Arena Labs · Text-length experiment
Does Text Length Affect AI Humanizer Performance?
Yes. Across 18 source texts, performance generally peaked between 100 and 250 words, then declined on very short and very long inputs. GPTHuman recorded the highest aggregate pass rate at every tested length.
Test design
A controlled comparison across six input lengths
The experiment used default settings and the same seven humanizers represented in the current Humanizer Arena benchmark.
18 source texts
Six target lengths across three content families: student or essay, business or informational, and SEO or article writing.
6 input lengths
25, 50, 100, 250, 500 and 1,000 words.
7 humanizers
GPTHuman, StealthGPT, Humanize AI Pro, HIX AI, BypassGPT, Netus AI and NoteGPT AI Humanizer.
3 named detectors
Pangram, GPTZero and Originality.ai.
Overall results
Rank | Humanizer | Average pass rate | Best length | Worst length |
|---|---|---|---|---|
1 | GPTHuman | 90.6% | 250 words · 93.4% | 25 words · 86.4% |
2 | StealthGPT | 81.7% | 250 words · 87.6% | 25 words · 72.1% |
3 | HIX AI | 79.8% | 250 words · 86.1% | 25 words · 71.3% |
4 | Humanize AI Pro | 79.2% | 250 words · 85.2% | 25 words · 68.9% |
5 | BypassGPT | 74.6% | 250 words · 81.4% | 25 words · 64.8% |
6 | Netus AI | 71.5% | 250 words · 78.3% | 25 words · 61.2% |
7 | NoteGPT AI Humanizer | 51.6% | 250 words · 67.2% | 25 words · 22.4% |
Performance by input length
Humanizer | 25 | 50 | 100 | 250 | 500 | 1,000 | Average |
|---|---|---|---|---|---|---|---|
GPTHuman | 86.4% | 89.7% | 92.1% | 93.4% | 91.8% | 90.2% | 90.6% |
StealthGPT | 72.1% | 79.8% | 85.3% | 87.6% | 84.9% | 80.4% | 81.7% |
HIX AI | 71.3% | 78.6% | 84.2% | 86.1% | 81.7% | 76.9% | 79.8% |
Humanize AI Pro | 68.9% | 76.4% | 83.7% | 85.2% | 82.6% | 78.1% | 79.2% |
BypassGPT | 64.8% | 71.5% | 79.1% | 81.4% | 77.8% | 73.2% | 74.6% |
Netus AI | 61.2% | 68.9% | 75.6% | 78.3% | 74.1% | 70.8% | 71.5% |
NoteGPT AI Humanizer | 22.4% | 41.7% | 58.9% | 67.2% | 63.5% | 55.8% | 51.6% |
What changed with length
The strongest patterns in the data
The test points to a mid-length sweet spot, but the size of the penalty at the extremes varied considerably by tool.
250 words was the peak
All seven humanizers reached their highest detector pass rate on the 250-word inputs.
Very short text was hardest
Every tool produced its weakest result at 25 words. GPTHuman led that group at 86.4%; NoteGPT fell to 22.4%.
Long inputs also lost ground
From 250 to 1,000 words, results fell by 3.2 points for GPTHuman and by 7.1 to 11.4 points for the other tools.
GPTHuman was least length-sensitive
It led every length group and retained 90.2% at 1,000 words, the smallest decline from its own peak.
Methodology
How to interpret this experiment
This compact Arena Labs test isolates one practical variable: input length. It is a point-in-time experiment rather than a replacement for the main benchmark.
Scope
18 source texts: three writing families at each of six target lengths, producing 126 planned tool-input combinations across seven humanizers.
Test period and settings
Testing was conducted in September 2026. Every humanizer was run using its default settings.
Detector pass rate
For each length, the reported score is the average share of evaluations classified as human across Pangram, GPTZero and Originality.ai.
Failures and limitations
Submission failures are reported separately from detector pass rates. Results describe this fixed 18-text test and may change as tools and detectors update.
Text-length experiment FAQ
- Does text length affect AI humanizer performance?
- Yes. All seven tools improved between 25 and 250 words, then declined on the 500- and 1,000-word inputs.
- What was the best input length?
- Every tested tool recorded its highest detector pass rate at 250 words.
- Which humanizer performed best?
- GPTHuman led all six length groups and averaged 90.6% when the six length groups were weighted equally.
- Does this prove one tool will always work better?
- No. This was a point-in-time test using 18 source texts. Individual documents, product updates and detector changes can produce different results.
Continue exploring
Compare the experiment with the main benchmark
The primary leaderboard uses a broader benchmark cycle and should remain the main source for overall rankings.