Arena Labs · Repeated-humanization experiment
Does Humanizing Text More Than Once Work Better?
Not reliably. Repeated passes caused small, irregular detector changes, but no tool improved consistently. Meanwhile, meaning preservation fell and factual errors accumulated.
Test design
The same outputs were repeatedly humanized
Three 250-word source texts were processed sequentially, with outputs saved after the first, second, third and fifth pass.
3 source texts
A student essay, business article and SEO article, each approximately 250 words.
7 humanizers
GPTHuman, StealthGPT, Humanize AI Pro, HIX AI, BypassGPT, Netus AI and NoteGPT AI Humanizer.
84 saved outputs
Four checkpoints for each of 21 tool-and-text combinations, requiring 105 sequential humanizer runs.
252 detector evaluations
Every saved output was checked by Pangram, GPTZero and Originality.ai.
Detector results by tool
Humanizer | 1× | 2× | 3× | 5× | Meaning at 5× | Factual errors at 5× |
|---|---|---|---|---|---|---|
GPTHuman | 88.9% (8/9) | 88.9% (8/9) | 77.8% (7/9) | 88.9% (8/9) | 8.7 / 10 | 0.67 |
StealthGPT | 77.8% (7/9) | 77.8% (7/9) | 77.8% (7/9) | 66.7% (6/9) | 6.7 / 10 | 2.00 |
Humanize AI Pro | 77.8% (7/9) | 66.7% (6/9) | 77.8% (7/9) | 66.7% (6/9) | 6.3 / 10 | 2.33 |
BypassGPT | 66.7% (6/9) | 77.8% (7/9) | 66.7% (6/9) | 66.7% (6/9) | 5.7 / 10 | 2.67 |
HIX AI | 66.7% (6/9) | 66.7% (6/9) | 55.6% (5/9) | 66.7% (6/9) | 5.7 / 10 | 3.00 |
Netus AI | 66.7% (6/9) | 55.6% (5/9) | 66.7% (6/9) | 55.6% (5/9) | 6.0 / 10 | 2.67 |
NoteGPT AI Humanizer | 44.4% (4/9) | 33.3% (3/9) | 44.4% (4/9) | 33.3% (3/9) | 4.3 / 10 | 3.33 |
Key findings
Extra passes added damage without reliable detector benefit
Detector results fluctuated by one evaluation at a time, while every measured quality dimension deteriorated as the text was rewritten again.
Detector changes were small and irregular
No tool showed a sustained upward trend. BypassGPT improved temporarily at 2×, while other tools oscillated or declined.
Meaning declined for every tool
The 1× to 5× decrease ranged from 0.6 points for GPTHuman to 1.8 points for HIX AI, Netus AI and NoteGPT AI Humanizer.
Factual errors accumulated
Average errors per text increased for every tool; the 5× values ranged from 0.67 to 3.33.
Text length continued drifting
Six tools expanded their outputs with repeated passes, while NoteGPT progressively compressed them.
Meaning preservation by pass
Humanizer | 1× | 2× | 3× | 5× | Change |
|---|---|---|---|---|---|
GPTHuman | 9.3 | 9.0 | 8.8 | 8.7 | -0.6 |
StealthGPT | 8.3 | 7.8 | 7.2 | 6.7 | -1.6 |
Humanize AI Pro | 8.0 | 7.4 | 6.8 | 6.3 | -1.7 |
BypassGPT | 7.4 | 6.9 | 6.3 | 5.7 | -1.7 |
HIX AI | 7.5 | 6.9 | 6.3 | 5.7 | -1.8 |
Netus AI | 7.8 | 7.2 | 6.6 | 6.0 | -1.8 |
NoteGPT AI Humanizer | 6.1 | 5.5 | 4.9 | 4.3 | -1.8 |
Factual errors by pass
Humanizer | 1× | 2× | 3× | 5× |
|---|---|---|---|---|
GPTHuman | 0.33 | 0.33 | 0.67 | 0.67 |
StealthGPT | 0.67 | 1.00 | 1.67 | 2.00 |
Humanize AI Pro | 0.67 | 1.33 | 1.67 | 2.33 |
BypassGPT | 1.00 | 1.67 | 2.00 | 2.67 |
HIX AI | 1.00 | 1.67 | 2.33 | 3.00 |
Netus AI | 0.67 | 1.33 | 2.00 | 2.67 |
NoteGPT AI Humanizer | 1.67 | 2.33 | 2.67 | 3.33 |
Readability drift by pass
Humanizer | 1× | 2× | 3× | 5× |
|---|---|---|---|---|
GPTHuman | +0.3 | +0.4 | +0.5 | +0.6 |
StealthGPT | +0.8 | +1.2 | +1.6 | +2.1 |
Humanize AI Pro | +0.9 | +1.4 | +1.9 | +2.5 |
BypassGPT | +1.2 | +1.7 | +2.2 | +2.8 |
HIX AI | +1.1 | +1.6 | +2.1 | +2.7 |
Netus AI | +0.8 | +1.3 | +1.8 | +2.4 |
NoteGPT AI Humanizer | +1.6 | +2.2 | +2.8 | +3.4 |
Output-to-input word ratio by pass
Humanizer | 1× | 2× | 3× | 5× |
|---|---|---|---|---|
GPTHuman | 1.04 | 1.05 | 1.05 | 1.06 |
StealthGPT | 1.09 | 1.13 | 1.17 | 1.22 |
Humanize AI Pro | 1.11 | 1.16 | 1.20 | 1.26 |
BypassGPT | 1.15 | 1.20 | 1.25 | 1.32 |
HIX AI | 1.08 | 1.14 | 1.19 | 1.25 |
Netus AI | 1.10 | 1.15 | 1.20 | 1.26 |
NoteGPT AI Humanizer | 0.93 | 0.90 | 0.87 | 0.84 |
Methodology
How the numbers were calculated
Detector results are binary rates; the remaining measures describe changes against the original source text.
Detector pass rate
Each tool had nine evaluations per checkpoint: three texts across three detectors. Valid rates therefore move in 11.1-point increments.
Meaning preservation
The reported 1–10 score is the average of three ratings, one for each source text, measured against the original.
Factual errors
Errors were counted across the three texts and divided by three, so averages move in increments of 0.33.
Scope and limitation
This small point-in-time test used three texts and default settings. It demonstrates what happened in these 252 evaluations, not a universal rule for every document or future product version.
Repeated-humanization experiment FAQ
- Does humanizing text twice improve detector performance?
- Not reliably. Some results moved by one detector evaluation, but no tool showed a sustained improvement across later passes.
- What changed after additional passes?
- Meaning preservation declined, readability drift increased, factual errors accumulated and output length continued changing.
- Which tool performed best?
- GPTHuman passed 8 of 9 checks at 1×, 2× and 5×, never fell below 7 of 9, and showed the smallest decline in the measured quality indicators.
- How many times should text be humanized?
- Within this experiment, one pass was the rational stopping point because later passes added quality risk without producing dependable detector improvement.
Continue exploring
Compare this experiment with the main benchmark
The primary leaderboard uses a broader benchmark cycle and remains the main source for overall rankings.