Arena Labs · Source-model experiment
Do AI Humanizers Work Equally Well on ChatGPT, Claude and Gemini Text?
No. Every tested humanizer recorded its highest detector-performance index on ChatGPT source text. GPTHuman led all three model families, while the gap between each tool’s strongest and weakest source model ranged from 5.2 to 10.1 points.
Results supplied
A comparison across three source-model families
The supplied tables compare seven humanizers across ChatGPT, Claude and Gemini using detector-performance and factual-preservation indices.
3 source-model families
ChatGPT, Claude and Gemini source text were reported separately.
7 humanizers
The same seven tools appear in each source-model comparison.
21 model-tool aggregates
Each humanizer has one aggregate result for each of the three source-model families.
2 reported indices
Detector performance is the primary comparison; factual preservation is reported separately.
Overall results
Rank | Humanizer | Detector performance mean | Factual preservation mean | Detector max-min spread |
|---|---|---|---|---|
1 | GPTHuman | 89.7 | 98.3 | 5.5 |
2 | StealthGPT | 80.2 | 89.2 | 8.3 |
3 | HIX AI | 79.4 | 90.0 | 7.4 |
4 | Humanize AI Pro | 77.1 | 84.2 | 5.3 |
5 | Netus AI | 74.8 | 85.6 | 6.5 |
6 | BypassGPT | 73.6 | 81.4 | 5.2 |
7 | NoteGPT AI Humanizer | 48.7 | 69.4 | 10.1 |
Detector performance by source model
Humanizer | ChatGPT | Claude | Gemini | Mean | Max-min spread |
|---|---|---|---|---|---|
GPTHuman | 93.1 | 88.4 | 87.6 | 89.7 | 5.5 |
StealthGPT | 85.0 | 78.9 | 76.7 | 80.2 | 8.3 |
HIX AI | 84.2 | 76.8 | 77.2 | 79.4 | 7.4 |
Humanize AI Pro | 80.3 | 75.0 | 76.0 | 77.1 | 5.3 |
Netus AI | 78.6 | 72.1 | 73.7 | 74.8 | 6.5 |
BypassGPT | 76.4 | 71.2 | 73.2 | 73.6 | 5.2 |
NoteGPT AI Humanizer | 55.2 | 45.8 | 45.1 | 48.7 | 10.1 |
Key findings
ChatGPT text produced the strongest detector results
The source model changed the measured outcome for every tool, but it did not change the overall leader.
ChatGPT led for every tool
Each humanizer’s ChatGPT score exceeded its Claude and Gemini scores.
GPTHuman remained first
It led by 8.1 points on ChatGPT, 9.5 points on Claude and 10.4 points on Gemini compared with the second-ranked tool.
Second place varied once
StealthGPT ranked second on ChatGPT and Claude, while HIX AI ranked second on Gemini.
Sensitivity varied by tool
The reported strongest-to-weakest gap ranged from 5.2 points for BypassGPT to 10.1 points for NoteGPT AI Humanizer.
Rank by source-model family
Humanizer | Rank on ChatGPT | Rank on Claude | Rank on Gemini |
|---|---|---|---|
GPTHuman | 1 | 1 | 1 |
StealthGPT | 2 | 2 | 3 |
HIX AI | 3 | 3 | 2 |
Humanize AI Pro | 4 | 4 | 4 |
Netus AI | 5 | 5 | 5 |
BypassGPT | 6 | 6 | 6 |
NoteGPT AI Humanizer | 7 | 7 | 7 |
Factual preservation by source model
Humanizer | ChatGPT | Claude | Gemini | Mean |
|---|---|---|---|---|
GPTHuman | 99.0 | 98.1 | 97.8 | 98.3 |
HIX AI | 91.2 | 89.3 | 89.5 | 90.0 |
StealthGPT | 90.1 | 88.7 | 88.8 | 89.2 |
Netus AI | 86.4 | 85.0 | 85.4 | 85.6 |
Humanize AI Pro | 85.0 | 83.6 | 84.0 | 84.2 |
BypassGPT | 82.3 | 80.7 | 81.2 | 81.4 |
NoteGPT AI Humanizer | 71.8 | 67.9 | 68.5 | 69.4 |
Evidence boundaries
What these tables can and cannot establish
This page reports the supplied aggregate tables without adding unsupported test details.
Aggregates, not raw outputs
The supplied file contains model-family indices but not the individual rewritten texts or detector readings.
No sample count supplied
The file does not specify the number of prompts, outputs or detector evaluations behind each aggregate.
No genre breakdown supplied
Genre-level results were not included, so the page does not claim that the pattern is identical across every writing type.
No significance claim
The observed differences describe these reported results and are not presented as proof of a permanent model-family effect.
Source-model experiment FAQ
- Do AI humanizers work equally well on ChatGPT, Claude and Gemini text?
- No. All seven tools recorded their highest detector-performance index on ChatGPT source text in the supplied results.
- Which humanizer performed best across all three source models?
- GPTHuman ranked first on ChatGPT, Claude and Gemini and also recorded the highest factual-preservation mean.
- Which tool was most consistent across source models?
- BypassGPT had the smallest detector-performance spread at 5.2 points, although its mean detector and factual-preservation indices were lower than the leading tools.
- Does this prove humanizers are optimized specifically for ChatGPT?
- No. The pattern is consistent with stronger performance on the supplied ChatGPT text, but the aggregate tables alone cannot establish why the difference occurred.
Continue exploring
Compare this result with other Arena Labs experiments
Source model is one variable. Text length, repeated passes and factual preservation can also change the outcome.