A rewrite can sound cleaner while quietly changing the facts. That is the failure mode I wanted to make visible when building Humanizer.best, an AI humanizer review directory.
We tested tools with one fixed paragraph, saved the first complete output and compared the result against explicit factual constraints. The three highest editorial scores in this first batch were The Prompt Index, Voilà and MyDetector. All three preserved the main facts in this sample. Their differences were in how much they changed, how easy the changes were to inspect and what their free access actually covered.
The ranking describes this short case study. It cannot establish universal model quality. I maintain Humanizer.best, and the testing workflow used AI assistance. This article was drafted and edited with AI. Tests and pricing were checked on October 27, 2026. The screenshots are real browser captures; the ratings are editorial judgments. We did not validate detector evasion or call paid APIs.
Treat the paragraph like a test fixture
The input was a fictional, AI-generated report containing 93 words:
A local library tested a four-day schedule for six weeks. It kept the same 40 weekly opening hours, closing Wednesdays and extending Monday and Thursday evenings. Visits rose from 1,200 to 1,320 a week, although Wednesday visitors had to change plans. Staff reported fewer interruptions, but productivity was not measured. Longer evenings may help people who work during the day; this does not prove every library should open fewer days. Before making the change permanent, the team will survey residents, compare costs and repeat the trial in winter, when travel habits may differ.
There is enough structure here to catch several kinds of semantic drift. A model must preserve the six-week duration, the 40-hour total, the named weekdays and both attendance figures. It must also preserve the distinction between an observation and a measured outcome.
Fewer interruptions do not establish higher productivity. Longer evenings may help daytime workers; that is not a demonstrated benefit for every library. Comparing costs does not imply that savings have already been found.
Those qualifications belong in the test fixture alongside the numbers. A rewrite that keeps every number but changes “may” into a confident conclusion can still fail the task.
The downloadable test kit includes the same paragraph. We stored a SHA-256 hash of the trimmed UTF-8 input so an accidental edit would be detectable before comparing outputs. The source fixture, captured outputs and scoring records are separate files, which makes an unintended change easier to trace.
Keep access failures outside the quality score
The discovery set came from the first 30 returned Google organic results for each of three queries: ai humanizer, humanizer ai and undetectable ai. Deduplication produced 35 product domains.
We obtained 21 complete, comparable outputs. The remaining 14 encountered sign-in, verification, payment or output failures. Those records remain visible but unscored. A login screen is evidence about access, not evidence that the underlying rewrite is bad.
For completed tests, we retained the first complete captured output in the stated mode. We did not rerun a disappointing result until it looked competitive. Where a technical retry was necessary, the product profile records it.
If one tool gets five attempts and another gets one, the comparison has uneven generation budgets. Saving the first result makes the evaluation easier to audit, even though it does not measure output variance.
Separate the scoring dimensions
The composite score uses six dimensions, each rated from zero to ten:
| Dimension | Weight | What it measures here |
|---|---|---|
| Meaning and facts | 30% | Numbers, schedule, qualifications and unsupported additions |
| Readability | 25% | Grammar, punctuation and ease of comprehension |
| Natural voice | 15% | Rhythm and phrasing without forced informality |
| Editing support | 15% | Observed comparison and editing controls |
| Value | 10% | Completed free access and clarity of limits |
| Integrations | 5% | Verified official humanizer integration documentation |
Here is the calculation, with integer weighting before the final rounding:
const weights = [30, 25, 15, 15, 10, 5];
function overallScore(values) {
if (values.length !== weights.length) throw new Error('Six scores required');
if (values.some(v => !Number.isFinite(v) || v < 0 || v > 10)) {
throw new Error('Scores must be between 0 and 10');
}
const weighted = values.reduce((sum, value, i) => sum + value * weights[i], 0);
return Math.round(weighted / 10) / 10;
}
console.log(overallScore([10, 9, 8, 8, 8, 0])); // 8.5: The Prompt Index
console.log(overallScore([10, 9, 8, 5, 8, 0])); // 8.0: Voilà
console.log(overallScore([10, 8, 7, 7, 8, 0])); // 7.9: MyDetector
The decimal makes sorting convenient. It does not make the underlying judgment statistically precise. In particular, a 0.1-point gap should not be read as a proven difference across future inputs.
The full methodology explains the evidence threshold. An integration score of zero means that no qualifying official humanizer interface was verified in the reviewed material. It is not proof that no private or future API could exist.
1. The Prompt Index: useful editing, visible changes
The Prompt Index scored 8.5/10 in the guest General / Standard configuration, with its “Remove AI fingerprints” option enabled. The full sample completed without registration.
Its rewrite was restrained. The first sentence stayed unchanged. The weekly total, Wednesday closure, Monday and Thursday evenings and attendance increase all survived. The productivity caveat also remained, expressed in direct language rather than converted into a claim about improved output.
The most useful interface feature was the Compare tab. Inserted and deleted words had separate visual treatments, making it easier to locate changes before judging whether they were acceptable.
A word diff cannot establish semantic equivalence. It can, however, direct attention to the clauses that need review. That is valuable when the goal is controlled editing rather than maximum lexical change.
We did not verify selective sentence generation or a separate line-level diff. Those are different capabilities from highlighting changed words, so they were not inferred from the Compare tab.
The observed guest allowance was five runs per calendar month, with 5,000 characters per run. The free account allowance was listed as ten runs per 30 days and 10,000 characters per run. Paid plans used runs plus per-request character limits, not a simple monthly word bank.
The first result needed little factual repair. The score advantage over Voilà came mainly from inspection support, not from a claim that heavier rewriting was inherently better.
Full Prompt Index review · Screenshots, output and parameter table
2. Voilà: a light copy edit that kept the constraints
Voilà scored 8.0/10 using its free base model, English language and Neutral tone. The observed editor cap was 1,000 characters, which was enough for the unchanged sample.
The result stayed close to the source. It adjusted punctuation and phrasing while retaining the figures, schedule, productivity limitation and conditional conclusion. It also kept the plan to survey residents, compare costs and repeat the trial in winter.
I would keep most of this rewrite. The input was already clear, and a light copy edit suited it. A more elaborate rewrite has more opportunities to introduce assumptions that were never present.
The tradeoff was inspectability. We verified a complete rewrite and a Copy action, but not an inline word diff or a completed selective sentence edit. The source and result still needed to be compared manually.
Voilà's broader suite pricing uses requests and model-specific conditions. Its suite Free plan listed 250 requests, while Premium listed 3,000 requests per month. Those suite allowances should not be treated as a separately verified renewal policy for the public guest humanizer.
For integrations, check which direction the API key travels. A suite accepting your provider key does not necessarily expose a humanizer endpoint your application can call. We did not verify a dedicated humanizer API, MCP server or agent skill for this tested workflow.
For this paragraph, Voilà was effective precisely because it did not try to add a new argument. It gave us a readable version of the same report.
Full Voilà review · Screenshots, output and parameter table
3. MyDetector: more conversational, with more text to inspect
MyDetector scored 7.9/10 in Standard / Normal / General Writing, using the Base model and Natural setting. Output language was set to match the input.
The rewrite broke the report into more short sentences. It retained the 40-hour total, both attendance numbers and the lack of a productivity measurement. The proposed benefit remained conditional, and the ending still called for a survey, cost review and winter repeat.
The draft was easier to read in small pieces, but it also added conversational commentary. The phrase “Of course” before the Wednesday inconvenience did not supply evidence or clarify the schedule. It was a voice choice we would probably remove from a neutral report.
The saved text contains 131 whitespace-separated words; the interface displayed 133. We keep that distinction rather than pretending every product uses the same tokenization rule. Neither number should be substituted for model tokens or billing credits.
The interface provided source/result comparison, Copy and Download actions, and clickable sentence controls. Clicking a sentence did not establish a verified replacement in this session. A visible control is not the same thing as a completed editing operation.
Pricing had a further implementation detail: quotas consumed characters at different rates depending on the model. The official table listed 1,000 characters per quota for the Basic humanizer model and 200 for Pro. That difference matters more than comparing subscription headlines alone.
The guest sample completed, and the registered Free plan listed 20 quotas per month. Paid Basic and Pro plans showed $6.99 for 1,100 quotas and $14.99 for 3,000. We did not buy them or benchmark their paid output.
Full MyDetector review · Screenshots, output and parameter table
What this suggests for an evaluation harness
For the next version of this evaluation, I would keep the same separation between generation, evidence capture and scoring. Store the original input, configuration, first output, date and screenshots before attaching a score. Otherwise, it becomes too easy to overwrite an inconvenient result while experimenting.
Some checks can be automated cheaply. Verify that expected numbers and named days appear, flag a large increase in length and record whether a full output was returned. Treat those checks as alerts rather than a semantic pass: a paragraph can contain the right numbers in the wrong relationship.
The next layer needs an explicit checklist for claims. Did a possibility become a certainty? Did a missing measurement become a positive result? Did a future comparison acquire an assumed outcome? Those are more useful questions than asking whether the rewrite merely “sounds human.”
For a stronger follow-up study, I would add multiple genres, independently scored outputs and repeated runs with a fixed budget. That would let us estimate consistency and reviewer agreement. This first batch does not provide either, so its ranking should remain a starting point for inspection.
The result I would use today
For this sample, The Prompt Index offered the best combination of factual preservation and visible change inspection. Voilà was a good example of restrained editing. MyDetector preserved the evidence while producing a more conversational, longer draft.
Detector bypass and plagiarism clearance were outside this test. When choosing a rewrite for an actual report, I would inspect its claims before judging its stylistic variety.
The linked profiles preserve the screenshots and complete rewrites, while the separate review articles explain the judgments. Product updates will revise those existing pages rather than create a new article for the same domain. That gives future tests a stable place to show what changed.
Top comments (0)