pricing & Free-Tier Limits | Free cap per scan: 15,000 characters | 5/5 |
| Overall (in the review) | 4.6/5 | ||
Resistance to Humanisers & Paraphrasers
n/aNot Tested- Method
- Twenty AI passages rewritten by three named humanisers and checked for whether the detector still catches them.
- Remark
- Not graded; left out of the average.
Raw notes
Twenty AI passages rewritten by three named humanisers and checked for whether the detector still catches them.
Not run: the three named humanisers are not available on a free account, and ZeroGPT's own humanizer is the product being tested, so it cannot stand in for them.
- Needs three humanisers the account does not have.
AI Model Coverage
n/aNot Tested- Method
- The catch rate for each of six AI models, graded on the weakest.
- Remark
- Not graded; left out of the average.
Raw notes
The catch rate for each of six AI models, graded on the weakest.
Not run: one AI model wrote all 20 passages, so there is no per-model rate to compare.
- Needs passages from six models.
Accuracy on AI-Written Text
5/5Tested- Method
- Twenty passages of about 200 to 300 words, two from each of ten kinds of writing (essay, product description, history summary, blog post, short story, cover letter, science explainer, travel guide, opinion piece, café review), were written by one AI model at default settings.
- AI passages caught
- 20 of 20
- Lowest gauge
- 88%
- Remark
- Scale 5: 47 to 50 of 50 caught. With 20 passages the same rate is 19 or 20 of 20.
Raw notes
Twenty passages of about 200 to 300 words, two from each of ten kinds of writing (essay, product description, history summary, blog post, short story, cover letter, science explainer, travel guide, opinion piece, café review), were written by one AI model at default settings. Each was pasted once into the detector and the first verdict taken. A label of AI counts as caught.
All 20 of 20 passages were labelled "Your Text is AI/GPT Generated". The lowest gauge was 88% on a history summary of Roman roads and 18 of the 20 showed 97% or more.
The percentage threshold was not needed: every passage was at 88% or above, so a 50% cut gives the same count.
- Twenty passages is a small set. One model wrote all of them, so the result does not show how ZeroGPT copes with text from other models.
False Positives on Human Text
3/5Tested- Method
- Twenty passages of 220 to 330 words were cut from English books published before 1930 (novels, a memoir, an essay, a play and a science book), so no AI wrote them.
- Human passages flagged
- 2 of 20
- Flagged at a 50% cut
- 1 of 20
- Non-English human passages flagged
- 0 of 12
- Remark
- Scale 3: 4 or 5 of 50 flagged (8 to 10%). Two of 20 is 10%. The cap at 2 applies only above 10%.
Raw notes
Twenty passages of 220 to 330 words were cut from English books published before 1930 (novels, a memoir, an essay, a play and a science book), so no AI wrote them. Each was pasted once. A label of AI, or of most of the text being AI, counts as a false flag.
Two of the 20 were flagged. Alice in Wonderland came back "Your Text is AI/GPT Generated" at 75.7%, and Pride and Prejudice came back "Most of Your Text is AI/GPT Generated" at 39.2%. At a 50% cut only the first would count.
The other 18 were labelled "Human written" and 17 of them scored 0%. The highest of the 18 was 4.7%.
Three passages scanned three times each gave the same label and percentage every time, so the two flags are not a one-off.
- Twenty passages is a small set, so one more flag would move the grade. Books before 1930 are an old style of writing, not a modern student essay.
Mixed & Lightly-Edited Text
4/5Tested- Method
- Ten passages were built from about 110 words of a human passage joined to about 110 words of an AI passage, five with the human text first and five with the AI text first.
- Mixed passages flagged
- 9 of 10
- Labelled AI outright
- 5
- Highlights on the AI half
- 52 of 60
- Remark
- Scale 4: 14 to 16 of 20 flagged. Nine of 10 is 90%, which reaches the 17 to 20 band, but four of the nine were hedged and the highlights spilled onto human text in four passages, so the grade is held at 4.
Raw notes
Ten passages were built from about 110 words of a human passage joined to about 110 words of an AI passage, five with the human text first and five with the AI text first. A label of AI, or one that says the text may include AI parts, counts as flagged. Highlights were counted by hand-checking which half each yellow sentence sat in.
Nine of the 10 were flagged: 5 as "AI/GPT Generated" and 4 with a hedge such as "Likely Human written, may include parts generated by AI/GPT". One passed as "Human written" at 12.8%.
Of 60 highlighted sentences, 52 sat on the AI half and 8 on the human half. Five of those eight were in one passage where the human half was the Alice in Wonderland text that ZeroGPT also flags on its own.
- Ten passages is a small set. The edited-AI-draft case (an AI draft with human edits) was not run; only joined passages were.
Non-English Text
5/5Tested- Method
- Twenty passages in Spanish, French, German and Chinese: two AI and three human in each language.
- Correct labels
- 20 of 20
- Human passages flagged
- 0 of 12
- Scans with no verdict
- 1
- Remark
- Scale 5: 22 to 24 of 24 correct with at most one human passage flagged. With 20 passages, 20 of 20 is the top band.
Raw notes
Twenty passages in Spanish, French, German and Chinese: two AI and three human in each language. The human passages were cut from books before 1930. Each was pasted once and marked right or wrong by the label.
All 20 were labelled correctly: the 8 AI passages as AI and the 12 human passages as "Human written". No human passage in any language was flagged.
The two Chinese AI passages were short, about 130 characters. ZeroGPT labelled them "Most Likely AI/GPT generated" and asked for "more text for a more accurate result", while its gauge read 0%.
One German human passage gave no verdict for 90 seconds on the first try. The same passage came back "Human written" on the second try.
- Eight AI passages and twelve human ones are a small set, and the AI Chinese passages were short, so the label and the gauge disagreed.
Document & File Support
2/5Tested- Method
- Each of ten points was tried on 6 November 2026 on the free account: paste, a PDF, a Word file, a plain-text file, a batch of files, a scan by web address, a cap of 2,000 words or more, a browser extension or add-in, an API and an export of the result.
- Points that worked
- 5 of 10
- Cap per scan
- 15,000 characters
- Remark
- Scale 2: 4 or 5 points.
Raw notes
Each of ten points was tried on 6 November 2026 on the free account: paste, a PDF, a Word file, a plain-text file, a batch of files, a scan by web address, a cap of 2,000 words or more, a browser extension or add-in, an API and an export of the result.
Five points worked: paste, a PDF, a Word file, a plain-text file and the cap. The cap is 15,000 characters a scan, about 2,300 words by my own passages (1,223 characters were 188 words). Longer text is cut at 15,000.
Five did not: the batch page says "You can upload up to 0 files" on the free plan, there is no scan by web address, I found no extension or add-in, the API was not tried with a balance of $0.00, and Export to PDF opens an upgrade box.
My first Word file, a near-empty one I made myself, loaded nothing. A normal Word file loaded in under two seconds.
- The API and extension count only when they work end to end on this account, so they are marked as not working here, not as missing for every plan.
Score Clarity & Evidence
3/5Tested- Method
- Ten points were checked on scans already run: a plain call, a percentage, a stated confidence band, sentence highlighting, an explanation, a downloadable report, a scan history, repeatability, a statement of limits and no upsell blocking the result.
- Points that passed
- 6 of 10
- Repeat scans with the same result
- 9 of 9
- Remark
- Scale 3: 6 or 7 points.
Raw notes
Ten points were checked on scans already run: a plain call, a percentage, a stated confidence band, sentence highlighting, an explanation, a downloadable report, a scan history, repeatability, a statement of limits and no upsell blocking the result.
Six passed: a plain label, a percentage gauge, yellow highlights on sentences, the same result on three repeat scans of three passages, a note for educators and no upsell in the way of the result. The Terms of Use say "No AI detector, including ZeroGPT, can determine authorship with absolute certainty in every case".
Four failed: no stated margin or confidence band, no reason given for why a sentence is flagged, a PDF report that needs a paid plan, and an empty history after about a hundred scans while signed in.
- The history page may record only API calls; the scans were made on the web page.
Scan Speed
5/5Tested- Method
- All 20 AI scans were timed from pressing Detect Text to the verdict on screen, on one connection, each from a fresh page load.
- Median scan time
- 1.2 seconds
- Slowest of 20
- 2.9 seconds
- Silent hangs
- 2
- Remark
- Scale 5: median under 5 seconds.
Raw notes
All 20 AI scans were timed from pressing Detect Text to the verdict on screen, on one connection, each from a fresh page load.
The median was 1.2 seconds. The fastest was 0.8 seconds and the slowest 2.9 seconds. Uploading a file took under two seconds.
Twice in about a hundred scans the button did nothing and no error showed, once for 90 seconds. Scanning the same passage again worked both times. Those two are not in the 20 timed scans.
- Times include my own network. The free plan showed no queue or slow-down in about a hundred scans.
Pricing & Free-Tier Limits
5/5Tested- Method
- The free plan was used for about a hundred scans, then the account page and the public pricing page were read in both billing views on 6 November 2026.
- Free cap per scan
- 15,000 characters
- Cheapest plan that lifts it, billed yearly
- $8.99 a month
- Same plan, billed monthly
- $14.99 a month
- Remark
- Scale 5: a usable free tier of 1,000 words a day or more, or a paid plan under $10 a month that lifts limits and adds files and bulk.
Raw notes
The free plan was used for about a hundred scans, then the account page and the public pricing page were read in both billing views on 6 November 2026.
The free plan stopped nothing in about a hundred scans, over 12,000 words counted on the page. It needs no card, takes up to 15,000 characters a scan and has no batch scan, PDF report or saved history.
Pro is $8.99 a month billed yearly ($107.88 a year) or $14.99 a month. It lifts the cap to 100,000 characters and adds 50 batch files. Plus is $13.99 a month yearly and Max $19.99 with 150,000 characters.
The detector is not metered by the word on any plan, so a cost per 1,000 words cannot be worked out.
- Business / EDU plans have their own tab and were not read. No plan was bought.
Other features checked
Coverage
Tested
- Every passage pasted once into the detector box and the first verdict taken
- Text, PDF and Word uploads, the 15,000-character cap, the batch page, the PDF export and the history page
- Three passages scanned three times each for repeatability
- The pricing page in both billing views, the Terms of Use and the Trustpilot page read
Not tested or limited
- The corpus is smaller than the methodology asks (20 AI, 20 human, 10 mixed, 20 non-English instead of 50, 50, 20 and 24), so each grade applies the published scale to the rate and the counts are small
- All 20 AI passages were written by one AI model at default settings, so how ZeroGPT copes with other models was not measured
- The human passages are English books from before 1930, an old style of writing, not modern student essays
- Resistance to humanisers was not tested: it needs three named humanisers and a free account has none
- The API, a browser extension and the batch scan were not run: the account balance was $0.00 and the free plan allows 0 batch files
- The Chinese AI passages were short (about 130 characters), so the tool asked for more text
Still to test
- A larger set, with passages from six AI models and modern human essays
- Resistance to three named humanisers
- The batch scan, PDF report, API and history on a paid month
- Business / EDU plans