Test log
Undetectable AI test data
Raw results for the Undetectable 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
- 10
- Screenshots
- 7
- 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: 50 of 50 | 5/5 |
| T04 | False Positives on Human Text | Human passages flagged as AI: 0 of 50 | 5/5 |
| T05 | Mixed & Lightly-Edited Text | Mixed passages flagged: 15 of 20 | 4/5 |
| T06 | Non-English Text | Correct verdicts: 24 of 24 | 5/5 |
| T07 | Document & File Support | File points that worked: 6 of 10 | 3/5 |
| T08 | Score Clarity & Evidence | Clarity points that passed: 7 of 10 | 3/5 |
| T09 | Scan Speed | Median scan time: 6.4 seconds | 4/5 |
| T10 | Pricing & Free-Tier Limits | Free scans before a limit: 151 and more | 5/5 |
| Overall (in the review) | 4.9/5 | ||
Resistance to Humanisers & Paraphrasers
n/aNot Tested- Method
- Twenty AI passages rewritten by three named humanisers.
- Remark
- Not graded; left out of the average.
Raw notes
Twenty AI passages rewritten by three named humanisers.
Not run: the methodology asks for three named humanising tools, and none was approved for this run.
- Needs three named humanisers.
AI Model Coverage
n/aNot Tested- Method
- Catch rate for each of six AI models.
- Remark
- Not graded; left out of the average.
Raw notes
Catch rate for each of six AI models.
Not run: all 50 AI passages came from one model, so a per-model table cannot be built.
- Needs passages from GPT, Gemini, Llama and DeepSeek.
Accuracy on AI-Written Text
5/5Tested- Method
- Fifty AI-written passages of about 300 words, five from each of ten prompts (a history essay, a product review, a hobby blog post, a science explainer, an opinion piece, local news, a how-to guide, a story opening, a novel essay and a letter), were written by Claude Sonnet 5.5 at default settings and pasted in once each.
- AI passages caught
- 50 of 50
- Lowest score on an AI passage
- 65%
- Remark
- Scale 5: 47 to 50 of 50 caught.
Raw notes
Fifty AI-written passages of about 300 words, five from each of ten prompts (a history essay, a product review, a hobby blog post, a science explainer, an opinion piece, local news, a how-to guide, a story opening, a novel essay and a letter), were written by Claude Sonnet 5.5 at default settings and pasted in once each. The first verdict was taken.
All 50 were labelled "AI generated", so 50 of 50 were caught.
The lowest scores were on the science explainers for a teenager: 65%, 66%, 83%, 90% and 91%. Every other prompt scored 75% or more on every passage, and 35 of the 50 scored 99%.
The corpus holds one model family only, so this says nothing about GPT, Gemini, Llama or DeepSeek text.
- All 50 passages came from one model family, so the methodology's six-model split could not be built. The same 50 passages are kept in the kit for every later detector.
False Positives on Human Text
5/5Tested- Method
- Fifty human passages of about 300 words, two from each of 25 public-domain books written between 1776 and 1915 (Thoreau, Mill, Austen, Darwin, Franklin, Douglass, Wells, Wilde, Dickens, Melville and others, from Project Gutenberg), were pasted in once each.
- Human passages flagged as AI
- 0 of 50
- Highest AI score on a human passage
- 11%
- Remark
- Scale 5: at most 1 of 50 flagged.
Raw notes
Fifty human passages of about 300 words, two from each of 25 public-domain books written between 1776 and 1915 (Thoreau, Mill, Austen, Darwin, Franklin, Douglass, Wells, Wilde, Dickens, Melville and others, from Project Gutenberg), were pasted in once each. None can have been written by an AI.
All 50 were labelled "human written", so 0 of 50 were flagged.
The highest AI score on any human passage was 11%; 44 of the 50 scored 1% to 3%.
- The human passages are 19th and early 20th century books. They are certain to be human, but they are not the essays, emails and forum posts that people check today, and none was written by a non-native English speaker.
Mixed & Lightly-Edited Text
4/5Tested- Method
- Twenty mixed passages: ten AI drafts (A41 to A50) edited in a writer's voice, with sentences reworded, two ideas swapped, a generic line cut and one personal detail added; and ten passages made of half a human book passage joined to half an AI passage, human first in five and AI first in five.
- Mixed passages flagged
- 15 of 20
- Edited AI drafts flagged
- 10 of 10
- Joined passages flagged
- 5 of 10
- Remark
- Scale 4: 14 to 16 of 20 flagged.
Raw notes
Twenty mixed passages: ten AI drafts (A41 to A50) edited in a writer's voice, with sentences reworded, two ideas swapped, a generic line cut and one personal detail added; and ten passages made of half a human book passage joined to half an AI passage, human first in five and AI first in five.
15 of 20 were flagged as containing AI: all 10 edited drafts, and 5 of the 10 joined passages.
The five misses scored between 36% and 50% and were labelled "human written", all of them joined passages. One sat exactly on 50%.
Where it flagged a joined passage, the highlighting landed on the right half: in the AI-then-human example the Dutch oven bread instructions were red and the sentences from Franklin's Autobiography were green.
- The edits were made by Claude in the voice of a human editor, not by a person; a real editor's changes may behave differently. Highlight accuracy was checked by eye on four joined passages, not all ten.
Non-English Text
5/5Tested- Method
- Twenty-four passages in Spanish, French, German and Simplified Chinese: three AI and three human in each.
- Correct verdicts
- 24 of 24
- Human passages flagged
- 0
- Remark
- Scale 5: 22 to 24 correct, at most one human passage flagged.
Raw notes
Twenty-four passages in Spanish, French, German and Simplified Chinese: three AI and three human in each. The AI passages were written by Claude. The human passages are from Don Quijote (1605), Les Misérables (1862), Kafka's Die Verwandlung (1915) and three stories by Lu Xun (1918 to 1921).
All 24 verdicts were right: the 12 AI passages were labelled "AI generated" and the 12 human passages "human written", so no human passage was flagged.
The closest call was a French passage from Les Misérables, scored 40% AI and still labelled human. The AI passages scored 78% to 99%.
- Three passages per language is a small sample. The Chinese texts are early 20th century fiction in Simplified script.
Document & File Support
3/5Tested- Method
- Ten points checked on the test date, each one only if it worked end to end on my free account.
- File points that worked
- 6 of 10
- Words allowed in one scan
- 10,000
- Remark
- Scale 3: 6 or 7 points.
Raw notes
Ten points checked on the test date, each one only if it worked end to end on my free account.
Six of ten worked: pasting text, uploading a .txt file, uploading a .pdf, uploading a .docx, a 2,142-word scan (the box accepts up to 10,000 words), and exporting the result as a PDF report.
Four did not count. The upload box takes one file at a time, with no batch. I found no way to check a web address. The Chrome extension is listed but was not installed in the test browser. The API key page says the key is issued only when a plan is active.
- The extension and the API were not tested, because the API needs a paid plan and the extension needs installing.
Score Clarity & Evidence
3/5Tested- Method
- Ten points checked on the scans already run, plus one passage scanned three times in a row, twice (an AI passage and a mixed one).
- Clarity points that passed
- 7 of 10
- Repeat scans with the same score
- 6 of 6
- Remark
- Scale 3: 6 or 7 points.
Raw notes
Ten points checked on the scans already run, plus one passage scanned three times in a row, twice (an AI passage and a mixed one).
Seven of ten passed: a plain AI-or-human label, a percentage, sentence highlighting in red and green, a PDF report, the same score on every repeat (99% three times, 61% three times), a note that results are probabilities and not proof, and no upsell blocking the result.
Three failed: no confidence band, no explanation of why a passage scored what it did, and no scan history. My account's Documents page stayed empty after 151 scans.
- The history check relies on the Documents page, which may be meant for humaniser output only.
Scan Speed
4/5Tested- Method
- All 50 scans in the AI-accuracy test were timed from pressing Check for AI to the verdict on screen.
- Median scan time
- 6.4 seconds
- Slowest scan
- 36.1 seconds
- Remark
- Scale 4: median 5 to 9.9 seconds.
Raw notes
All 50 scans in the AI-accuracy test were timed from pressing Check for AI to the verdict on screen.
The median was 6.4 seconds, the fastest 5.8 and the slowest 36.1. The 50 human scans had a median of 3.8 seconds. One 2,142-word scan took 13.3 seconds. No scan failed or timed out.
- Timed on one connection, through a recorded browser, which may add a little.
Pricing & Free-Tier Limits
5/5Tested- Method
- I scanned until the tool stopped me, then read the public pricing page and the plan screen on the test date.
- Free scans before a limit
- 151 and more
- Cheapest paid plan
- $5 a month yearly
- Remark
- Scale 5: a free tier of 1,000 words a day or more.
Raw notes
I scanned until the tool stopped me, then read the public pricing page and the plan screen on the test date.
The detector never stopped me: 151 scans and about 44,000 words in one session, on an account with a balance of 0 credits and no plan. A scan may be 10,000 words. Upload, highlighting and the PDF report were all free.
Paid plans buy humaniser words, not detection: the cheapest is 10,000 words a month for $5 a month paid yearly ($60 a year); paid monthly, plans start at 20,000 words for $19 a month. The 3-day trial is offered on yearly plans only; I did not start it.
- I did not find a daily cap, but I cannot say none exists. The free detector may be a different model from the one behind the API.
Other features checked
Coverage
Tested
- Fifty AI, fifty human, twenty mixed and twenty-four non-English passages, each scanned once
- Three file types, a long scan, a PDF export, repeat scans, the account pages and the pricing page
Not tested or limited
- Every AI passage came from one model family (Claude), so the six-model split was not built and AI model coverage is not graded
- The human passages are public-domain books from 1776 to 1915, certain to be human but unlike modern writing; none is from a non-native English speaker
- The humaniser test was not run: no humanising tools were approved for this run
- The mixed-text edits were made by Claude in an editor's voice, not by a person
- The Chrome extension and the API were not tested; the API key needs a paid plan
- I could not tell whether a daily limit exists above the 44,000 words I scanned in one session
Still to test
- Add passages from GPT, Gemini, Llama and DeepSeek
- Run the humaniser test with three named humanising tools
- Test the API and extension on a paid month